Part I — AI Arrived Halfway Through the Race
There are plenty of frightening stories about artificial intelligence. Perhaps one day a sufficiently capable model will help somebody design a virus that should never have existed, make chemical or biological weapons easier to create, penetrate critical infrastructure, or become autonomous enough for us to discover that handing important decisions to something we do not entirely understand was not one of humanity’s better ideas. These are serious possibilities, and the people working on AI safety are right to take them seriously.
Yet I increasingly wonder whether we are looking for the AI catastrophe in the wrong place. We naturally fear the spectacular event, because spectacular events are easy to recognise. There is a laboratory accident, an attack, a runaway machine, a date future historians can put in a textbook and say: this is when everything changed.
There is another possibility in which nothing like that happens. Technology simply gets better. Companies become more capable. Individuals acquire astonishing tools. Productivity rises in places where the technology can be absorbed. And underneath all this progress, some of the institutions on which society depends become relatively less capable of understanding and managing the world around them.
AI may not have to turn against us. It may be enough for it to accelerate some parts of civilisation until the rest can no longer follow.
The strange thing is that this story did not begin with AI. I first noticed a primitive version of it when I entered the IT industry at the end of the 1970s.
Commercial computing was already a world in which knowledge aged quickly. Hardware changed, operating systems changed, programming techniques changed, database technology developed, networks appeared and businesses gradually discovered that computers could do much more than automate accounting. You learned because you had to. Much of what mattered professionally came from colleagues, manuals, customers, experimentation and occasionally from discovering painfully that the elegant solution you had designed did not survive contact with reality.
What puzzled me even then was the distance between this world and parts of the educational system supposedly preparing people to enter it. Universities made fundamental contributions to computer science, so this is certainly not an argument that academics knew nothing about computing. But conversations with professors sometimes gave me the peculiar impression that they understood the formal subject while having surprisingly little feeling for how quickly commercial IT itself was changing or which skills organisations actually needed. The educational machinery moved at a different speed from the environment into which it released its students.
For a long time this was merely irritating. Companies trained people themselves, employees taught themselves, educational institutions caught up eventually and another generation of technology arrived before they had quite finished doing so.
What I failed to appreciate at the time was that something much larger was being built underneath all this.
During the following decades, many private organisations slowly constructed what we would now recognise as a digital nervous system. Accounting systems became connected to operations. ERP linked activities that had previously existed in separate administrative islands. CRM accumulated information about customers. Supply-chain systems connected production, inventory, warehouses and suppliers. Banks turned transactions into information almost as soon as they occurred. Supermarkets knew what had been sold before the shopper had reached the car park. Logistics companies could eventually follow almost every parcel. Websites stopped merely presenting information and began observing what customers actually did.
It was messy, expensive and frequently incompetent. Anyone who has worked inside a large corporation knows how much mythical perfection is contained in the phrase integrated enterprise. Banks still depend on software written decades ago, ERP projects consume fortunes and sometimes achieve remarkably little, data remain duplicated and senior management is perfectly capable of ignoring what its own systems are telling it.
But there was a pressure underneath all of this which mattered enormously. Competitors existed. Customers could leave. Costs eventually appeared in a profit-and-loss account. If another organisation found a dramatically better way of doing something, ignoring it forever became difficult.
Over time, the distance between something happening in reality and the organisation knowing that it had happened became shorter.
That may be one of the least appreciated consequences of the IT revolution. We did not merely replace typewriters with computers. We slowly connected organisations to the reality in which they operated.
Public administration digitised too, sometimes on an enormous scale. Governments possess extraordinary quantities of information. Tax systems, population registers, social security, healthcare, education, land registries, police and justice together hold datasets that would make most corporations envious. So the argument cannot simply be that companies have data and governments do not.
The real difference is subtler. It lies in what those data are for, how quickly they move, whether different parts of the institution can use them, and whether they eventually influence what the institution actually does.
Much public information was collected because a law required it, because somebody needed to process a case, because a department had a reporting obligation or because statistics had to be produced. Systems were developed at different times for different purposes. Definitions differed, access was restricted, old applications survived because replacement was dangerous, and information travelled upward through organisational layers where reality was progressively transformed into categories, averages, targets, reports and finally little red and green squares on management dashboards.
The OECD now describes much the same problem in rather more diplomatic language. Governments have built substantial digital infrastructure, but the components often still fail to operate as an integrated whole, while fragmented responsibilities, legacy technology and weak feedback mechanisms limit the ability to turn information into better services and decisions.
That distinction between possessing data and possessing feedback is crucial. An institution can own terabytes of information and still have remarkably little idea what is happening.
I have written about this before as part of what I call the Babel Syndrome: the strange tendency of large organisations to create increasingly elaborate conceptual representations of reality while progressively weakening their direct connection with reality itself. The abstraction is necessary at first; nobody running a country can personally inspect every classroom, courtroom or hospital. The problem begins when the model stops being a compressed representation of reality and quietly becomes a substitute for it. The Babel Syndrome, Part 1
The teacher knows what happens in the classroom, the clerk knows precisely where citizens become trapped in the procedure, the doctor sees the patient, the police officer sees the street and the social worker sees the family sitting opposite. As that experience travels upward it becomes information, then categories, then indicators and eventually policy. By the time the signal arrives at the top, much of the texture that made it useful has disappeared.
This is not uniquely a public-sector disease. Large private organisations suffer from it too. The difference is that markets sometimes deliver rather brutal reminders that the model has stopped corresponding with reality.
Government is less fortunate in that respect. A tax administration does not lose customers to another tax administration. A court does not discover that litigants have switched en masse to a more convenient competing justice system. A ministry does not go bankrupt because its information architecture is twenty years out of date. Public organisations therefore have to construct their own feedback mechanisms deliberately, because one of the strongest feedback mechanisms available to a commercial organisation simply does not exist.
Too often, we built something else instead.
We computerised the procedure.
One small example from Belgian justice captures this almost perfectly. Electronic filing sounds like one of the least controversial forms of digitalisation imaginable: instead of lawyers carrying paper documents to a court, the documents arrive electronically. Yet electronic submissions have in some circumstances still needed to become paper somewhere inside the court process, to the point where printing itself became an administrative burden.
The PDF arrived.
The organisation had not.
The printer had merely moved.
The joke hides something important. For decades digital government often meant taking an existing paper process and building technology around it. The paper form became an electronic form, the counter became a portal, the filing cabinet became document management and the signature acquired a cryptographic equivalent. The technology changed while the institutional logic survived remarkably intact.
Public institutions have a perfectly legitimate defence. They cannot behave like startups, nor should they. Government must respect due process, equality before the law, continuity, democratic accountability, privacy and rights that a commercial organisation can sometimes simply price out of its business model. Some friction is not evidence of failure. It is one of the prices we deliberately pay for living under law rather than under the whim of whoever happens to run the organisation.
But that explanation becomes less satisfactory when extended over half a century.
The mainframe revolution was followed by personal computers, relational databases, networks, the internet, the web, enterprise software, mobile computing, cloud systems, APIs and modern data analytics. Each wave created another opportunity to reconsider how information moved and how institutions worked. Yet again and again the new layer was attached to the old structure.
Necessary caution slowly became difficult to distinguish from institutional inertia.
A court has good reasons not to reinvent the rules of justice every six months. It does not need thirty years to discover that printing electronic documents is not digital transformation.
The distinction matters because caution and paralysis are not the same thing.
This becomes more serious when we consider what private organisations were accumulating during those same decades. An experienced IT architect might distinguish between systems that record what happened, systems that allow an organisation to observe what is happening, and systems that help decide what should happen next. The first great wave of enterprise computing created systems of record. The most capable organisations then built increasingly sophisticated systems of observation. They instrumented the business itself.
And now AI is beginning to provide the third layer.
It arrives as a system of intelligence.
This is why treating November 2022 as the starting line of an AI race is so misleading. Companies, governments, schools, courts, NGOs and universities were nowhere near the same starting line. AI landed on top of fifty years of accumulated information architecture, organisational habits, feedback mechanisms and technical debt.
It arrived halfway through the race.
Give a capable AI access to live sales, customer behaviour, production, inventory, complaints, supplier performance, logistics and finance and you have given it something resembling sensory perception. Give exactly the same model annual statistics, PDFs, incompatible databases, policy documents and reports produced months after the events they describe and you have created a brilliant intelligence staring through fog.
The model may be identical. The intelligence available to the organisation is not.
This is one reason why simple figures about “AI adoption” tell us less than they appear to. Stanford’s 2026 AI Index reports AI use in 88 percent of surveyed organisations. Governments are adopting it too. But the deeper question is whether AI sits on top of an organisation already capable of observing reality and feeding what it learns back into action. Stanford AI Index 2026 — Economy
A capable private organisation increasingly tries to observe what happened, understand it, change something, measure the result and adjust again. AI can reduce the cost and time of almost every stage of that cycle. Analysis becomes faster, software becomes easier to modify, experiments become cheaper and organisations capable of learning can therefore learn faster.
An institution in which recognising a problem produces a report, the report produces consultation, consultation produces policy, policy waits for a budget, the budget requires procurement, procurement produces implementation and an evaluation arrives several years later lives on another clock.
For decades the difference between those clocks was tolerable.
AI may make it compound.
The irony is that bureaucracy can appear to adopt AI very successfully while missing this point completely. Artificial intelligence is spectacularly good at processing the products bureaucracies already create. It can summarise reports, draft policies, classify documents, prepare presentations, digest consultation responses and generate beautiful explanations of why a programme is progressing according to plan.
Which raises an uncomfortable possibility.
AI may allow an institution to become far more articulate without becoming any more connected to reality.
The private company may use AI to remove bureaucracy. The bureaucracy may use AI to automate bureaucracy. Both can announce impressive productivity gains.
The difference is that one has shortened the distance between reality and action while the other has merely accelerated the circulation of abstractions.
I explored this problem earlier in Reality Has Entered the Chat, where the central issue was that increasingly adaptive societies are talking back to institutions still organised largely around command, specification and control. AI makes that mismatch considerably more consequential. Reality Has Entered the Chat
And this is where the story becomes larger than technology.
Because while institutions spent fifty years accumulating their technical and organisational differences, the world they were supposed to govern was becoming something else entirely.
It was becoming vastly more complex.
That is where Part II begins.