The industry measures progress along a single line: bigger model, more knowledge, another benchmark passed, another straight extension of the same ladder. Every serious institution I have worked in knows that intelligence does not actually work that way. We do not hand one individual origination, risk judgement and final sign-off and call it efficient. We call it a control failure waiting to happen. Yet the dominant model of artificial intelligence asks one system to propose, reason, calculate and certify its own answer, and calls the final pass self-critique. I think the more interesting question is not how much a model knows. It is whether the architecture around it resembles a governed institution or an unaccountable one.
Why the future of AI should resemble a well-governed institution, not an omniscient machine
The artificial intelligence industry has settled upon a seductive measure of progress. Build a larger model. Feed it more of the world’s knowledge. Give it more computing power. Demonstrate that it can answer questions which defeated its predecessor. Then repeat.
This has produced astonishing results. It has also smuggled in an assumption which deserves much more scrutiny: that the best form of artificial intelligence is a single, increasingly universal intelligence.
Perhaps it is not.
Most of us do not need an artificial mind capable of discussing the influence of Shakespeare on Picasso every time we ask a question about risk, regulation or liquidity. The capability is impressive. It is also largely irrelevant to the task. What we need is intelligence which understands the problem, knows its limits, uses reliable evidence and submits important conclusions to appropriate challenge.
That is not simply a question of efficiency. It is a question of governance.
Intelligence is not one thing
The present contest encourages us to think of intelligence as a single ladder. One model is cleverer than another because it performs better across an expanding collection of tests. The winner is the model nearest the top.
Human institutions do not work like that. A bank does not ask its most intellectually versatile director to make every lending decision, reconcile every payment, interpret every regulation and investigate every fraud alert. It creates different functions with different expertise. It separates decision from assurance. It defines authorities, preserves evidence and escalates matters which exceed local competence.
We do not regard this as an unfortunate substitute for having one omniscient individual. We call it good governance.
An effective institution is intelligent because specialist people and systems interact well. Its intelligence lies partly in what each participant knows, but also in how questions are routed, how evidence moves, how disagreement is surfaced and how decisions are challenged. Intelligence is a property of the network.
Artificial intelligence may be moving in the same direction. Instead of one vast model being used for everything, a smaller routing system can identify the nature of a task and direct it towards the appropriate specialist. That specialist can draw upon controlled information, use deterministic tools and submit its answer to another component for verification. A larger general-purpose model remains available when a genuinely novel or cross-domain problem arises.
The large model becomes the appellate court, not the magistrates’ court hearing every case.
The legitimacy ceiling
The case for this change is not only technical. Frontier artificial intelligence may encounter a legitimacy ceiling before it encounters a capability ceiling.
The companies building the largest models are asking society to accept an extraordinary concentration of power. A small number of private organisations determine what knowledge enters the models, what behaviour is permitted, how errors are corrected, which users receive access and how much physical resource is consumed. Increasingly, the same organisations suggest that their technology may displace a substantial amount of human work.
The usual response is economic. Productivity will rise. New work will appear. If displacement becomes sufficiently great, some form of universal income can distribute the abundance created by machines.
This may answer a question about consumption. It does not answer the question of human worth.
Work is not always rewarding and it is certainly not the sole source of identity. Nevertheless, it provides many people with agency, contribution, standing, relationships and evidence that they are needed. The OECD’s work on AI and job quality recognises that changes to autonomy, intensity and engagement matter alongside employment and wages. The International Labour Organization has similarly connected AI-driven surveillance and reduced autonomy with risks to workers’ mental and social wellbeing.
A society in which millions lose purposeful work and become dependent upon transfers from a machine economy owned by a few technology proprietors is not necessarily stable, however generous the transfer. People may reasonably ask why those who own the machines should also define the terms upon which everyone else retains an income.
Nobody wants to become a supplicant to Elon Musk, Sam Altman or any other technology magnate. This is not a judgement about the personal generosity of particular individuals. It is a judgement about power. Economic dependence should not rest upon the continuing wisdom or benevolence of people whom the public did not appoint and cannot remove.
The problem, therefore, is not only what the technology can do to us. It is the political and institutional settlement implied by its architecture.
Architecture is governance
A monolithic model concentrates capability. It also concentrates dependency.
Users cannot ordinarily know which parts of its vast training history shaped a conclusion. An error in the core model can affect every domain in which it operates. A change in policy made by its owner can alter the service received by millions of people and organisations. If the model becomes essential infrastructure, its proprietor acquires influence without assuming the public duties normally attached to such power.
Regulation can impose obligations from outside. It can require testing, disclosure, incident reporting and human oversight. Those protections matter, but external regulation has limits when the underlying system remains opaque and indivisible.
Governance should also be designed into the architecture.
A network of smaller specialist models offers a different possibility. Each component can have a defined purpose, an identifiable evidence base and an explicit boundary of competence. A prudential model need not make medical judgements. A conduct model need not interpret engineering tolerances. A verification model can be independent of the component which produced the original answer. Sensitive institutional knowledge can remain within the organisation rather than being submitted to a universal external system.
None of this means the models will be infallible. It means their failures can be located, understood and contained.
The analogy with financial governance is close. We separate the first line from independent risk oversight and internal audit because no single function should originate an activity, judge its own risks and provide final assurance. Yet we are remarkably willing to let one AI model propose, reason, calculate, challenge and certify its own answer. Calling the final pass a “self-critique” does not create independence.
Networked intelligence permits genuine separation of duties. One model can formulate. Another can test. A deterministic tool can calculate. A controlled source can establish the facts. A human can decide. The audit trail records not only the answer but how authority and evidence moved through the system.
That is a much richer meaning of human in the loop. It does not mean placing a tired employee at the end of an automated process and inviting them to approve a conclusion they cannot realistically reconstruct. It means preserving human authority over the design, the boundaries, the exceptions and the final decision.
Smaller does not mean parochial
There is an obvious objection. Risk does not respect professional boundaries. A pandemic becomes a credit event. A cultural shift changes consumer behaviour. A geopolitical dispute disrupts a technology supply chain. A piece of scientific research changes an insurance exposure. A model which knows only banking regulation may miss precisely the connection that matters.
That is why the choice is not between one universal model and a collection of isolated mechanical idiots.
The emerging technical field of model routing points towards a more flexible arrangement. Research presented through the Association for Computational Linguistics suggests that a capable router can improve collective performance as the pool of candidate models grows, including pools containing many relatively small open models. Other work is exploring small models equipped with tools and multi-agent collaboration to compensate for the limitations of any individual component.
The result could be specialist by default and generalist by exception. Routine questions remain within bounded, auditable systems. Unusual questions are escalated. Several specialists can be assembled for a cross-domain problem. A frontier model can be called where its breadth adds genuine value.
That resembles a well-run organisation. Most matters are handled by people who understand them. Difficult matters draw together several disciplines. Only exceptional questions reach the board.
The architecture will introduce new risks. A poor router may send a question to the wrong expert. Models may pass errors to one another. Responsibility can become fragmented. Several apparently independent specialists may share the same flawed training material. The network may become so complicated that nobody understands the whole.
These are governance problems, not reasons to abandon the approach. They require ownership, testing, common evidence standards and clear escalation. The alternative, placing every capability and every hidden dependency inside one enormous model, does not eliminate complexity. It conceals it.
The physical world will have a vote
The drive towards smaller and more selective systems will also be forced by the physical world.
The International Energy Agency now expects electricity consumption from data centres to increase from about 485 terawatt-hours in 2025 to approximately 950 terawatt-hours in 2030. Consumption by AI-focused data centres is expected to triple. At the same time, power use for each individual AI task is falling rapidly as the technology becomes more efficient.
That apparent contradiction matters. Efficiency per task will not be enough if the number and complexity of tasks grow even faster. Nor will a global percentage reassure the communities in which demand is concentrated. A data centre competes for a particular grid connection, a particular water supply and a particular piece of land.
The Local Government Association has called for verifiable, site-specific information about energy, water, emissions and land impacts so that councils can assess proposed developments against local needs. This is an early indication of a much larger argument. Communities will not indefinitely accept the diversion of scarce resources on the assertion that ever larger models are intrinsically in the national interest.
Necessity will drive improvement. Models will become smaller. Computation will become sparse, activating only the capabilities required for a task. More inference will happen on local devices. Specialist chips will replace general-purpose equipment for some functions. Stored knowledge and retrieval will reduce the need to reproduce everything through repeated computation.
The future AI system may consume less resource for any given act of intelligence while becoming more widely distributed. That is likely to be economically attractive. It may also be socially easier to accept because the relationship between purpose, benefit and resource use becomes more visible.
From artificial supremacy to useful intelligence
The language surrounding frontier AI often treats human equivalence, and then human superiority, as the destination. Each benchmark passed becomes evidence that another human capability has been conquered. This is commercially effective theatre. It is psychologically and politically corrosive.
People do not object to a calculator being better at arithmetic. They do not experience an aircraft as a judgement upon the inadequacy of their arms. Those machines perform bounded functions whose purpose remains intelligible and whose authority remains subordinate to human ends.
A purported universal intelligence is different. It presents itself not as a tool which extends a capability but as a more capable participant across the whole field of human thought. The industry’s language then invites people to imagine their economic and intellectual redundancy.
Some of the fear is exaggerated. The ILO’s assessment of generative AI and jobs concludes that transformation is more likely than wholesale replacement for most occupations because human input remains necessary. But reassurance based only upon employment forecasts misses the deeper concern. A person can retain a job while losing discretion, mastery and pride in the work.
The better ambition is not artificial supremacy. It is useful intelligence under legitimate authority.
That changes the questions we ask. Instead of asking how much a model knows, we ask whether it knows what this decision requires. Instead of admiring its confidence, we test whether it recognises uncertainty. Instead of treating breadth as an unqualified virtue, we decide when breadth is relevant. Instead of trusting one system to challenge itself, we construct independent challenge. Instead of adding a nominal human checkpoint, we identify the judgement which must remain human.
This is not resistance to innovation. It is the difference between innovation which demands that society reorganise itself around the machine and innovation designed around human purposes.
Governed intelligence
The largest general-purpose models will not disappear. They will remain valuable for research, unfamiliar problems, cross-domain synthesis and the development of smaller models. Their role may nevertheless become narrower than their builders currently assume.
The everyday unit of artificial intelligence could become a governed network: specialist models, trusted evidence, explicit permissions, independent verification and intelligent escalation. No component needs to know everything. No component should be authorised to do everything.
This approach is less dramatic than the creation of a universal artificial mind. It is also more consistent with everything we have learnt about resilient human institutions. Concentrated intelligence creates concentrated dependency. Distributed expertise, properly connected and governed, creates capacity without surrendering all authority to a single point.
The fundamental choice is not between accepting AI and rejecting it. Nor is it simply between large and small models. It is between two conceptions of intelligence.
One seeks a machine which knows everything and asks us to trust its owners. The other assembles the knowledge needed for a purpose, exposes the evidence, preserves challenge and leaves authority where it can be held accountable.
We should stop asking whether artificial intelligence can become omniscient. We should ask why we would want it to be.
Richard Anderson
Founder, AndersonRisk
This essay was first published on 26 September 2026 on my Substack.