Who Gets to Think With the Machines?

The race to build artificial intelligence is creating a second contest over who will be allowed to use it, with a small group of technology companies and governments gaining control over the hardware, data, models and rules that determine access. Emad Mostaque, the founder of Stability AI, argues that unless capable systems remain widely available, AI could turn the promise of abundant intelligence into a new hierarchy of power.

That warning is easy to dismiss as another dramatic prediction from a technology entrepreneur. It should not be. The central question is no longer whether artificial intelligence will change work, education, politics and business. It is whether ordinary people will possess meaningful control over the systems changing those things, or whether they will receive intelligence as a managed service from institutions with interests of their own.

The distinction matters. A tool that can be used, examined and modified by millions of people distributes power. A tool that is available only through a handful of corporations and public authorities creates dependence. The same machine may produce abundance in one political arrangement and obedience in another.

Mostaque’s argument is that the world is moving towards the second arrangement. He describes an emerging “intelligence stack” stretching from computer chips and data centres to foundation models, applications and government policy. Each layer is expensive, technically demanding and increasingly concentrated. The companies that control the stack will not merely sell software. They will influence who can analyse information, automate decisions, coordinate organisations and compete in the economy.

That is why the debate about AI cannot be left to questions of model performance or product launches. The issue is not simply whether a system writes better code or produces more convincing images. The issue is who sets the limits, who owns the infrastructure, who sees the data, who can appeal a decision and who is excluded when access becomes inconvenient, expensive or politically suspect.

The internet was supposed to prevent this

The modern argument over AI resembles an earlier argument about the internet. The first public vision of the web was not a controlled marketplace but a great opening of doors. Information that had been held by universities, newspapers, libraries, publishers and governments would become available to anyone with a connection. Geography would lose its force. A student in a poor country could consult material once reserved for an elite institution. A small business could reach customers without buying access from a national broadcaster or newspaper chain.

For a time, the promise looked credible. Search engines made obscure information easy to find. Open publishing allowed specialists to speak directly to readers. Cheap digital tools enabled people with little capital to build communities, businesses and political movements. The internet’s most important achievement was not that it made communication faster. It reduced the number of people standing between a citizen and the knowledge he or she wanted.

That reduction in gatekeepers did not last. The web became dependent on a small number of platforms that control search, social distribution, online advertising, cloud infrastructure and app stores. Users may still encounter an enormous volume of information, but they rarely decide how that information is ranked, recommended or removed. Access is abundant in appearance and managed in practice.

Governments have added their own barriers. China’s Great Firewall demonstrated how a state could combine technical control with political censorship, allowing the internet to exist while preventing citizens from moving freely across it. Other governments have pursued less comprehensive but still significant restrictions, including age rules, platform duties and limits on particular categories of speech or application. The stated objectives are often protection and safety. The result is a larger role for authorities in deciding which forms of participation are permitted.

The lesson is not that every restriction is illegitimate. Children can be exposed to genuine harm online, and platforms can cause damage when they distribute fraud, abuse or incitement at scale. The lesson is more basic: once access to communication depends on a few institutions, those institutions acquire power over the conditions of public life. Measures introduced for one purpose can be expanded for another. A system designed to remove harmful material can become a system for deciding whose ideas are too risky to circulate.

Artificial intelligence enters this environment after the crucial infrastructure has already been concentrated. Unlike the early web, it is not arriving on a blank landscape of simple publishing tools. It is being built on vast data centres, specialised chips, proprietary datasets and closed models. The companies that own those assets are not merely hosting conversation. They are constructing the machinery through which future conversation, research and decision-making may occur.

Intelligence is becoming a managed service

The phrase “rationing of intelligence” sounds exaggerated until the economic structure of advanced AI is examined. The most capable systems require enormous quantities of computing power. Training a frontier model can demand hundreds of millions of pounds or dollars in hardware, electricity, engineering and data. Serving responses to millions of users creates another continuing bill. A model is not a book that can be printed once and distributed at low cost. It is an industrial system that must be trained, updated, secured and operated.

Those costs create a powerful barrier to entry. A university researcher may have an excellent idea but lack the computing resources to test it. A small company may develop a valuable application but remain dependent on an infrastructure provider that can change prices or terms. A public-interest group may want to inspect a model’s behaviour but be unable to obtain the weights, the training data or sufficient access to reproduce the results.

The result is a market in which the most useful cognitive capabilities are often accessed through an account rather than owned by the user. A person may ask a model to summarise a legal document, design a product, write software or prepare a business plan, but the underlying capability remains elsewhere. The user receives an output. The provider retains control of the system, the interface, the monitoring rules and, in many cases, the record of how the system was used.

That arrangement is convenient, but convenience can disguise dependence. A provider can impose usage limits, alter a model’s behaviour, withdraw an application, restrict a subject or block a region. It can decide that certain requests are allowed for one class of customer but not another. It can reserve the most powerful model for subscribers able to pay higher fees. It can suspend an account without offering a meaningful route of appeal.

The practical effect is not the disappearance of intelligence. It is the conversion of intelligence into a controlled utility. Access exists, but on terms set by the owner. The difference between having a tool and renting a service becomes important when the tool affects a person’s livelihood, education or political participation.

Mostaque’s concern is not that every company will behave badly every day. It is that the structure gives a small number of companies the ability to determine what ordinary people can do. If a model is used to evaluate job applications, provide medical information, design public services or explain political choices, the provider’s rules begin to shape opportunities far beyond the technology sector. A commercial decision about access can become a social decision about who receives assistance and who does not.

This is why the word “intelligence” changes the scale of the debate. Earlier digital platforms controlled distribution. Advanced AI may control not only distribution but analysis itself. Whoever owns the system can influence how information is interpreted, which options are presented and which forms of reasoning are treated as acceptable. The gatekeeper is no longer standing only at the entrance to the library. It may be helping write the catalogue, answer the questions and decide which books are worth opening.

The intelligence stack is a chain of dependence

Mostaque uses the idea of an intelligence stack to describe the layers that must work together before a powerful AI system reaches the public. At the bottom are semiconductors, specialist processors, energy supplies and data centres. Above them sit cloud platforms, storage and networking. Further up are datasets, training methods, model architectures and software interfaces. At the visible end are chatbots, search tools, workplace applications and automated services.

Control at any one layer can affect the whole system. A shortage of advanced chips can prevent a research team from training a model. Dependence on a cloud provider can expose a company to price changes or contractual restrictions. A closed model can make it impossible for users to examine how a result was produced. An application store can decide whether a tool is distributed. A regulator can impose rules that are affordable for a global corporation but ruinous for a small competitor.

The stack therefore creates more than a group of large companies. It creates a chain in which power at the lower levels can be converted into control at the higher levels. The firm that supplies computing does not need to write every application to influence the applications that can exist. The company that operates the model does not need to control every user directly to establish the conditions under which millions of users work.

The concentration is reinforced by scale. The more customers use a system, the more feedback and usage data its operator receives. The more data and revenue the operator has, the more it can spend on chips, engineers and distribution. The larger the user base, the more attractive it becomes to developers who want to build on the system. This is a familiar network effect, but applied to a technology that performs tasks once associated with professional judgement.

There is a political consequence. Citizens may be formally equal before the law while remaining unequal in access to the capabilities that determine practical power. A well-funded organisation can purchase several systems, hire specialists to integrate them and use them around the clock. An individual may receive a restricted consumer interface. Both may be said to have access to AI, but the phrase conceals a decisive difference in capacity.

The same imbalance can appear within government. Public agencies may use private models to assess claims, detect fraud, allocate resources or monitor communications. If the models are proprietary, the people affected may not know what data was used or how the decision was reached. An official can describe the system as a neutral tool, while the institution that selected it retains the power to define its purpose.

The danger is not limited to dramatic scenarios involving autonomous machines. It can arise through thousands of ordinary decisions made by systems no one outside the supplier can properly inspect. A model need not possess perfect knowledge to wield power. It only needs to be cheaper, faster and harder to challenge than the human process it replaces.

Regulation may protect the public — and the incumbents

The political response to AI has focused on safety. That focus is justified. Powerful systems can produce false information, expose private data, assist fraud and amplify dangerous activity. Companies need obligations, and the public needs protection from products released without adequate testing. The difficulty lies in the design of rules and in the distribution of the costs they impose.

Large firms can absorb extensive compliance requirements. They can hire legal teams, establish testing departments, maintain records, commission audits and negotiate with regulators. A small company, independent researcher or open-source project may not have those resources. A rule written to control the most powerful systems can therefore have the side effect of excluding smaller systems, even when those systems pose less risk.

This is the problem of regulatory capture in its most modern form. A powerful industry does not need to oppose regulation in public. It can support a framework that appears responsible while ensuring that the required infrastructure is too expensive for serious challengers. The public hears that the rules are protecting people. The market becomes less competitive. The incumbent’s position grows stronger.

The European Union’s AI Act illustrates the broader tension. Requirements for high-risk applications are intended to force providers to document systems, assess harms and maintain oversight. Those objectives are defensible. Yet any framework that creates significant legal and technical burdens will affect companies differently. A multinational with billions in revenue can spread the cost across a global business. A small developer may abandon a promising tool before it reaches users.

Safety policy also raises the question of who defines danger. Some risks are clear: fraud, exploitation and the exposure of personal information are not matters of taste. Other decisions are more subjective. A provider may restrict a model because a topic is controversial, because a regulator might object or because the company wants to avoid reputational damage. Over time, the category of “unsafe” can expand from conduct that causes measurable harm to ideas that create institutional discomfort.

There is an irony here. Regulators may seek to stop private companies from becoming too powerful, yet rely on those same companies for technical expertise and access to the systems they are trying to oversee. Governments may ask a handful of providers to define standards for the entire industry because those providers are the only organisations with the money and data to do the work. That arrangement gives the leading firms a voice in writing the rules that will govern their competitors.

None of this means that an entirely unregulated AI market would be desirable. It means that the public should distinguish between rules that reduce concrete harms and rules that simply make participation impossible for anyone without a massive balance sheet. A safety regime that preserves competition, scrutiny and user choice is different from a licensing system that turns advanced intelligence into a privilege granted to a few approved companies.

The distinction will determine whether regulation acts as a guardrail or a gate. Guardrails prevent a vehicle from leaving the road. Gates decide who is allowed to enter it.

Democracy cannot survive a permanent capability gap

Mostaque’s most severe warning concerns democracy. His argument begins with a modest assumption: democratic government requires citizens to possess enough information and practical ability to judge competing claims and hold leaders accountable. Perfect equality has never existed. Wealth, education, social position and access to institutions have always differed. But elections and public debate depend on a broad enough distribution of knowledge that power cannot be exercised without explanation.

Artificial intelligence could widen the gap. A political organisation with access to advanced systems could produce tailored messages for individual voters, identify emotional weaknesses, test thousands of arguments and respond to events within seconds. It could generate convincing text, audio and video at a volume no human campaign could match. The public would not merely be exposed to propaganda. Each person could receive a different political reality, designed around the information most likely to move him or her.

The wealthy could also use AI to multiply their human capacity. A company executive might have teams of automated researchers, negotiators, programmers and analysts working continuously. A wealthy individual could obtain advice across law, finance, medicine and communications that would once have required several professional firms. A low-income worker might have access only to a limited general-purpose assistant, subject to restrictions and unable to act on the user’s behalf.

Again, the danger is structural rather than cinematic. Democracy does not fail only when a dictator abolishes elections. It can become hollow when citizens retain the right to vote but lose the ability to understand the forces shaping their choices. Institutions can continue to operate while the real advantage belongs to organisations that control the information environment and the tools used to navigate it.

The experience of social media offers a warning, though AI will extend the mechanism. Recommendation systems already influence which stories people see, which arguments gain attention and which groups become targets of manipulation. AI can make the process more personal and more difficult to detect. A human propagandist can create a campaign. An automated system can adjust the campaign after every response, producing a separate strategy for each audience.

This does not make voters helpless or every campaign fraudulent. People are capable of rejecting manipulation, and public institutions can develop safeguards. But the burden becomes heavier when synthetic material is cheap, plentiful and tailored to the individual. The citizen is asked to verify a flood of claims while the organisation producing them has access to systems that can anticipate objections before they are raised.

Mostaque describes the result as a possible form of “technocratic feudalism”: democratic forms remain, but meaningful influence follows access to cognitive enhancement. The phrase is severe, yet it identifies a real political question. If some groups can reason, organise and persuade at a scale unavailable to everyone else, equal citizenship becomes a legal statement rather than a lived condition.

Open models offer a route out, but not a free one

The alternative proposed by Mostaque is open-source or open-weight AI that can be run, inspected and adapted beyond the control of a single corporation. The objective is not simply to make software cheaper. It is to ensure that people can retain a degree of independence from the companies that build the largest systems.

An open model can be studied by researchers, modified by developers and deployed by organisations with different priorities. It can allow a school, business or public body to keep sensitive material on its own infrastructure rather than sending every request to a remote provider. It can make it harder for one company to withdraw a capability from the market without warning.

The strongest version of the open argument is political. A public that relies entirely on closed systems is dependent on the institutions operating them. A public with access to systems that can be copied and examined has more room to challenge those institutions. Open models do not guarantee good outcomes, but they distribute the ability to experiment, verify and build alternatives.

The technical obstacles are substantial. The most capable models may be too large to run on ordinary computers. Training requires specialist hardware and expertise. Open release can make it easier for malicious actors to use a system for fraud, harassment or harmful research. A model that cannot be recalled after its weights are distributed presents a different governance problem from a model controlled by one provider.

Those concerns are legitimate, but they do not settle the argument in favour of permanent centralisation. Closed systems can also be misused. A company can make a harmful decision at scale, conceal weaknesses behind commercial secrecy or suffer a breach affecting millions of users. Centralised control does not eliminate risk. It changes who bears it and who has the power to respond.

The question should therefore be asked in practical terms. Which capabilities must be restricted, and why? What evidence shows that a particular release creates an unacceptable danger? Can access be limited in a targeted way rather than through a general ban? What independent bodies can examine the decision? What rights do users have when a system refuses service or makes a consequential error?

Mostaque’s position is that safety cannot become a universal argument for taking capability away from the public. That does not require releasing every model without conditions. It requires recognising that concentration is itself a risk. A society that prevents people from possessing powerful tools may reduce some forms of misuse while increasing the power of the institutions that possess them.

The best open-source strategy may therefore be neither reckless release nor total secrecy. It may involve distributed infrastructure, transparent testing, access tiers that are narrow and justified, and public investment in models that can be audited. The principle is that restrictions should be proportionate and contestable, not simply announced by the companies that benefit from them.

What if the machines we already hold in our hands could become the very bulwark against the rationing of intelligence that Emad Mostaque describes?

Consider the trajectory he sketches: a functional duopoly of frontier models, KYC-gated access, prompts stored forever, and the quiet possibility that one day you might have to prove your patriotism before an AI will even speak to you. Now place your own machine beside that picture. A MacBook Pro with 128 GB of unified memory, running local models through Ollama, generating text, images, and video without ever phoning home.

Does that device already embody a modest form of the “sovereign AI” Mostaque calls for—intelligence that no corporation or government can revoke at will? Or is it still only a temporary refuge, dependent on the continued goodwill of Apple’s silicon roadmap and the open-source weights that larger labs may one day be forbidden to release?

If capital no longer needs labour, as Mostaque warns, what happens when the same capital also no longer needs to sell you access to its models—because the models live entirely on devices it no longer controls? Could a future generation of Apple inference engines, powerful enough to run the next wave of open models at useful speed, shift the balance from “intelligence as a service” to “intelligence as property”? And if they did, would the rest of the industry follow, or would the regulatory and lobbying machinery he describes simply invent new barriers—hardware attestation, signed model licenses, mandatory cloud check-ins—precisely to keep capability centralized?

From Open Highways to Rationed Minds

The internet was once sold as the ultimate democratiser—a superhighway that would place the world’s knowledge in every hand and turn the World Wide Web into a new commons for livelihood and expression. Instead, Google and Meta captured the flow of attention, converting views into advertising fortunes while governments raised cross-national barriers and creators discovered that a single post could trigger demonetisation or regulatory exile. That broken promise now casts a longer shadow over artificial intelligence. Emad Mostaque’s warning that intelligence itself risks becoming a rationed commodity—controlled by a handful of labs, gated by KYC, and subject to political loyalty tests—echoes the same trajectory: open protocols giving way to concentrated platforms, democratic access yielding to gatekeepers. Yet a counter-current already exists in the form of high-memory personal machines capable of running local models that generate text, images and video without ever surrendering a query to the cloud. Whether these devices can preserve a genuine right to intelligence, or whether the same forces that captured the open web will simply move the choke-points to silicon and signed model licenses, remains the decisive contest of the coming decade. The pattern is clear; the window to break it is not.

 

The economic cost of dependence will arrive before the political crisis

The debate is often framed around a distant future in which machines exceed human intelligence. The economic effects of concentrated AI will appear earlier and in less dramatic form. Businesses will decide which providers to trust. Schools will decide whether to permit automated tutors. Public services will decide whether to outsource assessment and triage. Workers will use systems that can change their terms overnight.

In each case, the central issue is bargaining power. A company that depends on one AI provider may have little leverage if prices rise or performance changes. A software developer may build an entire business around an interface that is later restricted. A public agency may be unable to explain a decision because the supplier treats the model as confidential. A worker may be judged by an automated system without knowing how to challenge it.

These arrangements can create a false impression of efficiency. A service appears cheaper because the cost of dependence is not recorded. The organisation saves money on staff but loses control over the process. A school gains a powerful assistant but becomes vulnerable to changes in subscription terms. A public body accelerates administration but cannot independently verify the judgments being made.

The impact on labour will also be uneven. Workers who can combine expertise with advanced AI may become more productive and more valuable. Workers who receive only limited access may face automated competition without the same tools to defend themselves. The question will not simply be whether AI replaces jobs. It will be who gets the AI that complements human work and who is left competing against it.

This creates a familiar pattern in which a technology is described as available to everyone while the quality of access follows income. A free version may be enough for casual questions. The most capable version may require a subscription. Advanced automation may be reserved for enterprise customers. The difference is not merely convenience. It determines whether a user can perform a task at all.

The market may eventually lower costs, as it has done with many technologies. But lower prices do not automatically produce equal power. A cheap service can remain closed, monitored and subject to rules the user cannot negotiate. Nor does competition arise naturally when the underlying chips, data centres and models are controlled by a small number of firms.

For that reason, the policy debate should include ownership and access, not only safety. Public procurement can require portability and audit rights. Researchers can be given computing resources independent of commercial providers. Smaller developers can be protected from discriminatory access terms. Citizens can be told when an automated system has played a significant role in a decision and given a meaningful route to appeal.

These are not technical details. They are the ordinary protections that determine whether people remain participants in an economy or become customers of an invisible authority.

Who will set the limits of thought?

The most unsettling feature of the AI debate is that its language can make political choices sound inevitable. People speak of “the model” as though it were a natural force, “the algorithm” as though it had no owner and “safety” as though its meaning were self-evident. These phrases conceal decisions made by engineers, executives, lawyers and officials.

A model’s limits are chosen. Its training data is selected. Its refusals are designed. Its commercial tiers are priced. Its geographic availability is negotiated. Its monitoring systems are built. Its outputs are described as reliable or experimental according to the interests of the organisation providing them. Even the decision to make a model open or closed is a decision about the distribution of power.

This is where the comparison with the internet becomes useful. The open web did not disappear in one act. It was gradually enclosed by terms of service, platform dependence, advertising incentives, government pressure and the convenience of centralised applications. Users kept the appearance of freedom while losing control over the infrastructure on which that freedom depended.

AI could follow the same path at greater speed. The public may be told that the systems are too dangerous to release, too complex to inspect and too expensive for independent organisations to operate. Some of those claims will be true in particular cases. The danger lies in allowing them to become a general licence for central control.

Mostaque’s warning is therefore less about one company or one regulation than about a direction of travel. If the hardware, models and distribution channels remain concentrated, then every later decision will be made from a position of dependence. The public will be asked to trust institutions that have both superior capabilities and a financial interest in limiting alternatives.

A different path remains possible. Governments can support open research without treating every system as harmless. Regulators can target demonstrable risks while preventing rules from becoming barriers against competition. Companies can offer models that users can move between rather than locking them into one provider. Universities and civil-society groups can receive the computing power needed to test the claims made by industry.

The objective should not be to pretend that artificial intelligence is neutral. It is to make its power contestable. People should know who operates the system, what it can access, what its limits are and how to challenge its decisions. Organisations should not be able to describe a consequential automated judgment as an act of nature. A machine may produce the answer, but an institution chose the machine.

The question now is not whether intelligence will become abundant. It is abundant already in the laboratories and data centres of a small group of actors. The question is whether that abundance will be distributed as a public capability or rationed as a private privilege.

The next settlement will decide more than the technology industry

The choices being made now will outlast the first generation of chatbots and image systems. Once schools, companies and governments build their processes around a small number of AI providers, reversing the arrangement will become difficult. Skills will be tied to particular platforms. Records will be stored in proprietary formats. Public decisions will depend on systems that cannot be replaced without disruption.

That is why delay favours concentration. Every new dependency strengthens the firms already at the centre of the stack. Every closed integration makes migration more expensive. Every rule written around the capabilities of the largest providers makes their size appear necessary. The market can still change, but the cost of changing it rises with each year of institutional dependence.

The public should be sceptical of both extremes: the claim that AI will solve every problem and the claim that only a small priesthood can be trusted to operate it. The first is marketing. The second is a political programme. Both remove ordinary people from decisions that will govern their work, privacy and access to knowledge.

Mostaque is right to insist that the distribution of capability belongs at the centre of the discussion. A society does not preserve freedom merely by declaring that citizens may use a technology. It preserves freedom by ensuring that access is affordable, alternatives exist, decisions can be examined and power can be challenged.

The stakes are larger than competition between technology companies. If intelligence becomes a service controlled by a handful of corporations and governments, the public will not merely be buying software. It will be negotiating for access to the means of understanding and acting in the world. If open and distributed systems survive, people may retain the ability to build, inspect and adapt the tools on which their lives depend.

The contest is already under way. It will be settled through procurement rules, research funding, model releases, infrastructure ownership and the quiet terms of the applications people use each day. By the time the consequences are obvious, the decisive choices may have been made years earlier.

The future of AI will not be determined only by how intelligent the machines become. It will be determined by whether human beings are allowed to remain owners of the intelligence they create.