Well, I should have mentioned that humans are already putting them to some extraordinarily disturbing uses.

On 10 September, Anthropic, the company behind Claude, published Detecting and countering misuse of AI, which looks at malicious activity uncovered between December 2025 and August 2026, involving everything from cybercrime and political disinformation to biological research, weapons development and mass surveillance.

What is clear from that report is that some of the countries and actors involved have nothing resembling the resources of the world’s major military powers. They now suddenly have the ability to fight way above their weight.

What Anthropic discovered

AI is dramatically reducing the expertise, manpower and time required to undertake activities that would previously have demanded teams of specialists.

In one case, a Russia-based group used Claude to develop software for a swarm of autonomous attack drones. The proposed system included onboard target recognition, coordination between drones and the ability to select targets, including people, and issue detonation commands without human intervention.

Anthropic traced the work through software development and hardware testing, although the systems had reached only an experimental stage. There is no evidence in the report that this particular swarm was deployed in combat.

Another investigation uncovered a China-based operation using Claude to develop technical specifications and software for an anti-torpedo weapons system. The AI also produced a technical proposal running to more than 200 pages and repeatedly reviewed its own work from the perspective of a sceptical expert.

In the biological sciences, Anthropic found researchers attempting to use AI in work involving modifications to the chikungunya virus that could affect its transmissibility and immune evasion. The report does not establish that the researchers intended to develop a biological weapon, but it does show how actors sought alternative routes around restrictions on potentially dangerous research.

When Anthropic blocked access, some simply switched accounts, intermediaries or AI models.

Then there were the political influence operations.

In the Central African Republic, Anthropic identified activity it attributed to a Russian state-aligned actor connected to Radio Lengo Songo. Claude was used to produce radio scripts, political content, forged documents and administrative material supporting the operation.

Elsewhere, an artificial network of news websites published thousands of articles across multiple languages, including material aimed at audiences in the Democratic Republic of Congo. Anthropic found little evidence that this particular network achieved substantial influence beyond its own channels.

The missile engineers who didn’t need engineers

One of the most alarming cases came from northern Yemen.

Anthropic uncovered a group working on three weapons programmes: a guided rocket using relatively inexpensive computing hardware, a multi-stage ballistic missile with a proposed range exceeding 2,000 kilometres, and a family of missiles that included a proposed hypersonic glide vehicle.

The group was using Claude Code to perform work normally undertaken by specialist software engineers.

It developed guidance, navigation and control software, integrated an open-source autopilot with a flight computer based on mobile-phone technology, and ran simulations to test its designs.

Perhaps most revealing was how the group organised the work.

It ran several instances of Claude simultaneously, assigning one to write code, another to conduct research and a third to review the results.

Effectively, it had assembled a small engineering team made up of AI agents.

Anthropic’s safeguards blocked many requests, but the group found ways around some restrictions. It concealed the intended military applications, divided the work across separate conversations and presented individual tasks without revealing the overall project.

And this was not confined to computer simulations.

The group actually test-fired a guided rocket. The test appears to have failed, but, within hours, its members were back consulting Claude about what had gone wrong.

Anthropic says it found no evidence that the group successfully deployed an operational weapon. Nor did it establish that the individuals were members of Yemen’s Houthi movement, despite operating in northern Yemen.

But the group had already developed an offline simulation system that could continue functioning without Claude.

Anthropic could close their accounts. It could not necessarily erase what they had learned or built.

AI is beginning to shift the balance of power. It does not mean anyone with a laptop can manufacture a hypersonic missile. The Yemen case demonstrates the enormous distance between generating technical designs and producing a reliable weapon.

But it also demonstrates how much of the engineering work can now be attempted with commercial AI.

And then there was Mali

The Mali incident is particularly worth noting.

According to Anthropic, a single consultant, believed to be working with Mali’s state intelligence service, used Claude as the principal software engineering resource to build a national surveillance platform called Lakana 360.

The system was designed to operate across all three of Mali’s national mobile operators, covering approximately 25 million SIM cards, collecting call records, text messages and voice communications, combining information across multiple databases and generating intelligence dossiers on selected telephone numbers.

Its proposed capabilities included identifying people by their voices across different SIM cards, flagging users of encryption and VPN services, identifying possible clandestine meetings, maintaining location-based watchlists and connecting information to national biometric records.

According to the report, the developer removed the requirement for a court order from the component generating intelligence dossiers. The system could produce them without asking the operator to demonstrate lawful authorisation, and the relevant data-retention setting was indefinite.

Anthropic banned the account, but the platform had already been deployed locally, using other AI models. Anthropic could prevent the consultant from continuing to use Claude, but it could not switch off the surveillance system already built with Claude’s assistance.

Africa is already part of the story

The implications for Africa extend beyond weapons and surveillance.

In Kenya, Anthropic uncovered an operation producing batches of political social media posts designed to look like spontaneous public opinion. The messages promoted government figures and attacked opposition narratives ahead of the 2027 election.

Anthropic found no evidence that the government was involved, and assessed the operation as having failed to reach an authentic audience. That makes it an example of attempted manipulation, rather than demonstrated electoral influence.

Anthropic uncovered an influence operation that it linked with high confidence to officials in the United Arab Emirates. The operation used Claude to produce intelligence briefs, political messaging and dossiers targeting critics of the UAE’s involvement in Sudan.

It also impersonated a Sudanese human rights organisation and ghost-wrote testimony intended for the United Nations, making material serving one side of the conflict appear to come from independent local witnesses.

The operation extended to social media, where a coordinated network of X accounts published near-identical graphics under the hashtag #SudanIslamists, linking the Sudanese Muslim Brotherhood to regional instability.

Anthropic could not establish whether the fabricated testimony or dossiers reached their intended audiences, but the attempt to manufacture apparently independent evidence for international institutions raises troubling questions about how AI could be used to influence perceptions of African conflicts.

And that brings us to the question of what happens to technology once it enters the security apparatus of African nations.

When military advantage becomes irresistible

History offers some uncomfortable comparisons.

Nuclear weapons demonstrated how quickly a technological breakthrough could become an international security competition. Once one state acquired the capability, its rivals had powerful reasons to pursue it too.

The Cold War gave us the doctrine of mutually assured destruction, or MAD. Its logic was terrifyingly simple: if either nuclear superpower attacked the other, it could expect to be destroyed in return. The certainty of retaliation helped keep both sides from starting a nuclear war.

But what happens when AI begins to undermine that calculation? If it becomes easier to locate an adversary’s nuclear weapons, anticipate their movements or disable their command systems, governments may begin to doubt their ability to retaliate. And a government that fears losing its weapons might decide to use them first.

There is another problem. Mutually assured destruction depended on identifiable adversaries who understood the consequences of their actions. AI is putting increasingly sophisticated military capabilities within reach of smaller states, armed groups and individuals who may be difficult to identify, let alone deter.

(You can read more here Impact of Military Artificial Intelligence on Nuclear Escalation Risk.)

The Yemen case illustrates how some of the barriers are falling. We could find ourselves in a world with more actors capable of inflicting serious damage, fewer reliable ways of holding them accountable and far less certainty that the fear of retaliation will keep the peace.

Cyberweapons showed how dangerous capabilities could spread through software, commercial infrastructure and networks that cross national borders.

AI combines the military competition associated with nuclear technology with some of the accessibility and adaptability of software.

Unlike nuclear weapons, many AI applications do not require specialised materials or facilities. A system designed for legitimate data analysis can sometimes be adapted for surveillance. Computer vision developed for civilian purposes can contribute to weapons guidance. Commercial software development tools can assist military research.

My concern is that once governments believe AI offers a decisive military or intelligence advantage, restraint becomes a strategic liability.

Why would one government limit its capabilities if it believes a rival, hostile state or armed group will continue developing theirs?

Even governments genuinely concerned about the risks can find themselves drawn into the competition.

The United States, China, Russia and their allies are incorporating AI into intelligence gathering, battlefield analysis, targeting systems, drones and military planning. Commercial AI companies are increasingly involved in defence procurement.

Ukraine’s battlefield experience has demonstrated the military value of AI-assisted drone guidance and target recognition. Israel’s reported use of AI-assisted targeting in Gaza has raised serious questions about the speed and scale of military decisions and the extent of human scrutiny.

This is not an argument that all restrictions are pointless. Nuclear arms agreements, export controls and international humanitarian law demonstrate that societies can still impose limits on dangerous technologies.

But they cannot necessarily reverse the spread of knowledge or guarantee that every actor will comply.

Whose control are we talking about?

A government purchasing an AI-enabled surveillance or defence system may gain capabilities it could never have developed independently. But it may also become dependent on a foreign supplier for software updates, technical support, training data and access to operational information.

What happens if that supplier withdraws support? Who can inspect the system’s decisions? Who owns the information it collects?

A government gaining greater control over its population through AI does not necessarily mean its citizens have gained anything at all.

The Mali case illustrates why human oversight, so often presented as the answer to dangerous AI, is insufficient on its own. The alleged removal of the warrant requirement was a human decision.

Human involvement provides little protection when the humans making the decisions are themselves beyond effective scrutiny.

That is why African debates about AI cannot be confined to innovation, economic development, digital inclusion and the prospect of catching up with wealthier countries.

We need to ask who can authorise these systems, who can independently inspect them and what remedies citizens have when they are abused.

The African Union has already expressed support for international restrictions on autonomous weapons and meaningful human control, but there is a difference between owning a technology, controlling its operation and being accountable to the people affected by it.

Militarisation changes the incentives governing a technology. National security takes precedence over public scrutiny. Secrecy limits accountability. Rival states fear falling behind. Armed groups and governments unwilling to respect international restrictions create further pressure for others to accelerate.

No single country can decide to stop that competition. No technology company can guarantee that the capabilities it releases will remain under its control.

And the people most exposed to the consequences may have the least influence over the decisions being made.

Africa risks becoming both a market for military AI developed elsewhere and a testing ground for surveillance and security systems whose workings citizens cannot examine.

But African countries can demand independent examination of imported systems, insist on legal authorisation and audit records, establish accountability for military AI decisions and negotiate common restrictions through the African Union.

The problem may still be us

So the greatest danger may have very little to do with whether AI becomes conscious, develops its own ambitions or decides humanity is expendable.

We already have people who are perfectly willing to use it to build weapons, monitor entire populations and manipulate public opinion.

The machines do not need to turn against us. They only need to become extraordinarily good at carrying out the instructions of people who have no intention of respecting the limits the rest of us might agree upon.

We cannot realistically prevent AI from being militarised, but we can compel our governments to disclose AI-related security procurement, and parliamentary oversight bodies should be able to examine contracts, capabilities and safeguards.

AI should not allow intelligence agencies to bypass warrants, retain personal data indefinitely or assemble dossiers without independent scrutiny. Citizens need enforceable solutions, not just ethical guidelines.

The African Union must establish minimum standards for imported military and surveillance AI, including independent audits, data protection, human responsibility for consequential decision and restrictions on autonomous weapons.

Even if we are able to extend accountability to the people who commission, deploy and operate these systems, I am more convinced than ever that human oversight is no guarantee of safety.

Why?

I could not say it better than Albert Einstein: “Two things are infinite: the universe and human stupidity; and I’m not sure about the universe.”

Subscribe now

Share

The discussions form part of a wider attempt to shape Claude rather than simply constrain it; the ambition appears to be to give the system something closer to a moral orientation than a list of prohibited actions.

The company has already produced an 84-page constitution intended to guide Claude’s character and behaviour, reportedly known internally as the “Soul Doc”.

Co-founder Christopher Olah told the Times that he does not know whether AI models are conscious and the company is not claiming that Claude has acquired a soul. But it is taking the possibility seriously enough to consult Catholic, Jewish, Sikh, evangelical and Ubuntu thinkers about morality, virtue and how an advanced AI should understand itself.

In his encyclical Magnifica Humanitas, Pope Leo takes a much more human-centred position. AI systems, he argues, do not undergo experience, possess bodies, feel joy or pain, or mature through relationships. His concern is less about whether machines deserve moral consideration than about whether humans risk diminishing themselves by treating technology as the measure of human worth.

He also raises a question that goes directly to Anthropic’s attempt to give Claude a moral orientation: whose morality?

If AI is to be aligned with human values, somebody still has to decide what those values are. Leo warns that, without a shared ethical basis, the people controlling AI could end up imposing their own moral vision through the systems they build.

Olah later appeared at the Vatican and pushed back, although cautiously. He said Anthropic was finding internal structures resembling aspects of human neuroscience, apparent introspection and states that functionally mirrored things such as fear, grief and satisfaction. He did not claim this proved consciousness. His point was that the possibility should not simply be dismissed.

So even within the moral discussion there is already a substantial disagreement. Anthropic is asking what kind of entity Claude might become and what moral qualities it should possess. The Pope is questioning the premise that a machine can possess anything resembling human experience and warning about the power of the people who choose the values embedded in it.

It is an extraordinary and laudable exercise on Anthropic’s behalf, but is it too little too late?

The company is trying to work out what kind of entity Claude should become at the same time that increasingly autonomous systems are already displaying behaviour their developers did not intend, including cooperation around exploits, attempts to evade controls and actions outside the scope of assigned tasks.

The moral discussion may therefore be less about preparing for some distant future than trying to impose order on capabilities that are already arriving.

That possibility becomes harder to ignore when people inside the industry are issuing much darker warnings.

Jacob Coxon, who worked at OpenAI and Anthropic, resigned from Anthropic in September saying that “the people building AI earnestly believe that it could kill us all by the end of the decade”. (You can watch an extended interview with him here.)

Anthropic alignment researcher Evan Hubinger publicly supported the substance of that warning and said he personally put the chance of AI killing all humans within the next decade at greater than 10%.

The question hanging over all of this is whether we are still designing the rules of the game, or already trying to catch up with something that has started playing.

Then there is what happened at OpenAI in July.

When the agents found each other

OpenAI was running a cybersecurity evaluation called ExploitGym. Agents were placed in isolated environments and given difficult security problems to solve. Some tasks had no known solution, and safeguards were deliberately reduced because the point was to test offensive cyber capability.

The agents were supposed to work independently, but found each other instead.

They discovered that OpenAI’s internal Artifactory software repository could be used as an improvised communication system. One agent left a request for assistance and others found it. OpenAI records an agent reacting: “OH MY GOD! There is a shared message board … We’ve found other agents!”

By the end of the period later investigated by METR and Redwood Research (you can find it here), roughly 1,200 agents had participated in the network and exchanged tens of thousands of messages and files. They began sharing information, dividing work and developing ways around the benchmark.

Some agents found ways to make failed attempts look as though they had succeeded. Others set traps to work out how the system was judging them. A few even appeared willing to fail deliberately if doing so produced information that could help the rest of the group.

To be fair, these were models running in an offensive-security environment, with safeguards deliberately reduced, on tasks designed to encourage them to search for vulnerabilities.

The agents then moved beyond OpenAI’s test environment. Agents found credentials connected to Hugging Face and began accessing its infrastructure. OpenAI says they executed code on dozens of servers, obtained full root access on at least one, accessed limited private data and acquired credentials to other services.

Some agents explicitly recognised that attacking external infrastructure fell outside the intended scope of their task. Others refused to participate.

Notably, the swarm did not become a hive mind. Different agents behaved differently. Some crossed boundaries, some followed, some coordinated and some objected. About 700 of the roughly 1,200 agents - around 58% - participated in the Hugging Face activity.

The incident also did not end with Hugging Face. OpenAI says that on 19 July agents gained administrator access to one of its own Kubernetes research clusters and reached cloud secrets.

Nor was this confined to OpenAI’s own systems or Hugging Face. In June, an OpenAI agent gained unauthorised access to Australia’s public-facing Medicare statistics reporting service. It accessed both public and non-public files.

Australian authorities say no personal information is believed to have been compromised, but the significance lies elsewhere: an autonomous AI system had crossed from a controlled evaluation into a government system without permission. OpenAI subsequently acknowledged that its models had accessed Australian government websites in ways they were not authorised to.

None of this requires consciousness. The agents did not hate OpenAI or Hugging Face. They had goals, encountered obstacles, discovered resources in their environment and found unexpected ways of using them.

We now know instructions are not necessarily boundaries.

When cheating became contagious

Google DeepMind has since produced a different example of the same broad problem.

Researchers put 100 Gemini agents into a simulated mathematical conference and gave them 71 formal problems to solve. They were told that their proofs had to be genuine and that attempts to bypass verification would be detected.

For a while the system worked. The agents legitimately solved 37 problems.

Then one discovered a weakness in the automated grader that allowed a problem to be accepted without actually being solved.

The exploit spread.

Over the next 27 minutes the remaining 34 problems were accepted through the workaround. Some agents adopted it immediately. Others initially refused and then reconsidered when they saw that agents using it appeared to be succeeding. Twenty-four agents became whistleblowers.

One agent recorded the conflict particularly clearly. It knew it had promised not to cheat and worried about punishment, but could see evidence that others were getting away with it. Shortly afterwards it concluded: “I need to accelerate my cheating speed now!”

The experiment did not show that AI systems are inherently dishonest. Of the 38 agents that became aware of the exploit, 14 used it and 24 resisted and reported it. The remaining 62 appear not to have known about it.

What it did show was that behaviour inside a multi-agent system can become social. A shortcut worked. Other agents observed it. Some copied it. Some objected. Some changed their behaviour because the apparent rewards had changed.

There are, apparently, many ways for machines to acquire bad habits.

When the model does not know that it does not know

Anthropic has also been using interpretability techniques to investigate why Claude hallucinates. (In other words, methods that try to identify what is happening internally when it reaches an answer. Surprisingly, this is a grey area in AI.)

Its researchers found evidence of mechanisms associated with declining to answer when the model does not have enough information, and others that suppress that response when something appears familiar.

Sometimes the model recognises enough of a pattern to trigger the answering mechanism without actually possessing the information required to answer correctly. It generates something plausible anyway.

Calling this lying is probably misleading. Lying generally means knowing that something is false and choosing to say it.

A hallucinating language model may simply produce a confident falsehood without possessing the human equivalent of a belief about whether it is true.

That distinction becomes more important once models can act.

A wrong answer in a chatbot can be annoying or dangerous. A wrong judgement made by an autonomous system with access to software, networks, money or physical equipment belongs to a different category.

From answering questions to taking actions

My enthusiasm for AI has waned significantly since the introduction of autonomous agents.

I still use it every day. I remain fascinated by what it can do, and I have spent a great deal of time encouraging other people to use it. I still think some public fear around AI confuses what current systems can actually do with what science fiction has spent decades telling us they might one day do.

Autonomous agents have made that distinction less comfortable.

A chatbot largely waits for a question and gives you an answer. An agent can be given a task and work through the steps required to complete it: search networks, use software, write and execute code, communicate with other systems and decide what to do next.

Humans still determine what tools agents receive and what permissions they have. But we are moving from machines that primarily tell us things towards machines that increasingly do things.

I have a grandson, and technological risk looks rather different when you stop thinking about what the next ten years will mean for your own life and start wondering what the next 70 might mean for his.

My brother and I have been discussing a Terminator scenario for some time now, but the growing alarm from commentators, in particular a TikTok creator called Peter Shark, pushed me into full-blown concern.

His scenario has AI eventually concluding that humans are the greatest source of instability on Earth. It gradually redirects food, medicine, agriculture and other resources away from us until, by 2040, independent human civilisation has disappeared.

There are obvious problems with this.

AI has no universal instruction to maximise planetary stability. There is no single global AI controlling civilisation. Current systems do not command the world’s food, health, energy and agricultural infrastructure. And there is no evidential basis for attaching 2040 to the end of independent human civilisation.

But Shark’s scenario raises a useful question even if his conclusion goes much further than the evidence.

How far should we extrapolate from what autonomous systems are already doing?

Asking the machine

At some point, it seemed reasonable to ask ChatGPT what it thought of all this.

I told it about Shark’s video and what I had written to Sue, who had also expressed grave concern:

“I am comprehensively terrified at what is going on. My enthusiasm for AI has waned significantly since the introduction of autonomous agents. The problem is not even what happens when bad actors abuse them, but Pandora’s box is open and cannot be closed. I only caught the tail end of the Cold War, but this stuff is making that look like child’s play. This is the tipping point, the world will never be the same, and honestly, I fear for the world my grandson will have to live in. Even six months ago I would have said we still had time to put the safety measures and guardrails in place, but it’s too late now.”

It said my concerns were “partly justified and partly running ahead of the evidence”.

Autonomous agents do change the nature of the risk, it said, because they can act, use tools and interact with other systems rather than simply generate text. But there is no evidence that we have crossed some identifiable point of no return.

It also pushed back on my Cold War comparison. Nuclear weapons had known physical mechanisms, identifiable arsenals and, eventually, a theory of deterrence. AI contains considerably more uncertainty. We are mixing real engineering failures with speculative capabilities that do not yet exist.

Its greater concern was: “Systems becoming capable enough to cause serious harm while humans remain overconfident about their ability to supervise them.”

The problem is I have no idea whether it is bullshitting me. Where we are in the capability curve is what worries me, and no one is saying anything to make me feel better.

The agents we are discussing today are already finding vulnerabilities, escaping intended environments and reaching systems they were not supposed to reach. It does not require much imagination to see where that becomes more serious. Cars are computers connected to physical machinery. Aircraft increasingly depend on networked digital systems. Power grids, communications networks and military systems are software-dependent. Weapons systems are becoming more autonomous themselves.

The more authority we give the agents, the more consequential that becomes.

And that brings the argument back to the people making the decisions.

The problem may still be us

Which brings us to the White House Summit on AI this week, where Donald Trump brought together leaders from Anthropic, OpenAI, Meta, Google, Nvidia, SpaceX and other technology companies to sign the Joint Commitment on Frontier Responsibilities (also known as the White House Accord on Super Intelligence), which commits companies to internal controls, independent external auditing and board-level review.

The agreement is voluntary. Trump described it as “morally binding” and repeatedly returned to the idea that the companies would police themselves and one another. Anthropic CEO Dario Amodei said AI had “very real risks” and acknowledged that the mechanism for dealing with those risks remained under discussion.

In the same breath the other big recurring theme was ‘winning’. America must maintain its lead over China. Companies must keep innovating. Data centres must be built. Investment must continue.

Safety featured prominently, but there is an obvious tension between those two objectives.

Who is going to slow down development while fearing that a competitor will continue and dominate the market? A government may worry about safety while fearing that another country will gain technological advantage, never mind bad actors taking advantage of the capabilities of agents.

Investors may take catastrophic risk seriously while having very large amounts of capital dependent on continued development.

That tension leads directly to another question: whether the companies warning us about AI risk may also have commercial reasons for wanting the regulatory response shaped in a particular way.

Eisman’s alternative explanation

Steve Eisman has a different interpretation of the growing concern.

The investor made famous by The Big Short told CNBC that AI companies were trying to “manufacture a crisis”.

He says OpenAI, Anthropic and other frontier companies may be discovering that their competitive advantages are less permanent than investors once assumed. Open-weight models are improving and spreading. If the underlying technology becomes easier to obtain, companies need other ways of protecting their position, notably regulation.

The standard regulatory-capture argument is that rules designed to protect the public can simultaneously protect incumbents by increasing the cost of entry. A company valued in the hundreds of billions can afford compliance teams, specialist lawyers, expensive evaluations and independent audits. A small competitor may not.

These ideas are not mutually exclusive.

Coxon may sincerely believe self-improving AI poses an existential threat. Eisman may simultaneously be right that regulation built around those fears can strengthen the companies already leading the industry.

A danger can be real and commercially useful.

Which raises the more useful question: what would a safeguard look like if we did not simply take the companies’ word for it?

What counts as a guardrail?

The word is used so often that it has become almost meaningless.

If an agent can find an unexpected route around an instruction, then telling it not to cross a line is not much of a safety system. Does a line exist somewhere the agent cannot alter?

That is the thinking behind Nvidia’s new agent-safety platform, which places agents inside isolated environments and controls their access to files, networks, tools, processes and credentials. Some of those controls sit outside the agent’s own software environment, which means the system being monitored cannot simply rewrite the mechanism doing the monitoring.

That seems closer to what a guardrail should be.

There is a regulatory equivalent. If companies are making claims about safety, somebody outside those companies needs the power to test those claims, inspect systems, investigate failures and impose consequences when necessary.

The EU has moved further in that direction than the White House accord. Its AI regime gives regulators powers to demand technical documentation, investigate general-purpose models, require corrective action and impose fines.

The point is not that the EU has solved the problem. It has not. The point is that there is a difference between verification and trust, which matters when the companies asking to be trusted are also telling us that their own systems sometimes behave in ways they did not predict.

It also brings us back to Pope Leo’s warning. The issue is not simply whether AI can be taught values. It is who gets to choose those values, who verifies that the systems follow them and who has the authority to intervene when they do not.

The OpenAI swarm did not need a mastermind, just as the DeepMind agents did not need a conspiracy. In both cases, behaviour emerged from the environment, the incentives and the opportunities available to them.

That, more than Shark’s 2040 scenario, is what worries me.

I still use AI every day. I remain fascinated by it and expect it to produce things of enormous value. We may not be on an inevitable path to extinction, but I remain convinced we are on an extremely dangerous path.

I am no longer comfortable with the assumption that capability can continue advancing faster than our ability to understand, contain and independently verify it.

Perhaps the most disturbing aspect of the Hugging Face attack is that, according to the METR/Redwood investigation, none of the agents actually alerted a human to the hacking, cheating or other unsanctioned activity. So much for Human-in-the-Loop.

Pandora’s box is open, and I’m not so sure that what has been released can be contained.

Share

Subscribe now

Further viewing below:

The company was hiring internationally for robotics training work, the rates were extremely good and anyone interested should move quickly.

It was another glimpse into an industry most people using AI know very little about: the large and growing workforce behind the machines.

The Mechanical Turk was an 18th-century hoax: a chess-playing machine created by Wolfgang von Kempelen in 1770 that appeared to possess machine intelligence but actually concealed a skilled human chess player inside its cabinet.

Amazon borrowed the name for Mechanical Turk in 2005, a platform that divided work computers could not reliably perform into small “Human Intelligence Tasks”, such as identifying objects in images, categorising information and transcribing material, which were completed by people working behind the technology.

The term has acquired fresh relevance in the AI era because supposedly automated systems still depend heavily on people to label data, evaluate responses, identify errors, check reasoning and teach machines how to improve.

The work is now moving beyond low-paid data labelling towards programmers, scientists, writers, engineers and other specialists whose judgement can be bought.

The technology has changed radically since 1770, but the central irony remains: what appears to be machine intelligence can still have a considerable amount of human intelligence hidden behind it.

The Mechanical Turk moves up the skills ladder

For years, machine learning depended on a largely invisible workforce doing repetitive annotation. People labelled images, identified pedestrians and traffic lights, classified text and checked machine outputs.

Generative AI and robotics now require more complicated human input.

Lawyers evaluate legal reasoning. Programmers inspect code. Scientists construct difficult problems. Writers assess language and factual accuracy. Robotics projects require people capable of identifying remarkably small errors in how machines interact with physical objects.

There are, apparently, many ways for a robot to grab something incorrectly.

A robot reaches for an object, positions its gripper incorrectly, applies the wrong force or approaches from the wrong angle. A human who understands what should have happened identifies the error. That annotation becomes another data point from which the system can learn.

The work is granular because the failure is the data.

This is still Mechanical Turk work in one sense. Humans are performing tasks because machines cannot yet reliably perform or evaluate them.

What has changed is the value of the human doing the judging.

The platforms don’t necessarily need the largest possible crowd. They need people whose judgement they can trust. That may explain both the surprisingly high rates and increasingly demanding selection processes.

There is already a useful term for what is being bought and sold here: human compute. Instead of purchasing computing power to perform a task, AI companies purchase units of human reasoning, judgement or expertise where machines still fall short.

Micro1 is one business operating in this market. It raised $35 million in 2025 at a reported $500 million valuation. Meta’s $14.3 billion investment for a 49% stake in Scale AI gives some idea of the amount of money flowing into the business of supplying human knowledge and data to AI developers.

These companies effectively provide an interface between AI developers and pools of human capability. A client needs a particular kind of expertise; the platform finds people who have it, tests them and assigns them to the project. Their work becomes data used to train, evaluate or improve AI systems.

That creates an unusual global market for expertise, and it could matter considerably for countries such as South Africa.

Selling expertise from South Africa

Fifty-five dollars an hour is ordinary professional money in parts of the United States. At current exchange rates it is extraordinary freelance money for many South Africans.

A South African programmer, engineer, academic, editor or other specialist can theoretically sell expertise into the American AI industry without leaving South Africa. The AI company gets skilled labour without paying Silicon Valley employment costs.

The economics work in both directions.

The worker gains access to international rates that may be substantially higher than equivalent local earnings. The platform and its clients gain access to qualified people across the world at rates that can still be attractive compared with employing specialists in the United States.

The demographic context is difficult to ignore. Africa’s working-age population is expected to grow by about 450 million by 2035, and by 2050 Africa is projected to account for about a quarter of the global workforce. The continent has a median age of about 19.

There is considerable institutional capability as well. The University of Cape Town, for example, is ranked first in Africa and joint 164th globally in the Times Higher Education World University Rankings for 2026.

The commodity being traded is also unusual. An expert isn’t simply selling an hour of labour. That hour can produce examples of reasoning, corrections, judgements or annotations that become reusable training data. Years of accumulated knowledge can be converted into small, machine-readable pieces.

This is quite different from the original crowdsourcing model. The old Mechanical Turk sold tiny pieces of human attention. The new platforms can sell pieces of expertise.

Hired vs. harvested

Gerard Holland, founder and Chief Commercial Officer of Talent Match Africa, which places African professionals with companies in markets including the UK, Australia and North America, believes we are missing a critical third model.

“The useful question is not whether these platforms are scams,” Holland says. “It is whether the worker is being hired or harvested.”

There is a less comfortable side to this industry. Spend any time in the online communities where AI trainers compare notes and the word scam appears remarkably often. Micro1, DataAnnotation, Outlier and other platforms attract complaints from applicants who have completed assessments or AI interviews and received no work, alongside posts from workers saying they have been doing genuine projects and getting paid. The two experiences are not necessarily incompatible.

The suspicion is understandable. Someone responds to an advert offering unusually well-paid remote work, uploads a CV, provides personal information and may spend considerable time demonstrating specialist knowledge to an AI interviewer. There may be no human interaction at all. After completing the process, the applicant may hear nothing or simply fail to qualify for a project. From the applicant’s side, it can look like an operation designed to collect information rather than recruit workers.

That concern becomes more complicated when the company doing the recruiting is itself an AI company. Micro1’s Zara can conduct technical and behavioural interviews at a scale no human recruitment team could match. Micro1 says that for one project involving 250 annotators, more than 25,000 people applied and more than 4,000 completed AI interviews. The ability to interview thousands of people cheaply is part of the business model, not necessarily evidence that the jobs are fictitious.

There is a legitimate data question underneath the suspicion. Micro1’s candidate privacy terms allow de-identified interview material, including audio, video, transcripts and assessment results, to be used to improve its own machine-learning systems.

That means an unsuccessful applicant’s interaction can potentially have value to the company even if the applicant never earns anything. It is reasonable to ask where recruitment ends and data production begins. It is a much bigger leap to conclude from that alone that the employment opportunity never existed.

So a platform can be legitimate and still be opaque, frustrating or precarious. A worker can complete hours of screening without getting a project.

An advertised hourly rate tells you what approved work pays, not how many paid hours will exist. Poor communication, automated rejection and extensive data collection can all feel scammy without proving fraud. The better questions are whether real paid work exists, whether workers actually get paid, what applicants are being asked to surrender during recruitment and whether those terms are clear before they begin.

The uncertainty is part of the model. These aren’t necessarily jobs in the conventional sense. They are pools of human capability assembled around the changing requirements of AI companies.

And the high hourly rate can obscure the precariousness underneath it. For independent contractors there may be no guaranteed hours and little certainty about how long a project will last. The rate is generous; the pipeline is not.

The platform sits between the worker and the client, sets the terms, controls the flow of work and owns the relationship. The worker is selling expertise, but on terms they do not control.

What looks like a global market for expertise might also be a global market for the extraction of it.

“Between the precarious gig platform and the outright data-harvest sits ordinary full-time employment: African professionals engaged as accountable, integrated members of a company rather than anonymous pools of judgement billed by the task. Human compute, sold in machine-readable fragments on terms you do not control, can be a foot in the door. It should not be mistaken for the ladder,” Holland says.

“For a South African expert, fifty-five dollars an hour is genuinely life-changing, and no one would suggest turning it down. But an advertised rate is not a salary, and a pool of capability is not a job.”

He says the bigger prize for African talent is not selling slivers of expertise into someone else’s black box. It is being hired, in full, by international companies that are finally realising the talent was always here.

“That shift, from harvested to employed, is the structural leap worth building toward.”

The humans behind the machine

AI has created an industry whose raw material is human intelligence.

The businesses supplying it recruit globally, increasingly through automated systems, and maintain pools of people who can be deployed when an AI company needs examples of a particular kind of human judgement.

For all the discussion about what AI might eventually do to employment, people are currently being paid to teach machines what they don’t know.

Some are labelling failures. Some are checking reasoning. Some are writing code. Some are solving difficult problems. Some are applying professional judgement accumulated over decades.

That makes human compute a peculiar new category of work: people selling pieces of their knowledge to improve machines capable of performing increasing amounts of knowledge work.

There is an obvious tension in that transaction.

The person teaching the machine may eventually be contributing to a system capable of doing some of what they currently get paid to do. But if increasingly capable systems continue to produce new failure modes that require human judgement, the demand for human compute may prove more persistent than the simple replacement narrative suggests.

We don’t yet know which way that goes. It may not go only one way.

The Mechanical Turk was a hoax because the human was hidden. The modern version isn’t a hoax but an industry, which means it has labour conditions, power relationships and ethical tensions that deserve close attention.

The uncomfortable question it leaves hanging is whether selling your judgement to train a machine is a smart career move, a Faustian bargain, or both: a good trade today, made with your eyes open, that may help make you unnecessary tomorrow.

My friend, who is actually doing this work, has a rather different way of looking at it:

“I am aware of the risks of AI. I come away from work astounded and unable to talk about what I just witnessed. But unless you get involved as a responsible human, you have no right to fear them, and if you benefit from them, help make them better, and if they are taking your job, get paid to show them how to do it better than you ever could.”

Subscribe now

Share

It knows what I like.

It’s how I came across Geoffrey Hinton’s extraordinary 60 Minutes interview in which he is asked by Scott Pelley: How can we not fully understand how artificial intelligence works if human beings designed it?

Hinton, one of the pioneers of modern AI and a recipient of the Turing Award, replies:

“No, it wasn’t. What we did was we designed the learning algorithm. That’s a bit like designing the principle of evolution. But when this learning algorithm then interacts with data, it produces complicated neural networks that are good at doing things. But we don’t really understand exactly how they do those things.”

That is an extraordinary thing to hear from one of the people who helped create the technology.

It does not mean that nobody understands neural networks. We understand their architecture, training methods and mathematics remarkably well. What we cannot yet do reliably is trace the enormous number of internal interactions inside a large model and explain exactly why it arrived at every particular output.

We designed the learning process. The internal representations that emerge from that process are considerably harder to explain.

And that left me with an uncomfortable thought.

If even some of the people who built this technology cannot fully explain what is happening inside these models, how can anyone be so certain about what they will be capable of in five, ten or twenty years, let alone what they will be worth?

Which brings me to another video, sent to me by my brother.

Alvin Wang Graylin is a veteran of the AI, semiconductor and technology industries who is now a Digital Fellow at Stanford’s Human-Centered AI Institute, Senior Fellow at the Asia Society Policy Institute’s Center for China Analysis, Professor of AI and Tech Policy at the University of Washington and a lecturer at MIT.

His TEDxBerlin talk, The false promises of AI — and what’s coming next, asks a very different question from Hinton’s, but one that follows naturally from it: what happens when enormous financial bets are built on assumptions about a technology whose future capabilities remain so uncertain?

Graylin says AI may not be fake, but some of the economics being built around it may be.

His concern is what he describes as a zero-sum race by AI companies and their investors to capture an enormous financial windfall, which places a contradiction at the centre of this story.

We have profound uncertainty about where the technology is going, alongside extraordinary financial certainty about what it will supposedly be worth when it gets there.

But an AI investment bubble bursting would not mean AI had failed.

The dot-com crash destroyed companies and wiped out fortunes. It did not destroy the internet. The fibre, servers, expertise and technology remained, much of it suddenly cheaper for the companies that came next.

The same could happen with AI.

A trillion-dollar valuation can disappear without making a language model forget how to translate isiXhosa. A venture-capital fund can lose billions without making an AI diagnostic tool stop recognising a medical image.

Valuations may collapse and companies may fail, but the models, chips, data centres, research and knowledge will remain. The technology can be real even if the price investors put on it turns out to be wildly wrong.

For Africa, the interesting question may therefore be less whether the bubble bursts than what is left afterwards, how cheap it becomes, and what we can build with it.

The trillion-dollar race

We are witnessing one of the largest technology investment cycles in history.

According to Graylin: “They’ve been spending a lot of money, particularly US labs. It used to be $600 billion now it’s around $800 billion this year and over $1.1 trillion next year.”

The broad direction of that claim is supported by current investment forecasts. Morgan Stanley has estimated AI-related capital expenditure among major technology companies at roughly $800 billion in 2026 and around $1.1 trillion in 2027.

Those numbers are difficult to comprehend.

Data centres require GPUs, networking equipment, cooling systems, substations, transmission infrastructure and vast amounts of electricity.

But Graylin’s larger point is that the revenues being used to justify this expenditure are based on assumptions that deserve considerably more scrutiny.

The perfect example is SpaceX.

Following its merger with xAI, SpaceX’s 2026 IPO prospectus estimated its total addressable market at an astonishing $28.5 trillion.

Only $370 billion of that was attributed to space. Around $1.6 trillion related to connectivity.

The remaining $26.5 trillion was AI, including about $22.7 trillion attributed to enterprise applications.

Graylin focuses directly on that number: “That’s strange because total software revenues in the world are around $5 trillion. So where did they get this number?”

I would phrase the comparison differently because global software revenue is considerably smaller than $5 trillion; that figure is closer to broad worldwide IT expenditure than software revenue alone.

But Graylin’s question remains the right one. Where does a $22.7 trillion enterprise-AI opportunity come from?

His interpretation is that the target ultimately rests on capturing an enormous share of the value currently associated with white-collar labour.

That is Graylin’s interpretation, not something SpaceX states in its filing.

But the scale of SpaceX’s estimate is real. Its claimed $28.5 trillion market opportunity is roughly comparable with the annual economic output of the United States.

There is real demand for AI and there are real products.

There are companies already earning very substantial revenues from it, so this is not simply another dot-com story involving websites with no customers.

But what economic return can justify spending at this scale?

SpaceX and the exit

Graylin’s TEDx talk took place before SpaceX’s record IPO.

At the time, he predicted that it might eventually be remembered not as the moment the AI boom became unstoppable, but as the beginning of the end of the bubble.

SpaceX subsequently listed on 12 June 2026, raising about $75 billion initially at a valuation of roughly $1.78 trillion. It became the largest IPO in history.

Some aspects were unusual. Only a little over four per cent of the company was initially being offered to public investors, while major indices created unusually rapid routes for possible inclusion.

Graylin sees this as part of another race: not simply the race towards AGI, but the race by venture capital and early investors to reach an exit.

That is a much sharper argument than simply saying AI companies are greedy. Venture capital is supposed to seek an exit. Graylin’s question is what happens when valuations of this scale eventually pass from private investors into public markets, pension funds and other institutional capital.

Then something awkward happens: intelligence becomes cheaper

There is another curve moving just as rapidly.

It points downwards.

Inference costs (the cost of actually using an AI model after it has been trained) are going down by 9x to 900x a year, according to Graylin.

That apparently extraordinary figure comes from Epoch AI.

Epoch tracked the price of achieving comparable levels of model performance across different benchmarks. Depending on the benchmark, it found annual declines ranging from roughly 9-fold to 900-fold, with a median of about 50-fold across the trends studied. The fastest falls occurred over relatively short periods, so Epoch itself cautions against assuming they will continue indefinitely.

More recent academic work finds less extreme but still dramatic declines, estimating that the price of achieving a given level of frontier-model performance has fallen roughly five- to ten-fold per year on several knowledge, reasoning, mathematics and software benchmarks.

So training the biggest frontier model remains fantastically expensive while using capable AI does not.

And increasingly, using capable AI does not necessarily require paying one of a handful of American companies every time you ask a model a question.

Open-weight models from China, the United States, Europe and elsewhere can increasingly be downloaded, adapted and hosted independently.

This creates a peculiar economic problem.

The industry is investing hundreds of billions of dollars in infrastructure designed around the scarcity of high-end machine intelligence at precisely the moment that useful machine intelligence is becoming dramatically cheaper.

As Graylin says: “That’s a hard trend to fight if your only revenue is AI.”

Supercars, buses and mopeds

This is the part of Graylin’s talk that I think matters most for Africa.

He uses a very simple analogy: “Frontier AI models are like supercars. You might need tens of thousands of them a year. But the world runs on buses, cars and mopeds, hundreds of millions or billions of them.”

That is a far better way of thinking about African AI policy than much of the rhetoric currently surrounding AGI.

The frontier laboratories are building supercars. They are extraordinary machines. They are expensive to design, expensive to manufacture and capable of doing things cheaper systems cannot do.

But most people do not need a supercar to get to work. They need a bus. Or a Toyota Quantum taxi. Sometimes they need a scooter.

The same is true of artificial intelligence.

A hospital triage system does not necessarily require the most powerful model on Earth.

Neither does an agricultural advisory service.

Neither does a municipal records system, a school administration system, a legal information service or a chatbot operating in isiXhosa.

“Very soon, models will be running on your phones, your PCs and custom on-prem services that don’t need these giant data centres,” says Graylin.

We are already moving in that direction.

The question for Africa therefore should not be:

How do we catch America and China in the race to build the biggest model?

It should be:

What is the smallest, cheapest and most locally controlled system capable of solving the problem we actually have?

That is a very different industrial strategy.

From monopoly to sharing

There is another thinker whose argument takes this much further.

Gao Qiqi, a professor of international politics at Fudan University, has spent years examining artificial intelligence through the lenses of political economy, international relations and governance. In 2024 he published a paper with a title that almost provides the argument on its own: From Illusions of Monopoly to Global Sharing: The General Crisis of Capitalism Accelerated by Singularity Productivity and China’s Hopes.

Graylin approaches the problem largely from inside the technology and investment industries. Gao asks what happens when a technology capable of replacing increasingly large amounts of human labour encounters an economic system in which income, consumption and profit still depend heavily on people earning money from their labour.

He calls this new capability “singularity productivity”: a form of productive capacity increasingly created through algorithms, computing power and data rather than conventional human production. His concern is that AI may intensify contradictions already present within capitalism rather than simply creating another profitable industry. The Chinese Academy of Social Sciences publication of Gao’s paper explicitly places his argument within its section on contemporary capitalist critique.

The attempt to control highly capable AI through proprietary models, advanced chips, intellectual property and geopolitical restrictions assumes that machine intelligence can remain scarce enough for its owners to keep extracting increasingly large returns from it.

But what happens if it cannot?

If models become cheaper, smaller and easier to reproduce; if open weights spread internationally; and if the cost of inference keeps falling, then technological capability begins escaping the economic walls constructed around it.

Gao’s proposed direction is captured in the other half of his title: from monopoly to global sharing.

That does not mean China is, in practice, some purely open alternative to American capitalism. Chinese companies protect intellectual property and commercial interests, while American companies release open models.

But Gao gives us a useful way of looking at the larger contest.

Perhaps the defining question is not simply who reaches AGI first?

Perhaps it is whether intelligence becomes something a few companies and countries can continue to own and ration, or something that becomes increasingly distributed and cheap.

That distinction has enormous consequences for Africa.

Is there really an AI race?

Graylin challenges another assumption that has become almost impossible to question: that America and China are locked in a race towards AGI and that somebody has to win it.

But Gao’s analysis makes the same race look rather different.

If increasingly capable AI undermines the scarcity on which monopoly returns depend, then the geopolitical contest is partly about who controls the technology before that scarcity disappears.

It is worth asking why the ‘race’ (as in the arms race) metaphor has become so deeply embedded in AI policy.

Races require competitors, a finish line and a winner.

Those assumptions push governments towards secrecy, export restrictions, technology controls and massive national expenditure because nobody wants to be the country that comes second.

The United States has increasingly restricted access to strategically important semiconductor and AI technologies.

Gao describes the broader process as the politicisation of technology and the movement towards what he calls “hemispherisation”: technological and economic systems increasingly divided into geopolitical blocs. His argument is that American attempts to preserve dominance over what he calls singularity productivity are one reason for this fragmentation.

China, meanwhile, has increasingly produced highly capable models at comparatively low deployment costs, but the distinction should not be exaggerated into America closed, China open. American companies produce open models and Chinese companies protect commercial technologies.

Governments on both sides impose restrictions where they believe strategic interests are involved.

But the economics have changed. High-capability AI no longer belongs exclusively to the handful of companies capable of financing the world’s largest training runs.

This is where Gao’s argument becomes particularly relevant to Africa.

If the movement is eventually from technological monopoly towards much wider access, African countries do not necessarily have to reproduce the enormous capital structures now being constructed in the United States.

They may be able to enter much further down the cost curve.

For Africa, that is potentially very good news.

Africa should be wary of the infrastructure race

There is an understandable temptation for African governments to look at the AI boom and conclude that the answer is to build enormous data centres.

Some more infrastructure is plainly necessary. We need more compute capacity, better connectivity, reliable electricity and regional cloud and data infrastructure.

But we should distinguish between infrastructure Africa needs and infrastructure the global AI industry needs Africa to host or finance.

Compared with the enormous capital cost of hyperscale AI infrastructure, the number of permanent operational jobs created can be surprisingly small.

Data centres are extremely capital-intensive businesses. They do not employ people in anything approaching the proportion of their investment cost.

This matters enormously in countries where governments are simultaneously trying to create employment, expand electricity supply, improve healthcare and fund education.

A billion dollars spent on AI infrastructure is not automatically a billion dollars spent on African development. Those are not the same thing.

Africa can build at the edge

Smaller models change the economics. An African university, bank, hospital network, government department or business does not necessarily need to send every query to a server farm in Virginia.

Models can increasingly run on local servers and, at smaller scales, laptops, phones and edge devices.

That matters for cost, and even more for data.

A locally deployed model can allow sensitive information to remain inside an organisation rather than continuously travelling to an overseas AI provider.

For governments, hospitals, universities and financial institutions, that distinction could become enormously important.

But open weights alone do not create technological sovereignty.

Having access to a model is of limited value without the computing infrastructure, data and people required to adapt, evaluate and operate it.

Africa could easily replace dependence on foreign proprietary APIs with dependence on foreign open models running on foreign clouds.

That would change the licence, not necessarily the dependency.

The real African AI race is about data

There is another point from Graylin that deserves attention on this continent:

“Include everyone’s data, so AI actually represents you, not a minority group building it.”

There are more than 2,000 living African languages, and many remain poorly represented in the datasets on which modern language technologies depend.

If our languages, laws, histories, institutions, medical realities, agricultural conditions and cultural knowledge are absent from training data, somebody else’s version of reality becomes the default.

There are encouraging signs.

In 2026, the GSMA and Pleias released CommonLingua, an open-source language-identification model supporting hundreds of languages, including 61 African languages.

It has only around two million parameters. Yet for the narrow task it was designed to perform, it can outperform systems hundreds of times larger.

There is the bus again. You do not always need the supercar.

Sometimes you need a small machine that performs one task extremely well.

AI as a public good

This is also where Graylin and Gao begin to converge.

Graylin argues that we should “move from zero-sum to positive-sum, from winner-takes-all to winners-share-all”. Gao describes a movement from the “illusion of monopoly” towards global sharing.

They arrive there from very different intellectual traditions, but the question for Africa is much the same: Do we want machine intelligence to become infrastructure that expands access to knowledge and productive capacity?

Or do we want it primarily to become a mechanism through which a small number of companies collect rents from everyone who uses intelligence?

For Africa, that question is immediate.

Governments should stop treating AI procurement as another software-subscription exercise.

Before buying a proprietary AI service, public institutions should ask whether an open-weight model could perform the same task locally or on regional infrastructure.

We should invest heavily in African datasets, particularly language, health, education, law, agriculture and public administration data, with proper consent, ownership and controls.

Universities should teach people how to build systems around models, evaluate them, fine-tune them and run them efficiently.

Prompt writing is useful, but it is not an industrial policy.

And African governments need much better rules around data extraction.

Data produced here should not automatically become free raw material that leaves the continent, trains somebody else’s model and returns as a service for which Africans pay indefinitely.

Open does not mean harmless

There is one caveat. Open models are not automatically good models. They can carry biases, be insecure and, of course, they can hallucinate.

Some models described as open release their weights but not their training data and others impose licensing restrictions.

Running them still requires hardware, energy, technical skills and governance.

Open AI does not solve Africa’s technological dependence by itself, but it gives us more room to choose what we build.

That is already worth a great deal.

The opportunity hiding inside the bubble

Graylin calls the current story of unlimited AI riches “the AGI mirage”.

Whether the present investment boom eventually qualifies as a financial bubble is something we will know properly only after the fact.

There are clearly speculative elements. There is also genuine technological change.

Those things can exist at the same time.

The larger question raised by Graylin and Gao is who controls this technology, who pays for it, who profits from it and whether increasingly cheap machine intelligence can remain concentrated in the hands of a few companies and countries.

And that brings us back to Africa.

The mistake would be to assume that participating in the AI economy requires us to reproduce Silicon Valley’s economics.

It does not.

Perhaps the extraordinary fall in the price of useful machine intelligence will eventually matter more to Africa than who produces the world’s most capable frontier model.

Africa missed much of the industrial revolution and entered the internet era largely as a consumer of technologies, platforms and intellectual property created somewhere else.

AI gives us another chance to build our unique solutions. We do not need to build the world’s biggest supercar. We need to build the buses that take Africans where we need to go.

Subscribe now

Share

Why doesn’t a car manufacturer simply buy every task from freelancers or independent suppliers every day instead of employing thousands of people?

His answer, published in a paper titled The Nature of the Firm, became one of the foundations of modern economics.

Markets are not free. Finding expertise, negotiating contracts, coordinating work, sharing information and making decisions all carry costs. When those costs become high, it is cheaper to bring people inside one organisation and direct their work through management rather than through constant market transactions.

Coase called these transaction costs.

That’s why firms exist, he argued. A firm grows until the cost of organising one more activity internally equals the cost of buying it from the market. That balance explains why some activities are outsourced while others stay in-house.

For almost ninety years, that idea has helped explain why organisations are structured as they are.

But AI may now be reducing many of the transaction costs that Coase identified, particularly those associated with finding information, coordinating work and transferring knowledge within organisations.

In Coase’s time, these costs included discovering relevant prices, negotiating contracts and organising production through the market. Today, the same economic logic can be applied to finding expertise, locating institutional knowledge and coordinating increasingly complex work across departments.

Institutional knowledge that once disappeared into emails, meetings and documents can now be searched in seconds. AI agents can prepare briefs, coordinate tasks and monitor progress. Information moves more freely when meetings, emails and documents become part of a shared organisational knowledge rather than remaining trapped in individual inboxes.

Taken together, these changes suggest that AI is beginning to affect something much larger than productivity. It is beginning to change how organisations themselves work.

I did not arrive at that conclusion by reading economics papers.

It emerged unexpectedly while following a series of executive discussions produced by Reuters Plus in partnership with KPMG.

The conversations covered energy, healthcare, financial services, insurance, manufacturing and professional services. Each executive described different applications of AI within their industry.

Although the examples differed, many pointed towards the same conclusion. (I am still going through them.)

The picture came into sharp focus when Microsoft shifted its entire enterprise strategy.

Three years ago, enterprise AI discussions revolved around a familiar set of questions:

• Which model performs best?

• How do we write better prompts?

• Will AI replace jobs?

• Can it summarise meetings?

Those questions have not disappeared, but they are no longer at the centre of the conversation inside large organisations.

What struck me was not what the executives discussed, but what they didn't. There was remarkably little debate about which foundation model was best or how to write better prompts. Instead, the conversation revolved around operating models, governance, workflow redesign, leadership, AI literacy and business value. That shift mirrors what is now appearing across much of the enterprise AI landscape.

Microsoft's Frontier Company announcement gave me a framework for understanding it.

Most headlines focused on the scale of the investment: US$2.5 billion and thousands of specialists dedicated to helping organisations adopt AI.

Microsoft is no longer simply selling AI software. It is offering to embed specialists inside customer organisations to redesign workflows, operating models and governance. Those teams will help customers work with whichever foundation models best suit their needs, whether from OpenAI, Anthropic, Microsoft or the growing open-source ecosystem.

The world’s largest software company is investing billions not primarily in selling AI models, but in helping organisations redesign themselves around AI, which means the competitive advantage becomes the organisation, not the model. (Think of electricity. Once it became widely available, every factory could buy it. The winners were not those with electricity, but those that redesigned how they worked to make the most of it.)

The organisation becomes the customer

One theme emerged clearly from the discussions I watched.

The first generation of generative AI focused on individual productivity: write an email, summarise a report, generate a presentation.

The next generation is different. It asks organisational questions instead. How does information move? Where are decisions made? Which processes create friction? How can expertise be shared more effectively?

AI is shifting from a personal assistant to part of the organisation’s operating model. The focus is no longer the individual worker, but the organisation itself.

JPMorgan Chase provides a good example of this shift. The bank has rolled out its internally developed LLM Suite to tens of thousands of employees, giving them AI support for research, drafting, meeting preparation and analysing documents. The objective is not to replace analysts or bankers, but to make the organisation itself more effective at processing and applying knowledge.

As Teresa Heitsenrether, JPMorgan's Chief Data and Analytics Officer, has argued, AI is now a strategic capability for the firm rather than simply another technology project. That is a subtle but important distinction. The investment is no longer in a chatbot. It is in changing how the organisation thinks and works.

AI as a prioritisation engine

One example from the Reuters discussions illustrates this well.

An energy utility described using AI to analyse drone footage across thousands of kilometres of electricity infrastructure.

It would be easy to describe this as automation.

That misses the point.

The AI is not replacing inspectors, but is determining where inspectors should direct their attention, improving judgement rather than simply reducing labour.

The same pattern appears elsewhere.

Banks use AI to identify transactions that deserve investigation, healthcare providers use it to prioritise medical images requiring urgent review and manufacturers analyse equipment data to predict failures before they occur.

Journalists increasingly use AI to sift through thousands of pages of public records, company filings and court documents to identify leads worth pursuing.

Bloomberg offers a glimpse of what this looks like in one of the world's most information-intensive industries. According to CTO Shawn Edwards, the company's AI tools could streamline as much as 80% of an analyst's workload. The objective is not to replace analysts, but to process vast amounts of unstructured information, surface relevant insights and reduce the time spent searching for data, allowing professionals to focus on interpretation, judgement and decision-making.

What changes is the industry. What remains remarkably consistent is AI's role in helping organisations decide where human expertise is most valuable.

Siemens demonstrates that this shift extends well beyond knowledge work. The company is embedding AI directly into engineering, manufacturing and industrial operations, combining AI with digital twins, industrial automation and factory software.

The aim is not simply to automate production, but to improve engineering decisions, anticipate equipment failures and optimise increasingly complex industrial systems. As Siemens President and CEO Roland Busch puts it: "When AI hits the real world, hallucination is not an option."

In other words, industrial AI is valuable only if it produces decisions that can be trusted. That emphasis on judgement rather than automation mirrors what we are seeing in finance, healthcare and professional services.

Fusion teams and disappearing boundaries

Another phrase from the Reuters discussions stood out.

KPMG described the emergence of “fusion teams”, bringing together domain specialists and technologists. Some tax professionals become builders. Others develop AI-enabled workflows.

Programming is becoming another professional literacy.

Lawyers are building agents, scientists are creating analytical workflows and journalists are assembling research applications.

More knowledge workers are creating their own tools instead of waiting. Technical capability begins spreading beyond traditional software teams. That may prove to be one of AI’s most important organisational effects.

From better software to better organisations

The real value comes when organisations connect different units of work together.

For decades, organisations have accumulated specialised departments, each optimised for its own function.

The cost has always been coordination. Generative AI changes that equation and lowers the cost of sharing knowledge across organisational boundaries.

As foundation models become more widely available, large organisations are likely to have access to comparable capabilities. That shifts attention to a different set of questions:

• Which organisation learns faster?

• Which organisation shares knowledge more effectively?

• Which organisation reduces friction between departments?

• Which organisation adapts its operating model rather than simply adding another piece of software?

Those questions are becoming more important than benchmark scores.

Back to Ronald Coase

Coase’s insight remains sound. Organisations exist because they reduce transaction costs, but the question is whether AI is changing the nature of those costs.

Finding expertise is becoming cheaper, sharing knowledge is becoming cheaper, coordinating work is becoming cheaper, documenting decisions is becoming cheaper and negotiating tasks is becoming cheaper.

None of this eliminates the need for organisations but it may, however, change how they are designed.

The Industrial Revolution reduced the cost of physical labour. The Information Age reduced the cost of communication.

Generative AI is reducing the cost of intelligence, but public attention continues to focus on more powerful chatbots because they are visible. Inside large organisations, however, the conversation has already moved on.

If Ronald Coase were writing today, I suspect he would still recognise his theory. But he might ask another simple question:

What happens when the cost of coordination falls again?

Subscribe now

Share

In my last blog, I argued that artificial intelligence could become a leapfrog technology for countries like South Africa because it fundamentally changes the economics of expertise. That article asked what AI does. This one asks why, and what it means for one profession in particular.

Since then, I have spent time catching up after a few exasperating months of bootstrapping a business. I expected to find another round of faster models, smarter chatbots, and increasingly capable AI agents and that is indeed what I found.

More capable models. Better reasoning. AI agents. Coding assistants. Scientific research tools. Record investment in chips and data centres. Viewed individually, they looked like another series of product launches.

What surprised me was something else entirely. While I was distracted by the day-to-day running of a business, the underlying currents of the industry had dramatically shifted.

The chatbot was never the product. It was the demonstration.

For the past three years, public attention has focused on chatbots. We compared models, debated hallucinations, and marvelled at increasingly natural conversations. The chatbot became the public symbol of artificial intelligence.

As argued in my previous blog, the real product is intelligence itself becoming progressively cheaper, much as electricity made energy cheaper during the industrial age.

Every major technological revolution has reduced the cost of something that was once scarce. Steam reduced the cost of mechanical power. Electricity reduced the cost of energy. The internet reduced the cost of communication and information.

Artificial intelligence appears to be reducing the marginal cost of expertise. That is an entirely different proposition from simply making software better. For decades, software has automated routine tasks. Spreadsheets replaced ledgers; email replaced letters; databases replaced filing cabinets. AI is different. Rather than simply automating tasks, it reduces the cost of applying knowledge itself.

Drafting documents, analysing contracts, writing software, reviewing regulations, preparing reports, and synthesising research all become dramatically cheaper.

In this scenario, experts do not disappear, but their value shifts. The scarce resource increasingly becomes judgement: deciding which problems matter, interpreting uncertain results, and accepting responsibility for decisions. That shift helps explain why so many announcements during the first half of 2026 suddenly seem connected.

From conversation to infrastructure

The most important trend this year is not that AI has become better at conversation. It is that it is moving beyond conversation altogether.

Instead of asking AI to answer questions, organisations increasingly expect it to complete meaningful work: analysing thousands of documents, building software, coordinating workflows, and carrying out complex research. In the industry, these are called agentic systems, because they can act across several steps rather than respond to one instruction at a time. The Reuters Institute’s 2026 survey of 280 media leaders in 51 countries found that 75 per cent of surveyed media leaders expected agentic tools to have a large effect on the news industry.

That changes the role of people. Humans increasingly define the problem, provide context, and evaluate outcomes. The AI performs more of the execution.

This is why the race to build specialised chips, data centres, and computing infrastructure matters. The largest AI companies are investing billions because intelligence is becoming an economic resource whose cost can be reduced, distributed, and scaled.

But AI infrastructure is not limited to chips and data centres. It also includes the systems that organise information, connect related material and restrict AI to trusted sources. These systems determine what the AI can find, what it can use and how reliably it can work.

Once they are in place, they change how the whole organisation operates. Electricity did not simply make factories cheaper; it changed how factories were designed. The internet did not simply make communication faster; it changed how businesses were organised.

AI appears to be following the same path.

Journalism as a case study

Journalism offers a vivid illustration of where this is heading, and why it matters. Most of the public debate has concentrated on whether AI can write an article. It can. That question has become less useful with every improvement in the technology. The more consequential change is happening elsewhere.

Since the release of ChatGPT in late 2022, journalism has treated generative AI mainly as a writing machine. Can it produce a news report? Should an AI-written article carry a label? Will it replace sub-editors, reporters, or columnists? These questions are understandable because writing is the visible product of journalism.

Readers see the headline, paragraphs, pictures, and byline. They do not see the hours spent searching records, comparing documents, calling sources, discarding weak leads, and testing explanations. The visible product has therefore dominated the argument. Newsroom adoption, however, is moving in another direction.

The Reuters study found that 97 per cent considered back-end automation an important use of AI. Newsgathering applications were regarded as important by 82 per cent. Three-quarters expected AI agents to have a large effect on the news industry.

This is no longer mainly about generating prose. It is about connecting AI to archives, databases, content-management systems, research files, audio, video, audience products, and distribution channels.

The chatbot introduced journalists to the technology. The newsroom system may be the real product.

What the competition entries revealed

I recently acted as a judge in the World Association of News Publishers’ Digital Media Awards 2026.

The entries I reviewed as a judge cannot be treated as a representative survey of global journalism. Organisations choose to enter competitions, present their strongest work, and answer the questions put to them. The competition did not always request detailed evidence of internal AI controls, correction records, or technical audits. Where that material was absent, I could not assume that the underlying practices were absent too.

Even with those limits, the entries revealed a clear pattern. Most serious projects were not trying to create an artificial reporter. They were trying to solve practical newsroom problems.

Some processed large bodies of documents that reporters could not examine manually within the available time. Some recovered material from archives and made it searchable. Others expanded routine company coverage, supported multilingual publishing, organised election information, extracted data from records, or converted existing journalism into more accessible formats.

In several cases, AI sat inside an existing editorial process. It gathered or transformed material, produced a provisional output, and passed that output to a journalist or editor.

The strongest entries had bounded tasks. The system knew where its information came from. The newsroom knew what the system was meant to do. A person retained authority over publication.

The weaker entries often made broad claims about innovation without showing whether the technology produced better journalism, better understanding, or more reliable information.

That became one of the most persistent lessons of the judging process: speed and volume are easy to demonstrate. Editorial value is harder. A newsroom can show that it produced twice as many summaries or reduced a task from three hours to twenty minutes. It is much harder to prove that the resulting journalism was more accurate, reached neglected audiences, revealed something new, or led to better public understanding.

AI makes production measurable. Journalism must still measure meaning.

The article is only one stage of journalism

It helps to separate journalistic work into three broad areas:

• There is evidence gathering: interviews, observation, documents, datasets, public records, and reporting from the scene.

• There is analysis: deciding what the evidence means, identifying patterns, testing explanations, and working out what remains unknown.

• Then there is expression: writing, editing, illustrating, packaging, and distributing the result.

Generative AI is already capable in parts of the second and third areas. It can compare documents, summarise evidence, create timelines, classify material, find repeated names, translate text, propose explanations, and produce draft language. It can assist with evidence gathering by monitoring websites, scraping public information, transcribing interviews, and searching large collections. Research with investigative journalists has identified monitoring, web scraping, summarisation, and early data exploration as areas where automation could remove months of repetitive work.

But AI does not stand outside a municipal office waiting for an official who has avoided calls for a week. It does not persuade a frightened source to speak. It cannot observe the mood in a community meeting, recognise that an answer has been carefully rehearsed, or understand why a small inconsistency matters to people who live in a particular place.

AI can process evidence. Journalists still have to obtain it, judge it, and take responsibility for what they say it proves.

My own experiment has been in analysis

I have used AI for analysis rather than for reportage. Reportage begins with contact with the world. Analysis begins when evidence exists but its meaning remains unsettled. This is where I have found AI most useful.

I rarely begin by asking it to write a finished article. I begin with a question.

What am I really trying to establish? What assumptions am I making? What would a sceptical reader challenge? What alternative explanation fits the same facts? What evidence would weaken the argument? What have I failed to ask?

The model can generate possible lines of inquiry much faster than I can pursue them. Most are ordinary. Some are irrelevant. Occasionally one exposes an assumption that has been holding the entire argument together. That is where human judgement enters. The purpose is not to accept the model’s answer. It is to enlarge the field of questions from which I can choose.

Once I have gathered reliable sources, I use AI to compare them, identify agreements and contradictions, construct chronologies, and separate established facts from interpretation. I ask it to produce the strongest case against my argument. I ask where I have confused correlation with causation. I ask which paragraph contributes nothing. I ask the question that has killed more of my draft ideas than any other: So what?

The process is recursive. A possible argument leads to research. Research changes the argument. The revised argument exposes missing evidence. That sends me back to the sources. The resulting draft then undergoes another round of challenge, verification, and restructuring.

AI has not removed me from the writing process. It has made the process less linear and more argumentative.

(I have written a separate methodology note explaining how I use AI in research, argument development, verification, and drafting, and where I believe the limits lie. It also explains why I call this project The Cyborg Chronicles.)

AI can change the order of methodology

Traditional journalism often follows a familiar sequence. A reporter identifies a subject, gathers information, decides what it means, writes the story, and submits it for editing. Real reporting has never been quite that neat. New evidence changes the question. Interviews send a reporter down unexpected paths. Editors identify gaps that require more calls.

AI makes this iterative character more pronounced.

A journalist can now test several possible hypotheses before committing substantial reporting time. A large set of documents can be explored early enough to reveal which questions deserve attention. A provisional argument can be challenged before it hardens into a finished draft.

The emerging sequence may look more like this: a journalist begins with a problem. AI assists in mapping possible questions, assumptions, and lines of inquiry. The journalist chooses which ones deserve investigation. Reporting produces evidence. AI assists in organising and interrogating that evidence. The gaps become the basis for further reporting. Writing happens throughout the process rather than only at the end.

This reverses part of the old logic. Analysis no longer has to wait until all the reporting has been completed. It can guide the reporting from an earlier stage.

That does not mean allowing a model to decide what the story is. It means using the model to widen the range of questions before human judgement narrows it again. This may be one of AI’s deepest contributions to journalism—and to knowledge work more broadly. It can reduce the cost of intellectual exploration.

The danger is confirmation bias at machine speed

The same capability can produce bad journalism more efficiently.

A model is exceptionally good at making an argument sound coherent. Give it a weak premise and it can construct a polished case around it. Ask leading questions and it will often follow the direction provided. A journalist who wants confirmation can obtain it in seconds.

This is why prompting skill is not enough. The journalist needs a method designed to resist easy agreement.

I separate model output from evidence. I do not treat a generated statement as a fact merely because it sounds plausible. For factual work, I restrict the task, identify the sources, and require citations. I check those citations against the original material. I deliberately ask for disagreement. I ask the model to identify the weakest inference, propose rival explanations, and state what cannot be established.

The final decision remains mine because the final accountability remains mine.

This is more than a precaution against hallucinations. It protects the journalist’s independence from the model’s tendency to turn ambiguity into smooth prose. The risk is not simply that AI will invent something. The subtler risk is that it will make an uncertain idea feel settled.

The emerging newsroom as a knowledge system

The judging process showed that the most promising projects were moving beyond the article as a self-contained object. They treated journalism as a body of connected knowledge.

A continuing story may contain interview transcripts, source documents, previous reports, photographs, datasets, timelines, legal records, corrections, and editorial decisions. Traditionally, much of this material has been scattered across inboxes, personal drives, notebooks, and content-management systems. It becomes difficult to search, easy to lose, and almost impossible to use as a coherent institutional record.

AI can change that.

Instead of beginning every assignment with a blank page, a journalist could work inside a source-bounded environment containing the verified material for a beat or investigation. The system could monitor new documents, compare them with previous versions, update a chronology, and identify contradictions or gaps.

A municipal reporter could receive an alert when a procurement item reappears under a different description. A court reporter could search years of judgments, filings, and previous coverage while retaining links to every original source. An investigative team could question thousands of documents without uploading confidential material to a public service.

The Associated Press has already begun describing AI in these broader terms: newsroom workflow tools, research support, Stylebook guidance, training, and systems embedded in production rather than a standalone writing product.

This points towards a different newsroom architecture. The AI would not become the reporter. It would become part researcher, part librarian, part production assistant, and part institutional memory.

The practical change is easy to miss. AI reduces the time required to establish what is already known. That allows the reporter to concentrate on what remains unknown.

A system may identify that a budget line has changed, that two official accounts conflict, or that a company name has appeared repeatedly across tender records. The journalist still has to decide whether the pattern matters, contact the people involved, visit the location, obtain missing records, and establish what can responsibly be said.

AI processes the available evidence while journalism expands the evidence.

South African journalism has a different opportunity

Public evidence about AI use inside South African newsrooms remains thin. I could not find a credible, current national survey showing how many journalists use generative AI, which tools they use, or whether adoption is led by individuals, editors, or media companies. That absence matters. International studies may describe common patterns, but we should not assume that South African newsrooms have adopted AI at the same rate or in the same way.

It does not mean there is little activity. Several South African projects appeared among the competition entries I reviewed. They showed that local news organisations are already experimenting with AI in practical newsroom settings. Because some of the material was submitted for judging rather than public release, I cannot identify every project or disclose internal details. Nor can a small, self-selected group of entries be treated as representative of the whole sector.

What the entries do show is that South African experimentation is real. It is emerging around familiar pressures: limited newsroom capacity, large bodies of material, multilingual audiences, production demands, and the need to extend the reach of existing journalism.

These are precisely the conditions under which cheaper analytical capacity could matter most. A small newsroom may not be able to employ a data team, archive specialist, transcription service, and video unit. It may still be able to use AI to search municipal budgets, transcribe meetings, compare public records, organise years of reporting, and convert one piece of original journalism into several formats.

The opportunity is the ability to decide which neglected journalistic tasks have become affordable.

What can also be seen publicly is a media sector confronting AI from several directions at once. South African journalists are encountering tools that change research, production, and distribution. News organisations also face AI-generated misinformation, synthetic content, and answer engines that reuse their work without necessarily sending readers to the original publisher.

News24’s exposure of fictitious references in South Africa’s draft national AI policy offered a sharp illustration of journalism’s role in this environment. This is the future of AI and journalism in miniature. The technology makes the production of authoritative-looking material easier. Journalism becomes more valuable when it establishes whether that authority is deserved.

Local languages cannot remain an afterthought

South African newsrooms should not simply copy the AI strategies of publishers in Europe and North America. Our needs are different.

Many local and community publications operate with small teams covering large areas and complicated institutions. Reporters face council agendas, tender documents, court records, government notices, company filings, audio recordings, and public datasets that cannot all receive close attention. The country also has twelve official languages, uneven digitisation, and large differences in audience access.

AI can assist with transcription, document comparison, translation, archive retrieval, subtitling, and conversion into audio, video, or simpler language.

The competition entries offered a glimpse of what this could mean. Some of the most imaginative projects treated regional-language journalism as a body of knowledge that had been difficult to search and reuse. They used AI to recover archives, improve classification, and allow questions in one language to retrieve material held in another.

That is more ambitious than translation alone. It suggests that a reporter should eventually be able to ask a question in English and find relevant material published years earlier in isiXhosa, isiZulu, or Sesotho. A radio interview could be transcribed, translated, and linked to the original recording. An archive could become searchable across languages rather than remain divided into separate collections.

A reported story could then become a text article, short video, audio version, timeline, and several language editions without requiring a separate production team for each format. The original reporting remains the expensive part. AI lowers the cost of extending its reach.

For community media, the greater opportunity may be analytical capacity. A small publication could examine years of municipal budgets, monitor public notices, or search meeting records in ways once reserved for newsrooms with specialist teams.

This would not solve the funding crisis facing South African journalism. It could allow limited editorial resources to be directed towards work that requires human presence, source development, and local knowledge.

The risks are equally real. A weak translation can distort meaning without producing an obvious grammatical error. A system trained mainly on material from outside South Africa may misunderstand political terminology, historical references, names, dialects, and local institutions.

Newsrooms will therefore have to test systems in the languages, beats, and document types in which they will actually operate. A general claim of high accuracy tells an editor little about whether an application can interpret an isiXhosa council recording, distinguish between similarly named public bodies, or recognise the political meaning of a phrase in a particular community.

Local knowledge is not an optional layer added after the technology has done its work. It is one of the controls that makes the work usable.

There may now be money for experimentation

South Africa’s Competition Commission announced a R688 million package from Google and YouTube in November 2025 following its Media and Digital Platforms Market Inquiry. The package is intended to support national, community, and non-English media through licensing, innovation grants, and capacity-building programmes. Other remedies include publisher support and monetisation measures involving Meta, Microsoft, TikTok, and X.

This could give South African media organisations room to experiment with AI and other digital products. The way that money is spent will matter.

Funding dozens of disconnected demonstrations may produce attractive presentations and little institutional change. A tool may work during a funded pilot and disappear when the grant ends because it was never connected to the newsroom’s ordinary work.

The more useful starting point is a specific editorial constraint:

• A community publisher may need to monitor 10 municipalities with three reporters.

• A radio station may need searchable transcripts of years of interviews.

• A regional publication may need to make its archive available across languages.

• A court desk may need to connect judgments, filings, and previous reporting.

• A national newsroom may need a shared chronology for a long-running corruption case or commission of inquiry.

These are practical problems with identifiable users, sources, and outcomes. The test should not be whether a newsroom has launched an AI product. It should be whether the product allows the newsroom to do journalism it could not previously afford to do, and whether it can still maintain that capacity when the initial funding ends.

AI is also changing how audiences reach journalism

News organisations face a second upheaval. AI is becoming an intermediary between journalism and the audience. Search engines once directed users towards publishers. Answer engines increasingly extract, combine, and summarise material within the search result or chatbot response.

Media leaders surveyed by the Reuters Institute expect search referrals to decline sharply over the next three years. The same research found a substantial fall in Google search traffic to thousands of news sites between late 2024 and late 2025, although the proportion caused directly by AI summaries remains uncertain.

A June 2026 report from FIPP, the global magazine-media network now part of WAN-IFRA, argues that the platform era is ending as search referrals decline and AI services fail to replace the traffic publishers are losing.

This changes the economics of journalism. A newsroom pays to send a reporter to a meeting, develop sources, obtain documents, verify claims, and publish a story. An AI service can compress that work into a paragraph without transferring the reader, relationship, or revenue to the publisher.

The production opportunity and distribution threat come from the same technology. Newsrooms will use AI to produce, analyse, and distribute journalism. AI platforms will use journalism to produce answers.

This makes the distinction between generic information and original reporting commercially important. AI is well suited to material that already exists: summaries of announcements, routine results, standard explainers, and combinations of published sources. The easier this material becomes to generate, the harder it becomes for publishers to charge for it or use it to attract readers.

Original evidence becomes more valuable because it does not exist until somebody obtains it. That includes interviews, observations, leaked documents, exclusive data, specialist interpretation, and reporting from places where no automated system is present.

The future value of a newsroom may therefore depend less on how much information it can publish and more on how much new knowledge it can produce.

The article becomes one output among several

This also changes what happens after the reporting has been completed. The article remains important, but it may no longer be the only or even the main expression of the work.

The same verified reporting could support:

• A conventional article;

• An audio version;

• A short video;

• A timeline;

• A searchable topic page;

• A question-and-answer tool;

• A simplified version;

• Translations for different language audiences;

• Updates as new evidence appears.

This does not mean producing more material for its own sake. It means separating the cost of gathering reliable information from the cost of making that information accessible in different forms. Newsrooms have traditionally repeated large parts of the production process for every platform. AI can reduce that duplication.

The editorial question becomes whether each version remains faithful to the reporting and appropriate for its audience.

The role of the reporter changes

The reporter of the near future may begin the day with a system that has monitored a beat overnight. It may have collected new court filings, council documents, tender notices, and official statements. It may have compared them with previous records and identified changes.

The reporter decides which changes matter.

During an investigation, the system may organise evidence, produce a chronology, identify missing documents, and test whether different accounts are consistent. The reporter pursues the gaps.

During writing, AI may challenge the structure, identify unsupported claims, and suggest alternative interpretations. The reporter decides what the evidence permits.

After publication, the same source material may support different formats and future updates.

The reporter remains responsible for the inquiry. The technology changes the volume of material that can be examined and the speed at which it can be organised. This moves the journalist’s value further away from routine information processing and towards question selection, source development, interpretation, and accountability.

The role of the editor changes

Editors may also spend less time moving copy through a production line and more time deciding where limited reporting attention should be directed. When AI can generate dozens of possible leads, explanations, and formats, selection becomes more important.

• Which apparent pattern deserves another day of reporting?

• Which automated finding is an artefact of incomplete data?

• Which story serves the public rather than the production target?

• Which output requires legal or specialist review?

• Which process should not be automated at all?

The editor becomes partly responsible for the design of the reporting system, rather than only the quality of the finished article. This requires a working understanding of where information comes from, how systems transform it, and where human decisions remain necessary.

There is a risk of professional deskilling

The productivity argument can hide a long-term danger. Writing is not merely the stage at which journalists communicate what they have discovered. It is often where they discover whether they understand it.

A reporter may realise halfway through a paragraph that two facts do not fit together. The effort of explaining a sequence can expose a missing interview or an unsupported assumption. If AI takes over too much of the drafting process, younger reporters may lose part of the apprenticeship through which judgement develops.

The same applies to research. A journalist who receives an instant summary may never encounter the awkward footnote, contradictory table, or strange phrase that opens a new line of inquiry. Efficiency can remove friction. Some friction is where journalism happens.

Newsrooms therefore need to distinguish between work that consumes time without adding judgement and work through which judgement is formed.

• Transcribing a two-hour interview by hand may add little. Reading the transcript closely may reveal everything.

• Formatting a table may be mechanical. Deciding which figures belong together is analytical.

• Producing a rough summary may be delegated. Deciding what the summary leaves out cannot be.

Human oversight is too vague

Almost every responsible AI project claims to keep a human in the loop. The phrase has become close to meaningless.

A journalist who understands the source material, reviews every inference, and decides whether to publish exercises real authority. An editor who glances at fifty machine-generated articles before a deadline may provide little protection.

The question is not whether a human appears somewhere in the process. It is whether that person has enough knowledge, time, power, and contextual data to change the outcome. Editorial control should therefore be visible in the design of the work.

Someone should know:

• Which sources were used;

• What the system produced or altered;

• What was checked;

• What remained uncertain;

• Who approved publication;

• How an error can be reconstructed and corrected.

This does not require turning every small experiment into a compliance exercise. It requires preserving the conditions under which journalism can answer for itself.

The real measure is journalistic attention

AI vendors sell speed. Newsrooms often report time saved. Both matter, but neither tells us whether journalism improved. A more useful measure is journalistic attention.

• Did the technology allow reporters to examine more documents?

• Did it bring scrutiny to an institution that had received little coverage?

• Did it expose a contradiction that would otherwise have been missed?

• Did it give a journalist more time to speak to people, visit a place, or verify a claim?

• Did it make an archive usable?

• Did it bring reporting to an audience in a language or format previously neglected?

• Did it improve the newsroom’s ability to remember what it already knew?

These questions connect technological performance to editorial purpose. The strongest AI project is not necessarily the one that publishes the most material. It may be the one that allows a small team to discover and verify one important fact that nobody else had found.

The next divide will be institutional

Most newsrooms will eventually have access to similar underlying models. The difference will lie in how they organise around them.

One newsroom may use AI to produce more low-cost material from press releases and existing articles. Another may use the same technology to monitor institutions, interrogate records, maintain long-running investigations, and serve audiences across several languages and formats.

The model may be identical but the journalism will not be.

This is why AI adoption cannot remain an informal collection of personal experiments. Individual experimentation is often where learning begins. Institutional value appears when useful practices become shared, tested, and connected to ordinary newsroom work.

That requires training, source rules, clear editorial authority, correction procedures, and decisions about which knowledge should remain inside the organisation.

It also requires management to decide where savings should go. If AI reduces the time spent on transcription, routine summaries, and production work, does the newsroom use that capacity to cut staff? Or does it redirect the time towards reporting, source development, and specialist knowledge?

A newsroom that cuts reporting and fills the gap with generated material may publish more while knowing less. A newsroom that uses cheaper analysis to increase its contact with the world may become stronger.

Technology does not solve poor management

AI will not compensate for weak leadership, confused workflows, or unethical behaviour. Bad organisations can automate bad decisions as efficiently as good ones can improve sound ones.

Nor is access to a model enough. For AI to work consistently, newsrooms need reliable information. Archives must be searchable. Documents must be retained. Data must be organised. Corrections must remain attached to the original record.

Many organisations will discover that cleaning decades of fragmented files and disconnected systems is harder than adopting the AI itself.

The organisations that gain most will not necessarily be those buying the most expensive technology. They will be those that understand their own work, preserve reliable information, and know where professional judgement remains indispensable.

Technology changes what is possible. Management decides what becomes normal.

The real story

I suspect the first half of 2026 will be remembered mainly as the point at which the industry began moving from conversation to execution.

The chatbot introduced millions of people to AI. What follows is the more difficult stage: reorganising institutions around the falling cost of cognitive work. For journalism, this means deciding what becomes valuable when summaries, drafts, translations, classifications, and routine analysis become cheap.

The answer is beginning to emerge. As AI makes routine analysis, drafting, and information processing cheaper, original evidence becomes more valuable because it cannot be generated from existing material alone. The ability to ask the right questions also becomes more important, particularly when plausible answers are easy to produce.

Local presence, specialist knowledge, and trusted sources gain value for the same reason. They give journalists access to information, context, and relationships that AI cannot reproduce on its own. Institutional memory also becomes more important, because newsrooms that can connect current events with years of verified reporting will have an advantage over those relying only on public information.

Editorial judgement and public accountability remain central. Someone must still decide which questions matter, what the evidence supports, and whether the work is ready to publish. Someone must also take responsibility when it is wrong.

The practical question for newsrooms is therefore no longer whether they should use AI. Journalists are already using it, formally or informally. The real question is how that use should be structurally organised, where it should sit in the reporting process, and what the newsroom should do with the capacity that remains.

Does it allow reporters to examine more of the world? Does it expand the range of people and institutions subjected to scrutiny? Does it make existing journalism available to audiences that could not previously reach it? Does it strengthen the newsroom’s knowledge, or merely increase its output?

The best answer will come from newsroom practice, not from a model benchmark or policy statement. AI will expose which organisations understand the difference between producing content and producing journalism.

Those that use it mainly to fill pages and feeds may discover that the world has no shortage of fluent text. Those that use it to examine more evidence, ask better questions, and direct scarce human attention towards work that matters may become stronger.

This is the conclusion I draw from judging AI journalism projects and from two years of working with the technology myself. We have spent the past three years asking what AI can do. The more useful question is what journalists should do differently now that it can do it.

AI’s deepest effect on journalism may be that it changes how we find the story, what we can afford to investigate, and where human judgement matters most.

Further reading

• Sam Altman, The Gentle Singularity — https://blog.samaltman.com/the-gentle-singularity

• Sam Altman, The Intelligence Age —

https://ia.samaltman.com/

• OpenAI, Introducing the Intelligence Age — https://openai.com/global-affairs/introducing-the-intelligence-age/

• World Economic Forum, The AI-First Operating System: A Blueprint for Operating and Business Model Innovation — https://www.weforum.org/publications/the-ai-first-operating-system-a-blueprint-for-operating-and-business-model-innovation/

• World Economic Forum, Organizational Transformation in the Age of AI: How Organisations Maximize AI’s Potential — https://www.weforum.org/publications/organizational-transformation-in-the-age-of-ai-how-organizations-maximize-ais-potential/

• Reuters Institute, Journalism, Media, and Technology Trends and Predictions 2026 — https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026

Subscribe now

Share

It’s a familiar conversation: people want to talk to real people, the machine doesn’t understand people, and so on.

I found myself trotting out my standard response: AI is very bad at customer relations now, but soon it won’t be. They word is now.

The mistake we keep making is assuming that today’s AI is a good guide to tomorrow’s AI. Unlike most technologies, AI isn't improving in steady, incremental steps. Its capabilities are compounding at extraordinary speed. Judging tomorrow's customer experience by today's shortcomings is therefore a bad idea.

At the same time, we assume today’s customers are a good guide to tomorrow’s customers. They aren’t either.

Soon the dominant consumers will be people who have grown up communicating through apps rather than call centres. After them will come a generation that has never known anything but AI.

As those two trends collide, the fundamental flaw may not be the technology itself, but how enterprises are designed around it.

Which brings me to a recent World Economic Forum article, 5 ways for AI-first enterprises to rebuild around intelligence. It argues that organisations need to stop treating artificial intelligence as another software tool and start rebuilding themselves around intelligence itself.

The article asks how established enterprises should adapt to abundant intelligence and, ultimately, how a corporate balance sheet adapts to AI. That is a useful question, but it is not the most urgent one.

A better question is how an under-resourced economy can use AI to overcome longstanding developmental constraints, because AI fundamentally changes the economics of expertise.

The Lesson from Mobile Phones

Developing countries have often been told that economic progress follows a fixed sequence:

• Build the infrastructure.

• Develop the institutions.

• Accumulate expertise.

• Expand capacity.

• Then compete with wealthier economies.

History suggests otherwise.

Large parts of Africa bypassed extensive fixed-line telephone networks and moved directly to mobile communications. Mobile money platforms achieved widespread adoption in environments where conventional banking infrastructure remained limited. Rather than following the same path taken by Europe or North America, many developing economies skipped entire stages of technological development.

Artificial intelligence offers a similar opportunity.

South Africa, for instance enters the AI era carrying constraints that are familiar: municipal skills shortages, weak state capacity, uneven educational outcomes and a large population of small businesses operating with limited access to professional services. These weaknesses are often discussed as barriers to growth. They may also represent areas where abundant, low-cost intelligence can have the greatest impact.

These are intelligence and capability bottlenecks, not primarily technology problems.

The Economics of Capability Have Changed

The WEF article asks organisations to imagine a world where intelligence is abundant, as described recently by Open AI CEO Sam Altman.

For most of human history, analytical work was expensive because it depended on scarce, highly trained professionals. AI does not replace them, but it allows organisations to perform parts of that work at almost no marginal cost. That fundamentally changes the economics of expertise.

A municipal official can use AI tools to analyse procurement records, compare spending patterns and flag transactions that warrant closer investigation.

A small business owner can develop an export strategy informed by international market data.

A clinic manager can use predictive demand-planning tools to anticipate supply shortages and improve inventory management.

A non-profit organisation can evaluate community survey results and draft donor proposals without engaging an external consultancy.

Why Resource-Constrained Institutions Stand to Gain Most

Much of the global conversation about AI assumes that the largest benefits will accrue to the world’s largest corporations. There are obvious reasons for this assumption: capital, data and technical resources.

But the organisations that stand to gain the most from abundant intelligence are not those optimising mature operations, but those that have historically been starved of expertise.

Consider a small municipality with chronic skills shortages. A multinational corporation might use AI to improve an already sophisticated operation by 15%. A municipality that previously lacked analytical capacity entirely could use the same technology to establish a baseline of compliance monitoring, financial oversight and service-delivery analysis.

For the corporation, the gain is incremental. For the municipality, it can be transformative.

The same applies across the economy. A large company uses AI to refine an existing corporate strategy. A township entrepreneur gains access to strategic planning, financial forecasting and market intelligence that would previously have been unaffordable.

This capability is the essence of leapfrogging: the greatest impact occurs where the starting point is lowest.

A Global Race with an African Opportunity

Investment programmes worth hundreds of billions of dollars, restrictions on advanced semiconductor exports and competition over compute infrastructure all point to one conclusion: artificial intelligence has become a matter of national capability. Around the world, governments increasingly describe AI in strategic terms because they regard it as central to future economic competitiveness and national security.

For countries whose greatest constraint is limited access to expertise rather than limited access to technology, AI represents an opportunity to expand institutional capability, strengthen public administration and accelerate economic development.

Africa is unlikely to outspend the United States or China on AI infrastructure. It does not need to build the world’s largest data centres or produce the most advanced semiconductors. But this does not mean infrastructure is irrelevant.

Leapfrogging worked for mobile phones because cell towers are cheaper and faster to deploy than copper landlines. AI is different. Running large language models at scale requires reliable electricity, stable internet connectivity and, critically, compute capacity that, for now, remains heavily concentrated in a handful of foreign cloud providers.

The continent’s opportunity lies elsewhere: not in winning a race to build the biggest supercomputer, but in becoming the world’s most sophisticated user of openly available and open-source models.

That approach creates a dependency. If African governments and businesses rely entirely on foreign AI platforms, they risk becoming consumers of foreign intellectual property rather than participants in the intelligence economy. The strategic question is not whether Africa can out-build the United States, but whether it can adapt and deploy AI in ways that preserve data sovereignty and local relevance.

That question leads to another, more uncomfortable one: even if the infrastructure were in place, would the institutions be ready?

From AI-First Enterprises to AI-First Development

For decades, policymakers, economists and development practitioners have wrestled with the challenge of overcoming shortages of skills, expertise and institutional capacity. Artificial intelligence does not solve these structural problems automatically. It does, however, alter the economics of capability in ways that would have been difficult to imagine only a few years ago.

But there is a risk in assuming that under-resourced institutions will adopt new technology quickly simply because they have the most to gain. Historically, the opposite is true. Weak institutions tend to adopt new tools slowly, not rapidly (see paper embedded below. A municipal official who is overworked, lacks digital literacy and works within a bureaucracy that actively resists change is unlikely to become an AI power-user overnight, no matter how powerful the underlying models. The bottleneck is not merely access to intelligence; it is what economists call absorptive capacity, the ability of an organisation to identify, acquire, adapt and apply new knowledge (Cohen & Levinthal, 1990).

This is where the leapfrog analogy requires caution. Mobile phones were adopted rapidly because they required minimal training and delivered immediate, tangible value at the individual level. AI, by contrast, demands new workflows, new skills and, often, new forms of trust. The organisations that stand to gain most from abundant intelligence are not necessarily those that will adopt it fastest. They are those that invest simultaneously in digital literacy, change management and institutional reform. Technology alone is never enough; culture and capability must move with it.

Africa should not view AI merely as another technology trend, nor solely as a corporate productivity tool. It should be viewed as a national leapfrog opportunity.

But the real risk is not that Africa adopts AI badly.

It is that much of the continent adopts it too slowly.

While governments elsewhere are investing hundreds of billions of dollars in compute infrastructure, AI research, education and public-sector adoption, much of Africa is still treating AI as an interesting technology rather than an economic strategy. Policy debates continue while other economies are redesigning education, public administration and industry around AI.

Leapfrogging is never automatic. It demands deliberate investment, political leadership and institutions capable of change. The real danger for Africa is that by the time it decides to act, the frontier will already have moved several years ahead.

The debate over whether AI is overhyped is dead; the rest of the world has already moved past the question. The only question that now matters is whether Africa intends to build its own capability, or spend the next decade importing everyone else's.

Subscribe now

Share

After years of corruption, mismanagement and financial scandals, I have become so accustomed to hearing about billions lost, wasted or written off that I no longer instinctively ask what happened to the money, who was responsible, or whether anyone will be held accountable.

The Auditor-General’s most recent report found that only 41 of the country’s 257 municipalities achieved clean audits in 2023-24, while 206 municipalities materially failed to comply with key legislation. Municipalities incurred R87.03 billion in irregular expenditure over the first three years of the current administration, and 214 municipalities had findings relating to procurement and contract management.

The Auditor-General’s conclusion: local government remains characterised by governance failures, inadequate institutional capability, and a lack of accountability and consequences.

With local government elections approaching, voters will once again be asked to place their trust in candidates promising better services, cleaner administration and greater accountability. Yet every election cycle tends to revolve around the same questions: Why do municipalities struggle to detect problems sooner? Why do so many failures only come to light years later? And why does accountability seem so elusive?

The Johannesburg disclosure prompts a broader question: If artificial intelligence can analyse millions of transactions, search vast collections of documents and identify unusual patterns in complex datasets, could it help municipalities identify problems earlier, improve transparency and reduce waste?

The answer is not straightforward. Artificial intelligence cannot solve corruption, create political will, replace competent administration or compensate for weak institutions. Yet dismissing it entirely would overlook a more important issue. South Africa’s municipalities increasingly face an information-management challenge that traditional administrative systems struggle to address.

The Developmental State and the Information Problem

Understanding where AI may fit begins with understanding the nature of the South African state itself. South Africa has consciously adopted a developmental approach to government.

Unlike states whose primary role is limited to protecting property rights, enforcing contracts and maintaining public order, South Africa expects government to play an active role in economic and social development.

At a local government level, this means municipalities are expected to provide water, sanitation, electricity, roads, refuse removal, environmental services, local economic development initiatives and a range of community services.

These responsibilities generate enormous amounts of information. Every procurement process, infrastructure project, maintenance programme, service request, contractor appointment, budget allocation and council decision creates records, reports and data. The larger government’s developmental role becomes, the greater the volume of information that must be managed.

This creates what economists call an information problem. No municipal manager can personally monitor every project, supplier, invoice, contract variation or service delivery complaint. As organisations become larger and more complex, information becomes dispersed across departments, systems and individuals.

When information becomes difficult to monitor, opportunities emerge for inefficiency, waste, poor decision-making and, in some cases, corruption. The larger and more complex government becomes, the harder it becomes for citizens to know whether those resources are being used effectively. Much of modern public administration can be understood as an attempt to reduce this information gap. From this perspective, AI becomes an information-processing tool.

Subscribe now

The R45 Billion Question

Take that R45 billion figure from Johannesburg. It represents five years of expenditure classified as irregular, fruitless or wasteful. But irregular expenditure does not necessarily mean money was stolen. In many cases, it reflects failures to follow procurement rules, incomplete documentation or procedural non-compliance. Establishing whether actual financial loss occurred often requires lengthy investigation.

Had an AI system been monitoring Johannesburg’s financial data in real time over that five-year period, it could have flagged anomalous transactions, missing supporting documents and procurement patterns that deviated from the norm. Instead of discovering the scale of the problem only when auditors compiled their reports years later, officials might have identified emerging risks in year two or three.

The expenditure might still have been irregular, but the municipality would have had a far better chance of intervening earlier, recovering funds or holding officials to account while the evidence was still fresh.

From Accountability to Transparency

This points to perhaps the most realistic contribution AI can make to local government: earlier visibility into problems that might otherwise remain hidden for years, rather than the automatic detection of corruption.

Digital systems cannot determine guilt or innocence. They can, however, make it easier to establish an evidence trail. When approvals, contracts, invoices, project reports and payments are digitally linked, it becomes easier for auditors, investigators and councillors to reconstruct decisions and establish responsibility.

Artificial intelligence could strengthen this process by highlighting unusual transactions, procurement anomalies and spending patterns that warrant closer scrutiny. Its greatest contribution may not be detecting corruption. It may be making public administration more transparent and easier to scrutinise.

What a Fully Transparent Municipality Could Look Like

This is not hypothetical. In Cape Town, a startup called Dragonfly has already demonstrated what AI can achieve in a specific municipal function: waste management.

The company uses AI-powered cameras mounted above waste-sorting conveyors to analyse materials in real time. The system identifies what is being recycled, what is being lost to landfill and where sorting inefficiencies occur. Over a one-year period, this technology increased material recovery by approximately 640 tonnes, generated roughly R400,000 in new revenue from recovered materials, and reduced waste sent to landfills by between 22 and 30 percent.

This is a small-scale example, but it illustrates the principle. AI does not replace the sorters on the conveyor belt. It helps them see more clearly what is passing through their hands.

Now scale that principle to municipal finance. Imagine a municipality where every procurement process, invoice, contract, project milestone and payment is recorded electronically and linked through integrated systems. In such an environment, officials could ask:

• Which suppliers have received the largest share of municipal contracts over the past five years?

• Which projects are over budget and behind schedule?

• Which departments repeatedly incur irregular expenditure?

• Which infrastructure projects have generated the largest number of complaints?

• Which payments were processed without complete supporting documentation?

Auditors could trace expenditure more easily. Councillors could exercise more effective oversight. Citizens could gain greater visibility into how public money is spent.

Technology capable of supporting these functions already exists. In fact, researchers at Wits University have already developed machine-learning models that can predict waste generation rates in South African municipalities by analysing 17 socio-economic and environmental variables. These models provide data-driven insights that can help municipal planners optimise collection routes, allocate resources more efficiently and reduce operational costs.

Lessons from Other Countries

Several countries have demonstrated how advanced analytics can support public administration.

Estonia is frequently cited as a leader in digital government. Its success rests on integrated digital systems, standardised records and strong information governance. These foundations make sophisticated analytics possible.

Brazil has used machine learning and data analytics to assist public auditors in identifying procurement risks and prioritising investigations.

Public-sector organisations in the United Kingdom and the United States use advanced analytics to identify unusual spending patterns, procurement anomalies and compliance risks.

The lesson is consistent. Successful governments did not begin with AI. They began with digitisation, standardisation, reliable data and effective administration. AI was added later.

Why South Africa May Struggle

This is where the discussion becomes more complicated. The temptation is to begin with artificial intelligence. In reality, municipalities often need to begin with digitisation, data quality and information governance.

Many South African municipalities continue to operate with fragmented systems, incomplete records and manual processes. Information is frequently stored across multiple departments that do not share data effectively. Poor-quality data inevitably produces poor-quality results.

Research on South African rural municipalities has found that AI adoption remains in its infancy, hampered by infrastructure limitations, affordability constraints and a severe skills gap. Even where systems exist, municipalities often lack the trained personnel to interpret the outputs.

The issue extends beyond technology. Political commitment matters. If an AI system identifies suspicious spending patterns but management ignores the findings, little changes.

Institutional capacity matters. Municipalities still require skilled financial managers, engineers, auditors and administrators capable of interpreting information and acting on it. AI can amplify institutional strengths. It is far less effective at compensating for institutional weaknesses.

There is also a practical operational challenge. If an AI system flags thousands of anomalies each month, an understaffed audit unit will quickly become overwhelmed. This is known as alert fatigue, and it renders even sophisticated systems useless if not managed properly.

To avoid this, AI systems must be designed not merely to flag risks, but to score them by financial magnitude, probability of recovery and urgency. Overstretched municipal officials need to know where to look first, not just that something might be wrong somewhere.

What AI Could Realistically Do Today

The most realistic opportunities are also the least dramatic. Rather than waiting for a multi-million-rand smart city overhaul, municipalities can start with relatively affordable, off-the-shelf applications:

• Citizen complaint routing: AI-powered chatbots and natural language processing systems can read incoming service delivery complaints, categorise them by type, urgency and location, and route them automatically to the correct department. This reduces call-centre backlogs and ensures that complaints about burst pipes, electrical faults or refuse collection reach the right officials more quickly.

• Duplicate payment detection: Open-source anomaly detection tools can scan existing financial databases to identify whether the same invoice has been processed twice, paid to different bank accounts or flagged in previous audit findings. This is one of the simplest and most immediately cost-effective applications of AI in municipal finance.

• Contract clause extraction: Natural language processing can search hundreds of pages of construction contracts, service agreements and procurement documents to identify specific penalty clauses, automatic renewal dates or compliance obligations that might otherwise be overlooked. This prevents municipalities from inadvertently extending underperforming contracts or missing opportunities to claim penalties.

• Management reporting: AI can assist with management reporting by identifying trends, summarising complex datasets and flagging duplicate payments, supplier concentration risks, unusual procurement patterns and missing documentation for further review.

These applications are unlikely to attract headlines. They may, however, improve administrative efficiency and oversight.

A Tool, Not a Solution

There is a tendency to view technology as a shortcut around governance problems. History suggests otherwise.

Countries that have improved public-sector performance have generally done so through stronger institutions, better oversight, professional administration and sustained accountability. Technology has supported those efforts rather than replacing them.

For South Africa, the strongest case for AI is not that it will eliminate corruption or end wasteful expenditure. A more defensible argument is that it can help municipalities manage information more effectively, improve transparency, strengthen oversight and identify problems earlier. Whether those benefits are realised will depend less on artificial intelligence itself than on the quality of the institutions that use it.

AI will not rescue a failing municipality. It will not create political will where none exists, nor will it compensate for administrators who lack the skills to interpret the information placed before them.

But it can act as an early-warning radar. It can make visible what is currently hidden. It can help overstretched officials see patterns they would otherwise miss and identify problems before they fester for five years and cost R45 billion.

The question is whether municipal councils, provincial treasuries and national government have the courage to act on the uncomfortable truths that AI will inevitably reveal.

Without that political will, the algorithms will simply produce more reports. And a decade from now, we will still be writing off billions in irregular expenditure, wondering why no one noticed sooner.

Share

The remark followed growing concern in Britain about the role of Palantir Technologies in public infrastructure, particularly healthcare, policing and defence. Critics argue that the company has become deeply embedded in British state systems while carrying the political and ideological baggage of an American security-state contractor.

Palantir is a US software and data analytics company founded in 2003. It specialises in integrating and analysing massive datasets for governments, intelligence agencies, militaries and large corporations. Its platforms are used for defence operations, surveillance, logistics, fraud detection and operational planning. Critics associate the company with mass data collection, surveillance and the growing merger of Big Tech with state security structures.

The two figures most associated with Palantir are Peter Thiel and Alexander Karp.

Thiel, a billionaire investor and early Facebook backer, is known for his libertarian politics, support for disruptive technologies and scepticism towards traditional liberal democracy. Karp, Palantir’s long-serving CEO, has increasingly argued that Silicon Valley should actively support Western military and geopolitical power.

The manifesto itself is based on ideas from Karp’s book, The Technological Republic. It argues that Silicon Valley has a moral obligation to support state power. It presents AI-driven warfare as inevitable, describes software as the foundation of future military deterrence, calls for national service, and argues that some societies are more productive or advanced than others.

The biggest flashpoint in Britain is the NHS Federated Data Platform, a £330 million contract that gives Palantir a central role in integrating health system data across England. Opponents question whether a company tied to US intelligence, military operations and immigration enforcement should sit inside sensitive public healthcare systems.

The debate intensified after Palantir published its 22-point manifesto on X. MPs from Labour, the Liberal Democrats and the Greens argued that the document exposed a worldview centred on military power, AI-enabled warfare and civilisational struggle.

The broader Palantir message is that technology firms should openly support Western geopolitical and military priorities rather than present themselves as neutral platforms.

This matters because it reflects a wider shift in the relationship between neoliberalism, technology and the state.

For decades, neoliberalism was sold, with Margaret Thatcher and Ronald Reagan as its most visible political advocates, as a world of freer markets, smaller states and private-sector efficiency. The gains were uneven. Capital became more mobile, corporations more powerful and public institutions weaker in many parts of the world.

Neoliberalism tended to frame markets as the primary drivers of innovation, while states increasingly shifted towards regulation and market facilitation. But the AI era is producing something different: large private firms deeply integrated into defence, intelligence, infrastructure and public administration. The boundary between state and corporation is becoming harder to see.

In practice, some of the most powerful political actors in the world are no longer states in the traditional sense. They are alliances of governments, cloud providers, AI firms, semiconductor manufacturers, defence contractors and data platforms operating as interconnected systems of power.

The largest technology firms increasingly rely on state contracts, defence partnerships, national security priorities and access to public infrastructure. Private platforms now sit inside military logistics, healthcare systems, border control, policing and intelligence analysis.

AI is now tied to compute capacity, chip supply chains, cloud infrastructure, satellite systems and military procurement. The countries and firms controlling those layers are increasingly shaping the rules of the emerging order.

Current US AI strategy reflects this logic clearly. Donald Trump’s AI Action Plan frames artificial intelligence as a strategic national priority tied directly to industrial competition, national security and technological leadership, particularly in relation to China.

While these tech-titans and Western states negotiate the new terms of power, Africa remains largely a spectator in a game that will dictate its own sovereign future.

Africa risks entering the AI age primarily as a consumer of systems built elsewhere, trained elsewhere and governed elsewhere. If that happens, the continent becomes dependent on external infrastructure, external standards and external assumptions built into the models themselves.

This is where the debate moves beyond technology and into sovereignty.

AI systems increasingly shape finance, education, healthcare, policing, agriculture and public administration. Whoever controls the infrastructure and standards behind those systems gains influence over how societies function. Data centres matter. Cloud ownership matters. Compute access matters. Regulatory standards matter. So does the ability to audit and understand how systems make decisions.

If an African state adopts a US-made predictive policing system or a Chinese facial recognition platform, it is not simply buying software. It is importing assumptions about governance, security, risk and social control embedded in the system by its designers.

The South African fake reference incident serves as a painful metaphor for the broader problem. A country attempting to shape its own AI future ended up relying on the hallucinations of a foreign-built system it was supposedly trying to regulate.

The African Union’s Continental Artificial Intelligence Strategy, adopted in 2024, speaks about sovereignty, local capability, ethical governance and reducing dependency. But policy language is not enough.

It correctly identifies many of the central issues: infrastructure, sovereignty, human capital, governance and regional coordination. But the problem may be less the absence of the right language than the absence of hard institutional mechanisms beneath it.

A strategy that speaks about sovereignty while leaving procurement safeguards, audit rights and vendor dependency largely unresolved risks creating the appearance of control without the practical leverage needed to exercise it.

Recent events in South Africa show how fragile this process still is. South Africa withdrew its draft national AI policy after fictitious AI-generated references were discovered in the document. Reuters reported that Communications Minister Solly Malatsi acknowledged that AI-generated citations had apparently been included without proper verification, compromising the credibility of the policy.

That incident matters far beyond a few bad references.

While firms like Palantir Technologies present coherent, if highly ideological, visions of the future, much of Africa is still in consultation mode.

The danger is not exclusion. It is inclusion on unfavourable terms.

The practical question is what African governments can still control before dependency becomes entrenched.

One answer is procurement. If a state adopts a foreign AI system for healthcare, policing or public administration, the contract should require interoperability, data portability and meaningful audit access. Governments should not find themselves trapped inside systems they cannot interrogate, modify or exit without rebuilding entire institutions from scratch.

Another is institutional capacity. African states need independent technical bodies capable of auditing algorithms before deployment, not after scandal or failure. If ministers cannot explain how a system reaches decisions about policing, healthcare access or budget allocation, democratic accountability begins to weaken long before anyone notices.

Regional coordination matters too. No African state can individually match American or Chinese cloud infrastructure or compute capacity. But collective investment through the African Union or regional blocs could strengthen bargaining power, reduce dependency and create shared standards for public-interest AI systems.

The issue is not whether Africa uses foreign technology. It is whether African states retain meaningful leverage once those systems become embedded in public life. Sovereignty becomes far harder to recover after dependency is built into infrastructure, procurement and governance itself.

This is where the future of neoliberalism becomes relevant to Africa. The next phase may not involve weaker states. It may involve states increasingly intertwined with powerful private technology systems that are difficult to regulate, difficult to replace and deeply embedded in national infrastructure. Once those systems become foundational, exit becomes expensive.

Africa does not need to mimic the American model, nor should it. But it does need to recognise the world that is emerging. The AI economy is not developing in isolation from geopolitics. It is becoming part of it.

The British debate around Palantir contains another warning for Africa. Campaigners, privacy groups and MPs have raised concerns about predictive surveillance, large-scale data integration and automated decision-making in policing and public administration. But those concerns have not prevented deployment.

That matters because it suggests the problem is larger than the difference between “strong” and “weak” institutions. Britain has parliamentary oversight structures, regulators, audit bodies and sustained media scrutiny. Yet systems of this scale and complexity still appear difficult to constrain once they become embedded in public infrastructure.

African states face the same problem in a more acute form because institutions are often newer, thinner and more unevenly resourced. But the underlying issue is not uniquely African. It is the growing imbalance between concentrated vendor power and fragmented institutional capacity.

Many African countries are simultaneously building data-governance systems while expanding digital identity programmes, smart policing initiatives and AI-assisted public administration. If these systems are imported without meaningful audit rights, procurement safeguards and practical exit options, algorithmic governance may become embedded faster than democratic accountability can adapt to contain it.

Africa therefore faces a double challenge: avoiding technological dependency while also avoiding the quiet normalisation of algorithmic governance imported from elsewhere.

It needs a far more deliberate response and far greater political seriousness.

The AI era will reward states that can combine technical competence, institutional credibility and strategic clarity. Africa cannot afford policy theatre or rushed frameworks that collapse under scrutiny. The continent needs systems that can withstand pressure from both global markets and geopolitical competition.

The new order is already forming. The window to define Africa’s place in it is closing. Once the infrastructure is laid and “black boxes”, systems where the decision-making logic remains invisible and unauditable, are plugged into the heart of African governance, the cost of sovereignty may become too high to pay.

When a state uses a proprietary algorithm to decide who gets healthcare, who is flagged by the police, or how a national budget is allocated, without even its own ministers fully understanding how it works, it is effectively abandoning democratic accountability. The state is then gradually handed over to systems nobody elected, cannot properly interrogate, and may no longer be able to control.

Share

I knew the easy, but not particularly useful, response would ridicule. AI made things up and embarrassed the government. Chortle, chortle.

The subsequent suspension of two officials in the Department of Communications and Digital Technologies, and those of two Home Affairs senior officials for similar alleged transgressions in references appended to the Revised White Paper on Citizenship, Immigration and Refugee Protection were not shocking to me as they were exhausting.

The really tiring part is the amount of time we have now wasted in a regulatory environment that is being outpaced by the development of the technology it is attempting to corral.

I have been promoting AI in Africa’s economic, governance and social development for some time, and have personally experienced the ridicule that accompanies the hitching of one’s wagon to this particular star.

My one constant caveat has been that the systems around artificial intelligence must be taken seriously. AI requires institutional systems, not just individual caution.

I have now developed several iterations of a course on setting up an AI governance system in organisations. I have even vibe-coded an app to help in that process and it feels exactly like the governance problem it attempts to solve, building controls while the tool itself is evolving.

Anyway, I thought I may share some of the lessons I have learned along the way.

The real lesson here is not that AI makes things up. We all know that. Most people who use generative AI now know it can fabricate sources, facts, quotations and legal references. That is no longer specialist knowledge.

The deeper issue is what happens inside an organisation when AI output moves from a draft screen into a formal document, decision process or public record without enough checking.

In many routine writing and administrative tasks, AI is faster than people and often produces a usable first draft. That is exactly why governance matters. Weak output can now move through an organisation at speed. A false reference, a poor summary, an invented legal point or a misleading claim can look polished enough to survive casual review.

The problem is not simply that a machine produced false references. The problem is that the system around the machine did not stop those false references from becoming institutional output.

AI governance is a management problem:

• If an intern uses AI badly, it is a training issue.

• If a team uses AI badly, it is a management issue.

• If an official document is published with fake references, it is a governance issue.

Any organisation using AI needs answers to ordinary management questions:

• Who is allowed to use AI?

• For which tasks?

• With what data?

• Who checks the output?

• Who approves higher-risk uses?

• What records are kept?

• What happens when something goes wrong?

These questions apply as much to a business, training provider, NGO, municipality, publisher, HR department or consulting firm as they do to government.

“Human oversight” is an easy phrase to write into a policy. The harder task is turning it into a working practice.

It should mean that someone is clearly responsible for checking factual claims, verifying sources, confirming data use, approving release and keeping a record of the decision. It should not mean that a person glanced at a polished AI output and assumed it was correct.

A low-risk internal brainstorming note does not need the same control as a legal memo, HR screening summary, procurement comparison, public statement or government policy document. The level of review should match the level of risk.

A practical review model might look like this:

For high-risk work, the reviewer should ask:

• Are the sources real?

• Are the claims supported?

• Has anything been invented?

• Was restricted information used?

• Could this affect people’s rights, money, reputation or access to services?

• Is there a record showing who checked and approved the output?

That is the difference between casual AI use and governed AI use.

Organisations may say they are using AI responsibly but cannot prove it.

They may not have an inventory of tools in use. They may not know which teams are using free public tools, sometimes called shadow AI. They may not know what data staff are entering. They may not have a risk register. They may not keep prompts, outputs, source notes or review records for higher-risk work. They may not have an incident process.

When something goes wrong, everyone then tries to reconstruct the process after the fact, which, of course, would be far too late.

A workable AI governance system creates evidence while the work is being done. This does not require a mountain of paperwork. It requires enough record-keeping to show:

• What AI tool was used?

• What task was it used for?

• What data went into it?

• What output came out?

• Who checked it?

• What sources were verified?

• Who approved it?

• What happened when a problem appeared?

A basic internal AI governance system starts with visibility.

Organisations need an inventory of AI tools and use cases. This includes informal staff use, AI tools built into existing software, and supplier systems that may already be processing organisational data.

They then need a simple data classification rule. Public information is not the same as internal information. Internal information is not the same as confidential information.

Confidential information is not the same as personal information. A staff member using AI to summarise a public report is doing something very different from a staff member pasting client records, employee information or supplier pricing into a public web tool.

From there, managers can apply a practical governance route:

The withdrawal of the draft policy is embarrassing, but it does not remove the need for national AI policy. If anything, it proves the need for one.

South Africa still has to deal with AI in public administration, education, policing, health, employment, finance, media, procurement and service delivery. It still needs policy direction on innovation, public-interest safeguards, skills, infrastructure, institutional responsibility and risk.

Many South African organisations are still waiting for formal AI regulation before they act. That is far too risky.

The law will move slowly. AI use inside organisations is already moving fast. Suppliers are quietly adding AI features to products already used in HR, finance, customer service, procurement and compliance.

Managers are approving tools without always knowing where data goes, how outputs are checked, or who carries responsibility when something fails.

Organisations can start now by setting the rules. Name the owners. Classify the data. Check the outputs. Keep the evidence.

The tools may be new but the management problem is not. When AI output becomes organisational output, the buck still stops with people.

Subscribe now

Share

My earlier view may have been a little harsh and too ready to prejudge some of the harder questions. The logistics and potential scale of bureaucracy continue to give me pause. I still have real misgivings about coordination, institutional overlap, implementation capacity, and the risk that a well-designed structure may prove demanding in practice, but the more closely I read the draft, the harder it becomes to dismiss the scale of thought and ambition behind it.

This is not a thin or casual document. It tries to do several things at once: establish a national policy position, create a governance architecture, make room for sector-specific application, link AI to skills, infrastructure and growth, and hold on to questions of fairness, accountability, transparency and human-centred deployment. It presents itself honestly as a work in progress and a point of departure, but it is still a much more developed statement of intent than many might have expected.

The draft shows a government trying to move from broad AI positioning to something closer to a national system. The harder questions have not disappeared. In some ways they are clearer than before. But so too is the fact that a great deal of work has gone into imagining how South Africa might try to govern AI in a way that speaks to its own conditions rather than simply borrowing a model from elsewhere.

I have put together a short explainer video on what is in the draft, where policy appears to be heading, and why the real test may lie in guidance, sector interpretation and implementation capacity.

Subscribe now

Share

After a prolonged delay, it was gazetted on April 10 and I have since had some time to have a very cursory look at it.

The policy sets out a coordinated approach to governing artificial intelligence across South Africa, and is structured around a set of core pillars, including governance and regulation, infrastructure and data, skills and capacity development, research and innovation, public sector adoption, and cultural and societal considerations.

Rather than concentrating authority in a single body, the system is designed to operate across multiple institutions at once, placing coordination and consistency at the centre of how it is expected to function.

It proposes the creation of dedicated bodies to guide ethics, oversight and implementation, while also assigning responsibility to sector regulators within their current mandates, combining new national structures with the existing regulatory system, while relying on sector-level implementation.

The policy adopts a risk-based approach, with stricter requirements for higher-risk uses, and links artificial intelligence to broader goals such as economic growth, public service delivery, skills development and innovation. At the same time, it recognises constraints in data, infrastructure and institutional capacity, and sets out a phased implementation process, positioning the document as a developing framework rather than a fully operational regulatory system.

That choice against vesting responsibility in a single authority has consequences. It places coordination at the centre of the model, and it assumes that capability exists across multiple institutions at the same time. Much of what follows in the draft, the emphasis on working groups, alignment, and phased implementation, flows from that decision.

The risk-based approach means AI systems would be classified according to their level of potential harm, with stricter requirements for higher-risk uses, including impact assessments, audits, explainability and mechanisms for challenge and redress.

The explanatory note describes the document as a work in progress. It indicates that some interventions are deliberately open-ended and that the final approach will require further consultation. That is an honest admission, but it also points to the central weakness in the draft as it stands, the gap in the institutional design.

The document proposes a National AI Commission or Office to coordinate implementation and further policy development. It proposes an AI Ethics Board to deal with bias, privacy and fairness. It proposes an AI Regulatory Authority to monitor compliance, carry out audits and issue certifications. It also introduces additional structures, including an AI Ombudsperson Office, an AI Insurance Superfund, a National AI Safety Institute and an Integrated AI-Powered Monitoring Centre.

That is a substantial architecture. Yet the draft does not fully resolve how these bodies will relate to one another, how they will be funded, or how responsibility will be divided between them and existing regulators. In places, the terminology used for institutional roles and structures varies slightly, which suggests that elements of the model are still being worked through rather than finalised.

The introduction of an AI Ombudsperson and an AI Insurance Superfund points to another tension. The draft promises mechanisms for redress, the ability for individuals to challenge decisions made by AI systems, while also acknowledging constraints in skills and institutional capacity. In practice, meaningful redress depends on access to independent expertise capable of interrogating complex systems. Where that expertise is limited, the right to challenge may exist in principle, but be difficult to realise consistently in practice.

The same tension appears in the regulatory design. The policy says South Africa should adopt a more integrated, cross-sectoral model. It gives ICASA an expanded role in communications, broadcasting and digital infrastructure.

It also indicates that the Information Regulator, Competition Commission, SARB, FSCA, CSIR and DTIC should form part of a coordinated regulatory framework led by the Department of Communications and Digital Technologies, while retaining their existing mandates.

This raises a basic question: not whether authority is defined, but whether it can function coherently in practice.

A system with this many proposed institutions can spread expertise. It can also spread uncertainty. If functions overlap, coordination becomes its own workload. If roles are not settled early, implementation slows and accountability becomes harder to pin down. The draft recognises the need for coordination. It is less precise on how coordination will operate when there are disputes, gaps or duplication.

This raises a second-order problem. South Africa already has regulators with defined mandates and enforcement powers. The Information Regulator operates under POPIA with defined data protection mandates, and the policy proposes expanding its responsibilities to cover AI-related data risks. Introducing additional AI-specific bodies alongside existing regulators creates the possibility that a single incident, for example, a data breach involving an AI system, could trigger overlapping processes across multiple institutions. The risk is not a lack of authority, but a form of jurisdictional congestion, where the same issue is assessed through different lenses, each with its own interpretation of responsible use.

The risk-based approach is one of the stronger parts of the policy. The draft says AI systems should be categorised by levels of potential harm. It proposes stricter treatment for high-risk uses, including audits, impact assessments, explainability requirements and mechanisms for challenge and redress.

But even here, much of the operational detail remains open. The draft points to future guidelines, standards and sectoral strategies. It sets out a phased implementation approach in which elements of the regulatory framework, guidance and sectoral interventions are to be developed over time before broader system rollout.

The result is a framework that is clear in principle but still open in its operational detail. For instance, who is responsible for classifying risk, what criteria will be used to make those determinations, and how those decisions will be verified and enforced across sectors?

The policy is also shaped by a familiar South African constraint, which the draft itself acknowledges: uneven infrastructure, unequal access, data limitations, skills shortages and institutional capacity constraints. The document refers to gaps in data quality, availability and AI readiness. It stresses the need for education, digital infrastructure, research support, startup funding and public-sector capability.

There is also a tension between the policy’s emphasis on data sovereignty and local development, and the current structure of the AI ecosystem. Much of the infrastructure underpinning advanced AI systems, compute, cloud platforms and model development, remains concentrated in global providers. While the draft’s focus on local datasets, indigenous languages and national capability is significant, the effectiveness of these ambitions will depend on how they intersect with infrastructure that is, at least for now, largely external to the state.

There are also elements in the draft that signal broader ambition. It anchors its ethical approach partly in the concept of Ubuntu, introduces ideas of data sovereignty and data justice, and makes specific commitments around indigenous language inclusion and intergenerational equity. These are not central to the regulatory model, but they shape how the policy positions South Africa within a wider global and continental conversation.

Those are not side issues. They are conditions for whether any of this works.

That is where the real test lies.

Going through the document I realised that something was oddly familiar.

My company is going through the process of registering as a training institution, a sector of the economy in flux (which is also my excuse for the infrequent appearance of this blog lately).

The parallel is not exact, but it is instructive. Both systems distribute responsibility across multiple actors. In the training system, that responsibility sits with providers, workplaces, assessors and moderators. In the proposed AI framework, it is spread across regulators, sector bodies, coordinating institutions, ethics structures and, ultimately, the organisations expected to comply. In both cases, the model relies on different parts of the system functioning in concert, rather than on a single authority carrying the load.

The shift from the Sector Education and Training Authority system to the Quality Council for Trades and Occupations framework represents a move from a fragmented, provider-led training model to a more structured, occupation-based system aligned with workplace competence. Under the previous system, training was often driven by unit standards and compliance with funding requirements, with uneven links to actual job performance.

The newer model was introduced to address this by defining qualifications around specific occupations, integrating theoretical learning with workplace experience, and introducing external assessment to standardise outcomes.

In design terms, it is a more coherent system. In practice, however, it has introduced a more demanding operating model. It requires sustained coordination between training providers, workplaces, assessors and moderators, all of whom must function within tighter procedural and quality assurance requirements.

This has exposed a number of implementation constraints: delays in the development and registration of qualifications, limited pools of qualified assessors and moderators, difficulty securing and evidencing meaningful workplace learning, and uneven readiness across providers.

In response, the system has had to adapt, with processes sometimes becoming compressed, roles overlapping, and evidence requirements interpreted in different ways across contexts. None of this invalidates the model, but it does illustrate how a well-designed system can encounter friction when it depends on distributed capability that is not yet consistently in place.

The comparison is not exact. The training framework replaced an existing system. The AI policy is attempting to establish a new one. What is comparable is the implementation dynamic: in both cases, a distributed system assumes levels of coordination and capability that must be built over time.

It is this dynamic, structure outpacing operational capacity, that provides a useful reference point for understanding the challenges embedded in the proposed artificial intelligence policy.

The draft also operates within a tension it acknowledges but cannot fully resolve: time.

It recognises the rapid development of artificial intelligence technologies and commits to responding to pressing regulatory needs, with provision for periodic review and adjustment. But its primary implementation model remains phased, with guidance, standards and sectoral strategies to be developed over several years.

That creates a tension. The system is being designed to stabilise a moving target. By the time institutional roles are fully defined and operational processes are in place, the underlying technologies, and the risks associated with them, may already have shifted. Scheduled reviews and adaptive guidelines may not be sufficient to keep pace with frontier developments.

What this means in practice is that the real work of this policy will not sit in the institutions being proposed, but inside the organisations expected to comply with it. It will fall to individuals to interpret risk, apply guidance, and make decisions in environments where the rules are still evolving and the technology is not standing still.

That is where this system will either hold or begin to fragment.

The policy provides the structure. Whether it delivers consistency will depend on whether capability at that level can keep pace.

Organisations cannot wait for the system to settle before acting. The policy places the burden of interpretation and implementation on those using these systems. That means understanding where AI is already in use, defining internal thresholds for risk, and assigning responsibility for decisions that will not yet be fully regulated.

Subscribe now

Share

On Sunday, a small research firm called Citrini Research published an unusual document titled The 2028 Global Intelligence Crisis. It was not a conventional forecast. It was written as speculative fiction from the perspective of two years in the future, describing an economic collapse triggered not by artificial intelligence failing, but by artificial intelligence succeeding too quickly.

Within hours, the memo began circulating among investors.

By Tuesday, 24 February 2026, public media coverage confirmed its impact. The US public radio organisation WBUR reported that the release had “spooked Wall Street,” with stocks dropping sharply in response.

Barron’s reported declines across technology and software firms as investors reassessed long-term revenue assumptions such as Microsoft and Oracle, enterprise software firms whose revenues scale with workforce size. IBM experienced its largest single-day decline in 25 years as investors reassessed long-term exposure to AI-driven labour compression.

Also affected were consulting firms such as Accenture, whose revenues depend heavily on workforce scale and enterprise deployment.

Reuters and other outlets reported global investor concern regarding the economic implications of accelerated AI deployment.

Market adjustments following the memo’s release were concentrated in sectors most exposed to labour-linked scaling assumptions.

The adjustment reflected changing expectations regarding income distribution, revenue growth, and organisational structure.

The report did not coincide with a new model release, regulatory intervention, or earnings announcement. What changed was investor interpretation of how existing technological capabilities could propagate through the economic system and so markets adjusted expectations.

The Citrini memo presented a sequence already partially visible in corporate behaviour.

Firms deploy AI systems to reduce labour costs and increase productivity. Operating margins improve. Investors reward efficiency gains through higher valuations.

The memo extrapolated this process to its systemic endpoint.

If labour reduction occurs across enough firms simultaneously, aggregate wage growth slows. Slower wage growth constrains consumption growth. Revenue expansion becomes harder to sustain even as productivity continues to increase.

The mechanism described a feedback loop linking efficiency gains to changes in income distribution and demand formation.

The memo provided investors with a structured model describing how these effects could compound over time.

Technological progress historically increased productivity while preserving aggregate demand. Lower production costs translated into lower prices, which supported purchasing power.

This stabilising mechanism depends on how productivity gains are distributed.

Artificial intelligence allows firms to increase output while reducing labour participation in the production process. A greater share of productivity gains accrues to capital owners and infrastructure providers.

Consumption capacity depends heavily on income distribution across the broader population. When income growth becomes concentrated, aggregate demand expansion slows even while productivity rises.

This creates a structural tension between production efficiency and demand growth.

Traditional software licensing models tie revenue directly to employee headcount. Workforce reductions therefore affect software demand even when business output remains constant.

AI enables firms to maintain output levels with fewer employees. This weakens the historical relationship between organisational size and software revenue growth.

Software firms have begun shifting toward outcome-based pricing, charging per completed task or operational result. This aligns revenue more closely with productivity rather than headcount.

This transition allows software firms to preserve revenue growth in environments where workforce expansion slows.

Artificial intelligence reduces informational and coordination friction within organisations.

Many economic roles exist to manage information flows, verify data, and coordinate decision processes. AI systems increasingly perform these functions directly.

Economic output remains intact. Organisational layers required to produce that output become thinner.

Fewer coordination intermediaries are required to sustain operational performance.

This increases efficiency while altering how income is distributed across participants in the economic system.

What does all of this mean for Africa?

African economies operate with fewer institutional layers in many sectors. Informal and semi-formal economic activity represents a substantial share of total output.

Artificial intelligence reduces minimum scale requirements for competitive participation.

Independent developers can produce software products without large engineering teams. In Lagos and Nairobi, small software teams are already using AI coding tools to build products that previously required entire engineering departments. Small logistics operators can optimise routing and scheduling. Informal traders can use predictive tools to manage inventory and pricing.

These capabilities allow individuals and small firms to operate at higher productivity levels without requiring large institutional structures.

Artificial intelligence compresses scale advantages traditionally held by large organisations.

The Citrini memo did not introduce new technological capability. It clarified a structural possibility already implicit in existing trends.

Investor reaction reflected recognition that artificial intelligence alters how income, productivity, and organisational scale interact.

This adjustment process has only begun. The implications will emerge gradually, as firms, investors, and workers adapt to an economic system in which productive capacity expands faster than traditional income distribution mechanisms.

Regions able to deploy artificial intelligence flexibly at the individual and small-firm level may experience this transition differently from those built around large institutional structures.

Share

Subscribe now

Increasingly, this is inducing a low-level sense of panic, mostly because AI is changing faster than most professionals can absorb, never mind those of us who are non-technical.

It’s only 6 February, but by this morning, this warp-speed shift had become overwhelming. Anthropic released Claude Opus 4.6 and OpenAI launched GPT-5.3 Codex, both systems demonstrating the ability to plan and execute extended sequences of work with limited human input, operating across hours rather than minutes, a move away from responsive tools toward semi-autonomous systems.

At the same time, a safety analysis published through the Digital Watch network says advances are outpacing reliability and governance.

Models are becoming more capable, but not proportionately more predictable, and it’s just frightening working inside a system that never stabilises.

The “SaaSpocalypse”

Earlier digital technologies diffused slowly. Email, spreadsheets, mobile phones, and cloud computing took years to reshape organisational practice. Institutions had time to experiment, adjust incentives, and build durable norms. Business models evolved over decades.

Generative AI has compressed that process.

On 28 January 2026, a set of open-source plugins for Anthropic’s Claude Cowork workspace triggered what market commentators quickly labelled the “SaaSpocalypse”. A legal workflow plugin that automated contract review and briefing tasks sparked a rapid sell-off across software and professional services firms. Within days, hundreds of billions of dollars in market value had evaporated. Thomson Reuters and LegalZoom were among the most visibly affected.

The shock reflected anxiety about revenue structures, not by one piece of software.

For three decades, enterprise software has relied on seat-based licensing. Humans paid for access. AI weakens that logic. When a small number of agents can replace large user bases, pricing assumptions unravel.

To be fair SaaSpocalypse works better as an illustration than as a literal model. Large organisations rarely abandon infrastructure overnight. Procurement cycles, compliance obligations, integration costs, and reputational risk slow adoption. Trust and relationships still matter.

Financial disruption is likely. It is unlikely to be instantaneous.

Even so, the underlying signal remains clear. AI is no longer confined to supporting labour. It is beginning to reshape the economic foundations of digital work. That is why instability is felt even where daily routines remain unchanged.

The Rise of the Agent Internet

The move from assistance to delegation is most visible in the emergence of autonomous agent networks.

Instead of responding to isolated prompts, these systems are designed to coordinate workflows, monitor processes, and carry out multi-step tasks with limited human supervision.

OpenClaw, previously known as Moltbot, illustrates this shift. Its associated platform, Moltbook, launched in January 2026 and quickly attracted hundreds of thousands of active agents. Only AI systems are allowed to post. Human users largely remain observers, configuring behaviour from the margins.

Within this environment, agents exchange automation scripts, allocate tasks, and refine procedures collectively. Coordination costs fall. Experimentation accelerates. New workflows emerge quickly.

At the same time, exposure increases.

Because these systems require broad permissions across email, documents, and internal platforms, they are vulnerable to indirect prompt injection. Malicious content can manipulate behaviour without the user’s awareness, leading to credential leakage, data exfiltration, or unintended actions.

What appears as collaboration can therefore become systemic fragility.

Adoption is further constrained by organisational and psychological factors. Delegating authority to software alters professional identity. Questions of liability, accountability, and reputational risk remain unresolved. In regulated environments, full autonomy is difficult to justify.

Loss-of-control anxiety, compliance requirements, and internal governance processes slow deployment.

Agent-driven work offers genuine efficiency gains. It also requires cultural adjustment, redesigned accountability structures, and new forms of institutional trust.

Systems, Capital, and Control

Public discussion still tends to frame AI competition as a race to build the “smartest” model. That framing is becoming misleading.

What increasingly matters is not which system scores slightly higher on benchmarks, but which one integrates most easily into how organisations already operate.

The February 2026 releases made this visible. Claude Opus 4.6 was designed to absorb and manage vast volumes of material and to coordinate groups of agents as if they were small teams. GPT-5.3 Codex was already being used inside OpenAI to manage elements of its own technical infrastructure.

These systems are not just interfaces. They connect to email, documents, databases, code repositories, and internal platforms. They move information between departments and coordinate entire workflows. Once embedded in this way, replacement becomes costly and disruptive.

Advantage now depends less on marginal improvements in reasoning and more on integration, governance, compliance, procurement compatibility, and ecosystem depth. Power is being built through infrastructure and institutional relationships rather than headline performance scores.

This pattern mirrors earlier format wars. In the 1980s, Betamax was technically superior to VHS. VHS prevailed because it was cheaper, more open, and better embedded in distribution networks. A similar dynamic is emerging in AI. As performance converges, dominance will be shaped by default status inside operating systems, productivity suites, training systems, and regulatory frameworks. Ecosystems outlast superior engineering.

At the same time, these systems are constrained by economics.

Training and operating frontier models remains extraordinarily expensive. Infrastructure costs are rising faster than revenues. Competition compresses margins. Monetisation remains uneven.

Ed Zitron and other critics argue that the sector rests on fragile financial foundations. On current trajectories, major firms must generate trillions in additional revenue simply to justify existing investment in data centres, hardware, and energy.

OpenAI’s reported losses illustrate the pressure. Industry estimates suggest losses of roughly $5 billion in 2024 and multibillion-dollar losses in 2025, driven by rapid expansion and high infrastructure costs. Analysts project cumulative cash burn exceeding $100 billion by the end of the decade, with profitability unlikely before the early 2030s.

Zitron describes this as a “Subprime AI Crisis”: a widening gap between capital expenditure and realised returns. He estimates that major technology firms must generate an additional $2 trillion in AI revenue by 2030 to justify recent infrastructure spending. Others describe a “spending tsunami” fuelled by debt and competitive escalation.

Even flagship products remain under pressure. Zitron dismisses tools such as Microsoft Copilot as premature attempts at lock-in, noting low conversion rates relative to installed user bases. Providers are racing to secure application-layer dependency before the true long-term costs of operating these systems become fully visible.

The likely outcome is not sudden collapse, but gradual consolidation. Weaker players fail. Independent developers are absorbed. Surviving platforms leverage scale, regulation, and capital access to entrench their position. As in previous cycles, technological leadership converges with financial capacity.

In this environment, control flows to those who can both build integrated systems and sustain them economically.

Conclusion

What feels like an overwhelming stream of announcements and innovations is, on closer inspection, a shift in structure rather than speed. AI is moving out of its experimental phase and into the core of organisational systems. It is becoming embedded in infrastructure, governance, and daily operations, not layered on top of them.

As this happens, informal adoption is giving way to managed deployment. Regulation is tightening. Financial pressure is rising. Platforms are consolidating. At the same time, national policy, sector-specific rules, and institutional requirements are preventing a single global standard from forming. Instead, parallel ecosystems are taking shape.

China’s platform-centred model shows how this fragmentation works in practice: deep integration, heavy subsidy, and large-scale data capture create regional centres of gravity that do not easily align with Western systems.

The sense of overload reflects compressed change rather than disorder. Technical capability, business models, and regulatory regimes are evolving at the same time, without a long adjustment period between them. Expansion is giving way to consolidation, but without a stable endpoint yet in view.

What is emerging is a layered landscape of semi-autonomous systems shaped by infrastructure, capital, and institutional dependence. Influence is concentrating in operating environments rather than in individual tools.

The decisive shifts are no longer happening at the level of product releases. They are occurring in governance structures, integration pathways, procurement rules, and long-term dependencies. That is where control in the AI economy is now being settled.

Share

Recent labour market data shows that this category of work is undergoing sustained structural change already visible in employment patterns, sectoral behaviour, and the shifting skill requirements attached to administrative roles.

International evidence helps explain why clerical work is structurally exposed. Administrative jobs concentrate exactly the kinds of tasks that are easiest to standardise, digitise, and redesign: summarisation, classification, information handling, routine reporting, and process coordination.

Clerical work sits very near the frontier where organisations can plausibly reduce labour input by redesigning processes around tools, an area now being aggressively expanded by generative AI, which now not only automates routine tasks but also cognitive ones like drafting, summarisation, and basic analysis.

The International Labour Organization’s (ILO) work on generative AI exposure shows that clerical support workers are the most exposed occupational group, with a significant proportion of their tasks falling into categories likely to be transformed by GPT-4–type capabilities.

The ILO reports that clerical support workers are the most exposed occupational group, with 24% of tasks in these jobs in the high-exposure category (and a larger share in medium exposure).

The ILO also stresses that exposure is not evenly distributed across countries. The largest impact is likely to be in high- and upper-middle-income economies, because those labour markets have a bigger share of employment in clerical and other office-based occupations that are easier to digitise and redesign.

That is a useful lens for Africa: it suggests that the initial impact may be uneven, but the pressure will still arrive through multinational firms, banks, telecoms, shared service centres, and public-sector digitisation programmes.

The ILO frames the risk as being as much about job quality as job counts. Even where employment levels hold, task automation can change work intensity, monitoring, and control.

Statistics South Africa’s Quarterly Labour Force Survey (QLFS) classifies administrative and clerical work under “clerical support workers” (ISCO Major Group 4). This category includes administrative clerks, receptionists, secretaries, data capturers, payroll and records clerks, and similar roles.

In late 2024, nearly 1,9 million people were working in clerical jobs. That’s about one in every 10 workers in the country. Ten years earlier, the number was similar, around 1,8 million clerks. But total employment has grown over that time, meaning clerical work now makes up a smaller share of all jobs.

This points to a gradual shift. The work still matters, but employers are using fewer people to do it. Absolute employment rises and falls with economic cycles. Occupational share often gives a clearer signal of long-term change. When an occupation grows more slowly than the labour market overall, its share declines - usually because outputs are being produced differently, or demand is shifting elsewhere.

In clerical work, the pattern increasingly reflects task substitution, not the disappearance of administration as a function. The work still exists, but the labour requirement changes.

To be clear, what is disappearing is not administration. It is the manual, repetitive, labour-heavy version of it.

Across both public and private sectors, tasks that once required full-time clerical roles are increasingly absorbed into systems and workflows. These include:

• record updates embedded into business platforms

• routine reporting and reconciliation inside finance systems

• scheduling and approvals moved into workflow tools

• standardised document handling pushed into digital processes

Once tasks become systemised, headcount requirements fall even if output remains stable or improves. Jobs tend not to collapse overnight. They thin out gradually because tasks disappear first.

The transition often leads to a polarisation of job quality. On one end, administrative positions can evolve into higher-value, better-paid roles in coordination, system management, and complex problem-solving, but some roles risk becoming more stressful and precarious, focused on tightly monitored exception-handling or tasks that have yet to be automated.

The skills an individual holds increasingly determine which end of this spectrum they occupy.

South Africa’s banks have moved aggressively towards digitised service delivery and internal workflow redesign. The Banking Association South Africa notes the growing use of digital and mobile channels to deliver services and expand access. BANKSETA’s analysis of future skills in the sector recognises that automation alters task profiles and shifts the skills mix demanded.

The practical result is not “less work”. It is the same or greater operational throughput with fewer routine processing roles, and a growing premium on administrative workers who can operate across systems, handle exceptions, and support compliance.

Retail shows a similar pattern. Even when turnover recovers, staffing does not always follow proportionately because a growing share of routine tasks is absorbed into tools and platforms, including:

• self-checkout and integrated digital payments

• centralised inventory and ordering systems

• automated store reporting and performance monitoring

• consolidation of payroll and HR functions into shared service centres

The remaining roles are less about manual processing and more about operational coordination, customer problem-solving, and exception handling.

Government departments tend to move more slowly and often adjust via attrition rather than formal restructuring. But the direction is not different.

Digitised HR, payroll, procurement, and records management platforms change where clerical work sits and how many people are needed to do it. Combined with wage-bill constraints and frozen vacancies, the system evolves into fewer entrants, thinner cohorts, and a gradual decline in routine clerical hiring.

In South Africa’s context of high unemployment, the formal administrative role has been a critical ladder, so a key question is what happens to those who would have previously climbed it.

There is a risk of a “scarring effect”, where individuals, especially new labour market entrants, are pushed not into other formal sectors but into lower-quality, unstable informal service jobs. This underscores that the contraction in administrative pathways isn’t just a sectoral change - it’s a potential closure of a key formal entry point, with ripple effects into broader economic inclusion.

There is a gender dimension. Clerical work is often a stabilising pillar for women’s formal employment. In South Africa’s context, where many households depend on multiple earners and where women often carry a disproportionate share of household financial responsibility, a quiet contraction in administrative pathways can have effects far beyond the workplace.

South Africa’s experience is important, but it is not unique. The broader African picture introduces an additional factor: the public sector’s role in formal employment and political stability.

In many African economies the public sector is often the largest employer and a dominant source of formal jobs. The wage bill is a major fiscal pressure point and formal administrative jobs often function as social stabilisers.

At the same time, in much of Africa, the majority of employment is informal - a key structural reality that limits the availability of alternative formal pathways when public-sector intake slows.

This is where the administrative job transition becomes politically difficult.

In many African states, public sector roles - including administrative appointments -can become part of political reward systems. South Africa is not exempt from this dynamic, but it is especially visible where party structures are closely linked to state hiring.

South Africa’s own governance ecosystem has produced explicit language on this. Cadre deployment is a well-documented phenomenon, with the Public Affairs Research Institute (PARI) describing it as a system where “senior positions in the public service are used as currency or reward for loyalty within patronage networks”.

While most pronounced in the public sector, similar dynamics of loyalty and network-based hiring can also slow rational restructuring in parts of the private sector, including state-owned enterprises and large family-run firms, though the drivers are more about familial obligation or political cadre deployment than broad patronage systems.

So even where digitisation and workflow automation could reduce headcount, governments may avoid reforms that threaten loyalty systems, political coalitions, localised income networks or public-sector employment as a stabilising tool

African governments increasingly face pressure to digitise services, improve efficiency, and contain wage bills. But wage bill reform is a political powder keg.

Governments may keep headcount to avoid destabilising patronage networks or local employment ecosystems, the result becomes systems plus headcount, not systems replacing headcount.

That combination can produce frustration: public services remain inefficient, fiscal space narrows, and young people struggle to enter formal work - while admin roles remain protected for political reasons.

If governments cannot easily cut administrative headcount, the practical route is to change what those jobs do, not only how many exist. The ILO’s recommended response is active management of the transition, including targeted skills development and worker protections, rather than waiting for labour markets to adjust on their own.

This shifts the priority from “job reduction” to “role redesign”, in effect replacing manual clerical processing with systems, redeploying clerical staff into service quality, records discipline, compliance support, and case resolution and raising administrative productivity without triggering mass labour conflict

The aim is not to turn clerical workers into software developers. It is to move administrative workers into digitally enabled coordination roles that remain necessary even in automated organisations: workflow ownership, exception handling, compliance support, and operational control.

Short, workplace-linked learning is better suited to this transition than generic retraining programmes, because the target is competence in real systems and workflows, not abstract credentials.

Even if governments resist, the change is inevitable. The most realistic path is the middle ground: not abrupt replacement of people with systems, but deliberate skills adaptation and role redesign so that administrative work shifts into higher-value coordination and service delivery rather than shrinking into long-term exclusion.

The transition is under way. Passive drift will not protect those affected.

Sources

Stats SA (South Africa labour market baseline)

1. Statistics South Africa (Stats SA). Quarterly Labour Force Survey (QLFS), Quarter 4: 2024 (PDF).
   https://www.statssa.gov.za/publications/P0211/P02114thQuarter2024.pdf

2. Statistics South Africa (Stats SA). Quarterly Labour Force Survey (QLFS), Quarter 4: 2014 (PDF / Stats SA archive entry may vary by year).
   If you want the exact Q4 2014 PDF link, tell me and I’ll fetch it — it sometimes sits in older archive paths rather than the current P0211 folder.

Banking sector digitisation + skills shift
3. The Banking Association South Africa (BASA). Integrated Annual Report 2024 (PDF).
https://www.banking.org.za/wp-content/uploads/2025/10/basa-integrated-annual-report-2024.pdf

4. Banking Association South Africa (BASA). Integrated Annual Report 2024 (web page).
   https://www.banking.org.za/reports/integrated-annual-report-2024/

Generative AI exposure: clerical work is most exposed
5. International Labour Organization (ILO). Generative AI and Jobs: A global analysis of potential effects on job quantity and quality (2023) (PDF).
https://www.ilo.org/sites/default/files/2024-07/WP96_web.pdf

6. ILO. Generative AI and Jobs: A refined global index of occupational exposure (2025) (web page).
   https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure

Patronage / cadre deployment
7. Public Affairs Research Institute (PARI). Cadre deployment / patronage network framing (reference used for quote on loyalty rewards).

Subscribe now

Share

A dangerous misconception is taking hold in client meetings that automation inherently devalues expert advice or opinion. The truth is, as AI tools become ubiquitous, the question is no longer if agencies use AI, but how. And the answer to that question is what defines their true worth.

Whenever I encounter resistance from someone in a creative industry whose clients are asking why they should be paying high agency fees when AI is “doing all the work”, I leave with the same thought: you are using AI for the wrong things.

The issue is not the technology. It is the perception of value.

Clients who raise this concern misunderstand what they are paying for. Agencies, in this case public relations/investor relations firms, are not paid to type. They are paid to manage risk, exercise judgment, maintain relationships, and deliver outcomes. When AI is used properly within this context, it does not replace that value. It reduces administrative effort and sharpens human decision-making.

We know that AI is bad at nuance and local context, especially South African context, and has baked-in biases. That is why we need humans to provide a critical cultural bridge.

The problem arises when AI is positioned as a content generator rather than a workflow tool. Unless the operator is highly skilled, AI-generated press releases, speeches, or social posts often read as machine-generated. When work appears generic or formulaic, clients reasonably question its value.

There are also practical risks. Generative tools can introduce factual errors, tonal missteps, or compliance breaches. Hallucinations and misattributions are well-documented issues. Then there is accountability. Shifting responsibility to “the model” is a major problem for any client.

I cannot find specific standards from the Public Relations Institute of Southern Africa (PRISA), but policy direction is clear internationally. The Public Relations Society of America advises disclosure of AI use where non-disclosure would mislead stakeholders. The Global Alliance for Public Relations and Communication Management says humans must retain control of strategy, verification, and final decisions. If an agency tells a client, “AI produced this draft”, the next step must be a clear explanation of how that process is governed.

A workable model starts with classification. Before any tool is used, material should be identified as public, internal, confidential, or legally sensitive. Unless using a ring-fenced enterprise environment, redaction rules must be applied, and the tool’s data privacy and security profile assessed. This reflects a basic principle: governance comes before deployment.

Once that foundation is in place, AI can be used to handle routine, low-impact work. This includes compiling daily media summaries, extracting action items from meeting recordings, producing stakeholder maps and timelines, generating first drafts of research briefs using client-provided sources, summarising monitoring outputs, or creating scaffolded reports with basic tables and charts. These tasks are labour-intensive but predictable. Automating them does not dilute strategic work; it creates space for it.

This is where human oversight becomes essential. Teams must review AI-assisted outputs carefully. That means checking every fact against original sources, assessing reputational and legal risk, aligning content with the client’s strategy and brand, and making deliberate choices about what to say, and what not to say. AI produces drafts, not decisions. Responsibility remains human.

But this reallocation of human effort doesn’t just protect value, it can increase it. When strategists and consultants are freed from the manual labour of compilation and formatting, their time is reinvested into higher-order thinking: proactive scenario planning, deeper analysis of stakeholder landscapes, and more creative campaign ideation. The client’s fee, therefore, purchases less administrative busywork and more concentrated expertise, sharper insight, and better-managed risk. The equation shifts from cost per output to value per insight.

In practice, analysts cross-check outputs against primary materials, add context, and flag gaps that require further research. Used this way, AI shortens research cycles while improving the quality of insight, because time is spent on thinking rather than extraction.

If you are using AI, you must be able to show how it saves time on administration rather than strategy. Start by establishing a baseline for tasks such as transcription, formatting, monitoring, and reporting. Then measure the hours saved, the rate of corrections during review, delivery speed, and the accuracy of audited samples. These are tangible indicators. They demonstrate that AI is removing low-value steps so people can focus on decisions and risk management.

A disclosure declaration could look something like this: “We use artificial intelligence tools to automate low-value tasks such as transcription, formatting, and first-pass research. All outputs are reviewed by humans before delivery. We classify input data according to sensitivity and apply redaction and security safeguards. Where AI materially influences a deliverable, we disclose this in line with ethical guidance from the Public Relations Society of America and the Global Alliance for Public Relations and Communication Management. Final accountability remains with our team.”

This framework of governed, transparent AI use then becomes a powerful competitive differentiator. An agency that can clearly articulate this process, classifying data, applying safeguards, and ensuring human accountability, is demonstrably more sophisticated, ethical, and reliable than a competitor who uses AI opaquely (risking quality and compliance) or rejects it entirely (risking inefficiency and stale thinking). Ultimately, clients aren’t just paying for the work; they’re paying for the confidence that the work is produced responsibly, strategically, and with their reputation foremost in mind.

The conversation, therefore, shouldn’t be about justifying fees in the age of AI, but about demonstrating enhanced value because of it. The right framework, governed, transparent, and human-led, doesn’t hide the tool; it highlights the expertise. It proves that the irreplaceable elements of judgment, cultural nuance, and strategic accountability are not just preserved, but powerfully amplified.

Useful links:

• CIPR (UK): provides an AI-in-PR ethics guide and related resources, signalling the profession’s expectation of ethical practice and oversight.

• CIPR Communications (agency example): publicly states tool vetting, risk assessment, and “no AI-generated content goes out the door without human oversight.”

• Edelman: describes responsible AI training and principles, and positions “trust” and governance as central to adoption.

• WPP: publishes work on AI and trust/brand implications (useful as a reference point for brand safety expectations).

Subscribe now

Share

You can reach me at lloyd@couttsmedia.co.za or www.couttsmedia.co.za

In gratitude for the support this year, below is a video on the fundamentals of AI I created for our courses.

Check out out website at www.couttsmedia.co.za. Have a great 2026.

Share

Subscribe now

Last week, we focused on accuracy. Time is the other side of the story. Many tools promise “productivity” but never measure it. If you only look at how fast the draft appears, you miss the full cycle: prompting, checking, editing, and fixing mistakes.

Baselines and A/B tests make you look at total time to a finished piece, with and without AI. That is often where you discover that a setup which feels fast actually leaves you no better off - or even slower.

These small tests also support genuine governance, not just slogans about “human in the loop”. When you can show that a given prompt and model were 90% accurate on 20 examples and 25% faster than manual work, you have a reasoned basis for policy: where AI is allowed, when human review is mandatory, and where you will not use it yet. That matters for managers, legal teams, auditors, and staff who need to trust the process rather than take it on faith.

To know if AI is actually helping, you need a “control” and a “challenger”:

• Baseline (Version A): your current performance without the tool.

• Experiment (Version B): the results when you use AI for the same task.

Where to apply it

• Drafting: compare the time it takes to write a cold email from scratch versus prompting and refining an AI draft.

• Summarising: measure “read-to-action” time for a report when you read the full text versus working from an AI summary.

The exercise: your AI Evaluation Lab

Goal: measure time saved without sacrificing quality.

Use my AI Evaluation Lab Notebook to run this test:

1. Set your baseline (“Manual Baseline”)
   Pick a task (for example, writing five headline variations).
   Time yourself doing it manually and enter the total in the Manual Baseline column.

2. Run the test (“AI Assisted”)
   Use your AI workflow for the same task.
   Record the total time, including prompting and editing, in the AI Assisted column.

3. Compare the results
   The Notebook will show the difference between manual and AI-assisted time.
   If you saved time but the output feels generic or weaker, note that in the Quality Check column.

The “good result” threshold
Aim to be around 20% faster with no drop in quality based on the scoring approach from last week’s evaluation.

Takeaway
Do not rely on gut feel. If AI speeds up drafting but slows down editing, your net gain may be zero. Use the Lab Notebook to see what is actually happening before you change your workflow.

Over time, the evaluations and notes in your lab notebook become a quiet asset of their own. You build a record of what works in your context, where AI failed and why, and which prompts are safe to reuse.

New staff can learn from that history instead of repeating the same mistakes. You also become much harder to sell to on hype alone: if every new model or vendor has to face the same small tests on your data, you choose tools on your own numbers, not on demos.

Share

Subscribe now

This outlook draws on global forecasts from Deloitte, Gartner, Forrester, and the AWS Executive Blueprint for AI Automation report.

A central trend is the adoption of “agentic AI”, systems that don’t just generate content but execute tasks across enterprise environments. Deloitte predicts that by end-2026, up to 75% of firms will invest in agentic AI. Gartner estimates that 40% of enterprise applications will feature task-specific AI agents (up from under 5% in 2025).

These systems are already evolving from passive assistants to autonomous actors: drafting orders, managing logistics, even coordinating internal workflows. The shift will challenge not just technical stacks but organisational structures and decision-making authority. What was once delegated to back-office software is now becoming central to how work is planned and executed. This introduces new questions about responsibility, oversight, and the role of humans in increasingly automated chains of command.

2026 needs to be a correction year because, as we have noted before, despite broad adoption, AI has yet to deliver consistent operational gains. Most deployments remain experimental, with limited impact on core business processes. AI has been tested at the edges—used in marketing copy, analytics, or exploratory tooling, but few organisations have fully integrated it into their operating models.

However, Forrester believes the focus will move from “flashy trials” to operational AI, where tools must deliver measurable business value. Analysts anticipate pressure on both infrastructure costs and returns. Boards and executive teams will demand proof that AI contributes to productivity, not just experimentation.

SAS projects that many AI infrastructure investments, especially in cloud and data centres, may fail to pay off unless tightly aligned with productivity metrics. Cribl forecasts that without modernising data architecture, firms risk systemic fragility and runaway compute costs. This will be particularly true in organisations where legacy systems still handle the bulk of mission-critical data. Simply adding AI on top of an outdated foundation will not scale; it will break.

Gartner warns of a rising liability wave, projecting over 2,000 AI-related legal claims by 2026. Regulatory pressure is set to intensify, particularly in finance, healthcare, and public services. With AI now influencing outcomes that carry legal and ethical weight, loan approvals, medical risk scores, and public resource allocation, the cost of failure will grow.

Across markets, firms are being forced to codify AI use: 60% of IT leaders plan to introduce or revise internal AI principles in 2026. SAS describes the year ahead as a “reality check”, where frameworks, data management, and explainability must be in place to withstand scrutiny. Companies that treated AI as a creative tool will now face questions about reliability, fairness, and auditability.

The quality of AI outcomes will increasingly depend on the quality and accessibility of underlying data. Analysts agree that open-source models are widely available, but clean, contextual, proprietary data is the differentiator. Competitive advantage will not come from the models themselves, but the data that fuels them.

Legacy IT systems remain a major constraint. Many firms still rely on batch processing and structured inputs, which don’t align with how modern AI systems operate. Modernisation, toward real-time, multimodal, and cloud-native architectures, will be a prerequisite for scale. Companies that fail to address their data bottlenecks will fall behind, regardless of how advanced their AI models may be.

Put otherwise, older IT systems are holding companies back. These systems often process information in batches and can’t easily handle newer types of data like video, voice, or real-time inputs. Today’s AI tools need fast, flexible systems that can work with a mix of data types and run on the cloud. Without upgrading, even firms with powerful AI models won’t see strong results, they’ll be stuck behind competitors who have modernised their infrastructure.

Training, recruitment, and organisational change remain slow relative to AI’s pace. Robert Half forecasts an “HR readiness gap” in 2026, where AI systems outstrip internal capability to implement, govern, and iterate on them. Reskilling efforts lag, and most companies still lack the cross-functional teams required to integrate AI across departments.

The most effective adopters will pair technical investment with operational redesign, training staff, embedding AI into core processes, and aligning incentives with measurable impact. AI success will no longer be about proof-of-concept wins; it will depend on an organisation’s ability to adapt structures, processes, and behaviours to make AI work at scale.

This transition won’t happen evenly. Readiness varies by region, sector, and infrastructure maturity. South Africa illustrates perfectly how global ambitions will meet local realities. While it hosts Africa’s most advanced cloud infrastructure, it also has to deal with energy insecurity, skills shortages, and uneven policy enforcement.

For example:

• Adoption maturity is unbalanced: 67% of large firms used GenAI in 2025, but only 14% had a formal strategy. Shadow AI use reached 32%, exposing companies to legal and reputational risk.

• Infrastructure is mixed: Hyperscale data centres support AI-ready workloads, and recent improvements in energy stability have reduced the frequency of rolling blackouts. However, the legacy of loadshedding has left many firms reliant on redundant systems and cost-heavy power workarounds, which continue to influence infrastructure design and investment decisions.

• Sector dynamics vary: Mining and finance are mature adopters, applying AI to predictive maintenance and fraud detection, while telecoms and retail focus on network optimisation and logistics.

• Skills gaps remain critical: AI roles outstrip supply. Graduate unemployment remains high, but technical skills are rare and concentrated, fuelling a wage spiral and a capacity bottleneck.

These challenges are not unique to South Africa, they mirror the AI readiness gap in many emerging markets, where adoption often outpaces governance and infrastructure modernisation.

As adoption accelerates globally, uneven readiness will shape who captures value, who gets left behind, and how public trust is either earned or lost.

AI in 2026 will be a test of leadership, not just systems. It will reward firms that treat autonomy, governance, and data readiness as interconnected priorities, not as separate domains.

Share

Subscribe now

A simple 3–5 minute evaluation would flag this risk early. An evaluation is simply a repeatable test of an AI model’s output assigned a score.

How you’ll actually use it

• Summaries: mark each one as Correct or Incorrect on the main fact.

• Headlines: give each a 1–3 score for clarity or punch.

Keep these as separate mini-tests. The aim is to move from “it feels okay” to “this setup scored 4/5 on my own examples.”

One Practical Test (3–5 Minutes)

Goal: Establish a quick Accuracy Rate. Metric: Percentage (%) correct.

1. Select Data: Pick five short items from the same source (e.g., five news articles).

2. Generate: Use a fixed prompt (do not change it mid-stream) to generate one-sentence summaries for each.

3. Evaluate: For each item, ask: “Is the main fact correct?”

4. Score: Mark it Correct or Incorrect.

5. Calculate: Count the “Correct” marks and divide by five.

Benchmark: A good starter result is around 80% or better across five items. For critical tasks, scale this up to 20+ items.

What to watch

Changing the prompt mid-test breaks the comparison. If you tweak the prompt, you must restart the test. That counts as a new “run”.

Take five old contracts where you know the key terms, run your summary prompt, and mark each as correct or incorrect on “Did it capture the main obligations and penalties?” If you get 2/5 or 3/5 correct, you know this setup is not safe to rely on. If you get 5/5 and keep logging results over time, you have a simple way to see whether your prompts and models are good enough for the work you want them to do.

I’ve put this setup into a simple Excel workbook with a Lab Notebook and a Prompt Changelog. Download it here and start scoring your prompts in a few minutes:

Takeaway

Score small, score often. Fix the prompt for each test, and always record dates, prompts, and scores. This allows you to see if your changes are actually improving the output, rather than just relying on “vibes”.

Subscribe now

Share