Against the Regulation of Artificial Intelligence

By Matthew Parish

Wednesday 23 September 2026

There is a familiar pattern in the history of technological change. Something new appears. At first it is regarded as a curiosity, then as an opportunity, then as a threat. Politicians discover it shortly thereafter and, having discovered it, announce that it must be regulated. Artificial intelligence has travelled through these stages with exceptional speed. Governments that scarcely understood large language models three years ago are now constructing elaborate administrative architectures intended to govern them.

This instinct should be resisted. Artificial intelligence is undoubtedly dangerous. It will disrupt professions, destroy established business models, facilitate fraud, transform warfare, alter education and permit forms of surveillance that would have astonished the architects of the twentieth-century police state. Yet none of these propositions establishes the case for regulating artificial intelligence as a technology. They establish the case for applying ordinary law to harmful conduct undertaken using artificial intelligence.

That distinction is fundamental. Murder committed using an autonomous machine remains murder. Fraud conducted using synthetic voices remains fraud. Defamation generated by a language model remains capable of being defamatory. An unsafe medical product does not become safe merely because its defects were designed by an algorithm rather than by an engineer. The law already possesses an enormous armoury of principles governing liability, negligence, contract, intellectual property, privacy, competition, crime and consumer protection. Where genuine gaps appear, those gaps can be addressed individually. There is no compelling reason to construct an overarching regulatory state around the technology itself.

Indeed there are powerful reasons not to do so.

Regulating what we do not understand

The first difficulty is embarrassingly simple. Governments do not know what artificial intelligence will become.

Nobody does. Contemporary frontier models are evolving at a speed profoundly different from that of legislation. A statute may require several years of consultation, drafting, negotiation, enactment and implementation. During precisely that period an entire generation of artificial intelligence architecture may appear and disappear.

The European Union illustrates the problem. Its AI Act entered into force in August 2024 and much of it became applicable on 2 August 2026, although significant requirements governing high-risk systems will not apply until 2027 or 2028. The legislation establishes obligations concerning such matters as documentation, risk assessment, human oversight, transparency, robustness and general-purpose artificial intelligence.

Yet the European Union has already had to alter the regulatory timetable. Its 2026 AI Omnibus extended implementation periods and introduced simplifications intended, according to the European Commission itself, to reduce compliance burdens and encourage competitiveness and innovation. It demonstrates something more interesting: regulating a technology changing this quickly inevitably involves aiming legislation at a moving target.

The more detailed the legislation becomes, the worse this problem becomes. Parliament legislates for the artificial intelligence systems it knows. Engineers develop the artificial intelligence systems that do not yet exist.

Regulation protects incumbents

There is another uncomfortable truth about regulation. Large companies can afford it.

A frontier artificial intelligence company employing thousands of people can hire compliance departments, external lawyers, auditors, safety specialists and former regulators. A group of four graduate students working in a university laboratory cannot. Every compulsory risk assessment, registration procedure, conformity document, reporting obligation and regulatory filing therefore has an economic consequence irrespective of its intention. It raises the price of entry.

This matters particularly in artificial intelligence because enormous concentrations of economic power are already emerging. Training frontier models requires extraordinary quantities of capital, computing power, electricity, data and specialised engineering talent. Governments should be concerned about this concentration. Yet imposing substantial regulatory costs upon the industry may perversely reinforce precisely the companies already powerful enough to absorb them.

The result could become regulatory oligopoly: a handful of enormous artificial intelligence companies sufficiently wealthy to satisfy governments, surrounded by a bureaucratic wall that prevents smaller competitors from challenging them. That would be an extraordinary outcome. Governments frightened of excessive technological power would have enacted legislation helping to consolidate it.

Permissionless innovation

Much of the modern world exists because previous generations were permitted to experiment before governments understood what they were doing. The early internet was chaotic. It contained pornography, fraud, extremism, copyright infringement, criminal marketplaces and endless quantities of nonsense. It also produced search engines, electronic commerce, digital publishing, online education, instant global communications and eventually much of the infrastructure upon which contemporary civilisation depends.

Had every internet service required regulatory approval before deployment, many of those innovations might never have survived infancy. The people who create revolutionary technologies are rarely the people best equipped to navigate ministries.

Artificial intelligence requires the same presumption of permissionless innovation. Someone should be allowed to build an algorithm because he has an idea, not because a government department has concluded that the idea falls within an approved risk category.

The presumption should therefore be liberty followed by responsibility. Build what you wish. Experiment. Publish. Compete. But if what you build injures somebody in a manner recognised by law, responsibility follows. That principle is neither radical nor anarchic. It is approximately how liberal societies have traditionally treated human ingenuity.

Regulate conduct, not intelligence

The strongest objection to artificial intelligence regulation is conceptual. Artificial intelligence is not really an industry. It is a general-purpose method of manipulating information.

Artificial intelligence can write a poem, identify a tumour, guide a drone, reconcile accounts, translate Ukrainian into Japanese, generate computer code, analyse satellite imagery, compose music or answer a legal question. The things described by the expression “artificial intelligence” have increasingly little in common except that sophisticated computational techniques lie somewhere behind them.

Regulating artificial intelligence therefore begins to resemble regulating mathematics. The appropriate legal question is not whether artificial intelligence was involved. It is what happened.

Suppose an artificial intelligence system discriminates unlawfully when processing employment applications. Employment discrimination law should govern the conduct. Suppose an autonomous vehicle negligently kills a pedestrian. Product liability, negligence and road traffic law should determine responsibility. Suppose somebody uses generative artificial intelligence to impersonate a bank employee and steal money. Criminal law should punish the fraud.

The machine does not make these harms conceptually novel merely because silicon participated in producing them.

There will undoubtedly be genuinely novel cases in which existing doctrine proves inadequate. Courts and legislatures can address those cases as they arise. Common-law legal systems have spent centuries adapting general principles to technologies their authors could never have imagined. We do not need an Artificial Intelligence Ministry merely because tort law occasionally requires development.

The international problem

International regulation is potentially still more problematic. The Council of Europe opened its Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law for signature in September 2024. The European Union ratified it in May 2026, while the Convention requires five ratifications, including three Council of Europe member states, before entering into force. The ambition behind such instruments is understandable: digital technologies cross borders, so their governance appears naturally to require international coordination.

But artificial intelligence is not merely a consumer technology. It is rapidly becoming an instrument of national power. It will affect intelligence analysis, cyberwarfare, military logistics, autonomous weapons, scientific discovery, industrial productivity and economic growth. States therefore have radically different incentives concerning its development. An international regulatory arrangement respected scrupulously by liberal democracies but ignored, interpreted creatively or secretly circumvented by their strategic competitors would not create global safety. It would create asymmetric restraint.

Nor is meaningful global enforcement particularly plausible. Algorithms travel easily. Model weights can cross frontiers far more readily than factories. Research can migrate between jurisdictions. Computing facilities can be relocated. The more burdensome regulation becomes in one jurisdiction, the stronger the incentive for research to move elsewhere. International artificial intelligence regulation therefore confronts an elementary geopolitical difficulty. The technology is global; political authority remains territorial.

Safety without regulation

Opponents of heavy regulation are sometimes caricatured as believing artificial intelligence presents no risks. That is unnecessary and implausible. The risks may be enormous. But regulation and safety are not synonyms.

Companies have powerful incentives to make systems reliable because customers prefer reliable products. Insurers can require security standards. Corporate purchasers can demand independent testing. Universities can develop evaluation methodologies. Professional associations can establish standards. Courts can impose liability. Governments can procure only systems satisfying specified safety requirements. Defence establishments can impose extremely demanding standards upon military artificial intelligence without purporting to regulate every civilian model.

Reputation matters as well. An artificial intelligence company whose products routinely reveal confidential information, fabricate disastrous financial advice or permit trivial cyber intrusion will discover that markets can impose sanctions considerably faster than legislation.

None of these mechanisms is perfect. Neither is government.

Regulators suffer from informational limitations, political pressure, bureaucratic incentives and regulatory capture. They can make mistakes just as engineers can. The difference is that a defective product may be displaced by a competitor, whereas a defective regulation can become compulsory throughout an entire jurisdiction.

The precautionary principle turned upside down

Much artificial intelligence regulation is ultimately founded upon the precautionary principle: where the consequences of a technology may be catastrophic, society should act before catastrophe occurs. There is superficial wisdom in this. But precaution has two directions. Regulation itself creates risks.

Artificial intelligence may accelerate pharmaceutical research, improve medical diagnosis, optimise electricity networks, develop new materials, increase agricultural efficiency, improve weather forecasting and expand access to education. Nobody knows how valuable these developments will become because many have not yet been invented.

Every regulatory obstacle therefore carries an invisible opportunity cost. A system prevented from being developed cannot cure anybody. A company that never exists employs nobody. An experiment prohibited in advance produces no discoveries. The question is consequently not whether unregulated artificial intelligence is risky. Of course it is. The question is whether governments possess enough knowledge to determine in advance which experiments society can safely afford to prevent. History supplies little reason for confidence.

The danger of regulatory censorship

Artificial intelligence regulation also creates a less discussed political danger. Language models mediate information. Once governments acquire authority to determine what artificial intelligence systems may say, what training material they may use, which answers are considered harmful and which models may be distributed, regulation approaches control over knowledge itself. The intentions may initially be admirable. Governments may wish to suppress fraud, terrorist propaganda, dangerous medical misinformation or synthetic political deception. Yet regulatory powers rarely remain confined forever to the controversies that justified their creation.

A government empowered to define unacceptable artificial intelligence output possesses an extraordinarily powerful instrument. Future governments may have different conceptions of what is unacceptable. Liberal societies should therefore hesitate before constructing administrative machinery capable of supervising the intellectual output of machines that may become humanity’s principal interface with accumulated knowledge. The danger is not that every regulator will become a censor. It is that we would have constructed the machinery with which censorship could efficiently be undertaken.

Artificial intelligence as speech and thought

There is also a deeper philosophical difficulty. Artificial intelligence increasingly occupies territory historically associated with intellectual freedom. Writing software is an expressive activity. Training models is a form of research. Publishing weights may be understood as disseminating information. Prompting a language model is increasingly analogous to consulting a library, research assistant or interlocutor.

The closer artificial intelligence moves towards being infrastructure for human thought, the more uncomfortable prior regulatory control should become. Nobody needs a licence to write an equation. Nobody should ordinarily require government authorisation to publish an argument. We do not submit novels for safety certification before publication.

Artificial intelligence complicates these categories because software can act as well as speak. That distinction justifies regulation of particular actions where necessary. It does not justify placing cognition itself beneath administrative supervision.

Let the revolution happen

Artificial intelligence may prove to be the most consequential technology since electricity. If so, attempting to regulate its development comprehensively at this stage is not merely premature. It risks freezing into law the assumptions of the technology’s infancy.

Governments should prosecute crimes committed with artificial intelligence. Courts should compensate people unlawfully injured by it. Existing rules governing medicine, aviation, banking, weapons, privacy and other inherently sensitive activities should continue to apply when artificial intelligence enters those fields. Legislatures should repair identifiable holes in existing law where experience demonstrates that they exist. But that is regulation of activities, not regulation of artificial intelligence itself.

The distinction should become the organising principle of public policy. No artificial intelligence licensing agencies. No international authority deciding which models humanity may build. No general requirement that researchers demonstrate the safety of ideas before testing them. No assumption that technological novelty transfers the burden of proof from the state to the inventor.

The burden should run in the opposite direction. If government wishes to prohibit something, government should identify a concrete harm and justify the prohibition. Artificial intelligence will make mistakes. It will be abused. It will eliminate jobs and create others. It will generate fortunes, bankrupt established businesses and embarrass governments. It will undoubtedly produce consequences we shall regret.

That is what transformative technologies do. The alternative is to imagine that committees of legislators, diplomats and civil servants can anticipate the course of an intellectual revolution whose destination is unknown even to the people building it. They cannot. Attempts to do so may produce an immense regulatory architecture whose principal achievements are protecting incumbents, exporting innovation, suppressing experimentation and providing governments with new powers over information.

Humanity is entering an extraordinary experiment. There will be accidents along the way, and ordinary law must deal firmly with them. But experiments are how civilisation advances.

We should regulate harms. We should punish crimes. We should enforce responsibility.

We should not regulate intelligence.