Is a law of artificial intelligence arriving, or only a patchwork?

On 6 October 2025 the European Commission published a short notice in the Official Journal, reference C/2025/5423, withdrawing a proposal it had made three years earlier: COM(2022) 496, the Artificial Intelligence Liability Directive. The stated reason was that no agreement among the institutions was foreseeable. The Commission added that it would consider whether to bring forward a revised proposal or take some other approach — the form of words used when nobody wants to say never.

That withdrawn directive was the only instrument in the European pipeline addressed to the question most people actually have about artificial intelligence: whom do you sue when one of these systems ruins your life? The AI Act, for all its length, is a market-access regulation. It tells a manufacturer how a system must be built and documented before it may be sold. About what happens afterwards — when a hiring model has quietly excluded a class of applicants for two years, or a clinical tool has been confidently wrong — it says almost nothing.

So two questions run at different speeds. The first is regulatory and much discussed: which capabilities should be permitted, under what conditions, with what paperwork. The second is jurisprudential and barely discussed: what kind of legal object an AI system is. Law is an apparatus for attributing conduct to agents. It wants someone who acted, a state of mind, a duty, a breach. Put into the middle of it a system that produces outputs nobody intended and nobody can fully explain, and the machinery does not break so much as slip.

The instrument that survived treats software as a thing you can buy#

What Europe has instead of an AI liability regime is a rewritten product liability regime. Directive (EU) 2024/2853 was adopted on 23 October 2024, published in the Official Journal on 18 November and in force from 8 December; it applies to products placed on the market after 9 December 2026, and it repeals the 1985 directive.

The decisive move is in Article 4(1). A product, the directive says, "means all movables, even if integrated into, or inter-connected with, another movable or an immovable; it includes electricity, digital manufacturing files, raw materials and software." Software. That single word drags trained models, and the services built on them, inside a strict liability regime — strict meaning that a claimant need not prove anyone was careless, only that the product was defective and the defect caused the damage.

Two further provisions matter more than their drafting suggests. Article 9 lets a court order a defendant to disclose evidence once the claimant has made the claim plausible. Article 10 stacks presumptions on top, the most striking being that a court may presume defectiveness or causation where the claimant faces excessive difficulties, in particular from technical or scientific complexity, and the claim is nonetheless likely.

That is a quiet admission. Legislators looked at the prospect of an ordinary claimant explaining a transformer model to a trial judge, decided it was hopeless, and shifted the burden. That is a sensible response to opacity. It also settles who loses when nobody can tell, which is a different thing from settling who is responsible.

The regime remains a product regime, built for injury and damage to property. The harms that dominate complaints about AI — the refused mortgage, the mistaken identification at a border, the reputation dismantled by a fabricated quotation — sit awkwardly outside a claim in defective goods. Closing that mismatch was what the withdrawn directive was for. It is now closed by twenty-seven national tort systems, each with its own answers about fault, causation and standing.

Not a person, and not quite a thing either#

The European Parliament once had a bolder idea. Its resolution of 16 February 2017 on Civil Law Rules on Robotics (2015/2103(INL)) asked the Commission to consider, at paragraph 59(f), "creating a specific legal status for robots in the long run, so that at least the most sophisticated autonomous robots could be established as having the status of electronic persons responsible for making good any damage they may cause".

The reaction was brutal. An open letter to the Commission, signed by more than 150 specialists in robotics, artificial intelligence, law, medicine and ethics, called a legal status of electronic person "ideological and nonsensical and non-pragmatic": it could not be derived from the natural-person model without absurdity — a robot with human rights — nor from the corporate model without becoming a device for evading liability. By the time Parliament returned to the subject, in its resolution of 20 October 2020 on a civil liability regime for artificial intelligence, the language had inverted. Any reform, the text now said, "should start with the clarification that AI-systems have neither legal personality nor human conscience, and that their sole task is to serve humanity". That is not a sentence a confident draftsman writes.

Still, the corporate analogy deserves better than the dismissal it got, because it is instructive about what legal personality is for. A company has personality without consciousness; nobody supposes Unilever has an inner life. Personality is a bookkeeping device, and it works because behind the abstraction stand directors who can be disqualified, shareholders who can lose money and a balance sheet that can be reached. Grant personality to a model and you have an entity with a docket number and no assets — the liability shield the letter's signatories feared.

So the operative answer, across every jurisdiction that has legislated, is the unglamorous one. An AI system is regulated as a product, with duties attached to the humans arranged around it: provider, deployer, importer, distributor.

Four regimes, and only one of them is a statute about AI as such#

Europe's is the statute. Regulation (EU) 2024/1689 was published on 12 July 2024 and in force from 1 August. Prohibited practices and the AI-literacy duty began on 2 February 2025; obligations for general-purpose models, governance and penalties on 2 August 2025; general application on 2 August 2026, six weeks ago.

Then the timetable moved. The Commission tabled its Digital Omnibus on 19 November 2025 and the institutions reached provisional agreement in the first week of May 2026. Regulation (EU) 2026/1744 was adopted on 8 July 2026, published on 24 July and in force from 27 July. Obligations for stand-alone high-risk systems under Annex III — biometrics, employment, education, migration, access to essential services — now begin on 2 December 2027 rather than 2 August 2026; those embedded in products already covered by sectoral safety law move to 2 August 2028.

The central obligation of Europe's flagship AI statute has therefore been deferred by sixteen months without ever having bound anybody. On 13 November 2025, 133 civil society organisations and trade unions called the omnibus "the biggest rollback of digital fundamental rights in EU history". Industry calls it overdue realism. Both are responding to the same document.

The United States has no comprehensive federal AI statute and shows little sign of acquiring one. An attempt to impose a ten-year moratorium on state AI laws through the 2025 budget reconciliation bill was stripped out by the Senate in July 2025 on a vote of 99 to 1. What Congress has passed is narrow, chiefly the TAKE IT DOWN Act of 19 May 2025, carried in the House 409–2, on non-consensual intimate imagery, synthetic included. Executive Order 14365, of 11 December 2025, then moved the other way, directing the Attorney General to set up an AI Litigation Task Force to challenge state AI laws.

Which makes the state experiments worth watching. Colorado enacted SB 24-205 in May 2024, the first American statute to impose a European-style duty of reasonable care against algorithmic discrimination. Commencement was postponed from 1 February to 30 June 2026 by SB 25B-004, after Governor Jared Polis raised cost concerns in a special session. Then, on 14 May 2026, Polis signed SB 26-189, which repealed and replaced it. The duty of care, the impact assessments and the presumption of compliance are gone; in their place sit disclosure duties, a sixty-day cure period and a fault-based scheme allocating liability between developers and deployers. The one American state to adopt the European approach abandoned it seven weeks before it was due to apply. California went narrower and stuck: SB 53, signed on 29 September 2025, reaches only models trained above 10^26 operations by developers with revenues over 500 million dollars, and asks chiefly for published safety frameworks, incident notification and whistleblower protection.

China's rules arrived earlier than most Western commentary concedes, and were built differently: by service type rather than risk tier, enforced through registration rather than conformity assessment. The Administrative Provisions on Recommendation Algorithms took effect on 1 March 2022, the Deep Synthesis provisions on 10 January 2023, the Interim Measures for generative AI services on 15 August 2023, and labelling rules for AI-generated content on 1 September 2025 — fifteen months before the corresponding European deadline of 2 December 2026.

Above all of these sits the Council of Europe Framework Convention on artificial intelligence and human rights, democracy and the rule of law (CETS No. 225), opened for signature at Vilnius on 5 September 2024 and signed that day by the United Kingdom, Israel, the United States and the EU. It had 47 signatories by early September 2026, and the EU ratified on 15 May. It is still not in force, which requires five ratifications including three Council of Europe member states. Two years on, the first binding international treaty on artificial intelligence remains a document states have signed rather than one by which they are bound.

Judges, meanwhile, are deciding copyright — and disagreeing about ontology#

Where statutes hesitate, litigation proceeds. Three American decisions in 2025 set the terms. On 11 February, Judge Stephanos Bibas in Delaware held in Thomson Reuters v Ross Intelligence that using Westlaw headnotes to build a competing research tool was not fair use, the first and fourth statutory factors cutting sharply against Ross because the output substituted for the source. On 23 June, Judge William Alsup in the Northern District of California held in Bartz v Anthropic that training a language model on lawfully acquired books was transformative — "spectacularly so" — while assembling a permanent library from pirated copies was not. Two days later, Judge Vince Chhabria, in the same district, found for Meta in Kadrey v Meta on the record before him, but added a warning cited ever since: the real danger may be indirect substitution, systems enabling the rapid generation of countless works that compete with the originals even where those works do not themselves infringe. Anthropic then settled for 1.5 billion dollars, roughly 3,000 dollars for each of some 500,000 works, with final approval on 20 July 2026 — the largest recovery of its kind, and authority for nothing at all, since Alsup's reasoning will never now be tested on appeal.

The sharpest disagreement concerns what a trained model is. On 4 November 2025, in Getty Images v Stability AI [2025] EWHC 2863 (Ch), Mrs Justice Joanna Smith rejected Getty's secondary infringement claim. Model weights, she held, "are purely the product of the patterns and features which they have learnt over time during the training process"; the models contained no copy of any work and so could not be infringing copies. The contrast she drew was arithmetical: 220 terabytes of training images, a 3.44 gigabyte model. Getty has permission to appeal.

One week later, on 11 November 2025, the Regional Court of Munich I decided GEMA v OpenAI (42 O 14139/24), a claim over nine German song lyrics including Kristina Bach's Atemlos and Reinhard Mey's Über den Wolken. The court held that lyrics recoverable from the model by simple prompting were reproduced within it, that the text-and-data-mining exception covers compiling a corpus but not the training phase, and that OpenAI is liable for outputs delivered to users. OpenAI is appealing.

Two serious courts, seven days apart, reached opposite conclusions about whether a set of learned parameters contains a copy of what it was trained on. That is not a disagreement about copyright doctrine but about what sort of object is in front of them.

The responsibility gap, and the people who say there is no gap#

The vocabulary for all this is older than the technology that made it urgent. In 2004, in Ethics and Information Technology 6(3), 175–183, Andreas Matthias described the responsibility gap: with systems that learn during operation, the "manufacturer/operator of the machine is in principle not capable of predicting the future machine behaviour any more, and thus cannot be held morally responsible or liable for it". He was writing about neural networks, genetic algorithms and agent architectures twenty-two years ago, which should temper any sense that this crept up on us.

The literature has since refined the claim and attacked it. Filippo Santoni de Sio and Giulio Mecacci, in Philosophy and Technology 34(4), 1057–1084, distinguish four gaps rather than one — culpability, moral accountability, public accountability and active responsibility — and trace them to organisational and legal arrangements as much as to machine opacity. Daniel Tigard, in the same journal at 34(3), 589–607, argues flatly that there is no techno-responsibility gap: moral responsibility is a flexible practice, and it has absorbed new kinds of agent before.

The strongest deflationary reply, though, is legal rather than philosophical. Unpredictability defeats negligence, which needs foreseeability; it does not defeat strict liability, which was invented for hazards whose particular manifestations nobody can foresee. Keep a dangerous thing, cause damage, pay. Under that scheme the gap never opens: responsibility lands on whoever placed the system on the market and whoever deployed it, and the fact that nobody predicted the output is beside the point. The revised Product Liability Directive is that argument in statutory form. The second reply is that diffusion of responsibility across many actors is old news. Dennis Thompson named the problem of many hands in the American Political Science Review in 1980 (74(4), 905–916), writing about public officials, and corporate law has run for a century on the premise that an organisation can act and be liable although no individual did the whole of the thing.

Both replies are good, and neither is quite enough. Strict liability settles who pays; it does not settle who is answerable, and those are different goods — criminal law, professional discipline and public administration all demand the second. The many-hands analogy is stronger, but it works in corporate law because a corporation has an attributable decision procedure: boards, minutes, delegated authority, a will assembled out of human willing. That is what makes it intelligible to say the company decided, and nobody claims a model has anything of the kind. We have looked before at Scott Shapiro's argument about law and code; the point carries over, since the question is not whether rules can be made executable but who remains answerable once they are.

Here Robert Alexy's position, which we examined earlier this year in his account of the dual nature of law, does real work. Law necessarily raises a claim to correctness, and a claim to correctness entails an obligation to justify. A determination for which no reasons can be given fails that test, and on Alexy's account fails thereby to be a legal determination at all. Read through Alexy, an explainability requirement stops being a technical specification and becomes a condition of legality — a stronger footing than consumer-protection language provides, and one available to a court without any new statute.

The dilemma, put honestly#

Regulate early and you legislate on thin evidence about a moving target. Compliance costs are largely fixed, so they fall hardest on the entrants who might otherwise discipline incumbents: the firms best placed to absorb a conformity-assessment regime are those that already have compliance departments. Europe's own retreat shows the difficulty. The AI Act's high-risk provisions were amended before they ever applied, by the institution that had spent three years drafting them.

Wait for better evidence, though, and the distribution problem takes over. Harms in credit scoring, hiring, policing and diagnosis are not speculative and they are not evenly spread; once a practice is embedded in an industry's ordinary operations, undoing it means unwinding contracts, procurement cycles and settled expectations. Anyone who watched thirty years of privacy law arrive behind the fact knows how this goes.

The argument about the Brussels effect is likewise a real disagreement rather than a pose. Anu Bradford's The Brussels Effect (Oxford, 2020) documented how the size of the European market lets one jurisdiction set de facto global standards, because multinationals find it cheaper to comply everywhere than to segment. Against that, Marco Almada and Anca Radu argued in the German Law Journal 25(4) (2024) that the AI Act may export its framework without exporting its protections, a product-safety architecture being poorly suited to defending fundamental rights. And the competitiveness case, put at its highest in Mario Draghi's report of 9 September 2024, is that Europe has been writing rules for an industry it does not have.

None of these positions can claim a decisive test. Regulation 2026/1744 supports the reading that Europe regulated ahead of the evidence and had to correct itself, and equally the reading that Europe wrote defensible rules and then lost its nerve. Both fit the text. That is what it looks like when a field is governed in advance of anyone understanding it.

Something is certainly arriving: statutes with numbers, deadlines with dates, agencies with budgets, a settlement fund large enough to interest a hedge fund. Whether it amounts to a law of artificial intelligence is another matter, and the copyright judgments give the answer in miniature. A London court and a Munich court, a week apart, could not agree whether the object before them contained a copy. Before law can regulate a thing it has to be able to say what the thing is. On that the courts are still out, and the statutes, which must be drafted before the courts report, are guessing.