Essay · Richard Anderson · 19 November 2026

Three rulebooks

AI regulation in the UK, the EU and the US

Tower Bridge rendered in loose painted strokes, its towers lit gold against a smoky green and amber sky, the river below in blues
Tower Bridge. Photograph by Richard Anderson

Why compare

The EU regulates AI as a product with one cross-economy law. The UK leaves it to the regulators it already has. The US federal government is trying to stop the states from regulating it at all. Any organisation that trades across all three now works to three different rulebooks, and the differences matter more than the headlines suggest.

This essay sets out, as of mid-November 2026, what each regime actually requires, how that plays out in five sectors, and what would change for the UK if it had to follow EU rules, whether by rejoining or by closer alignment. It is not an argument for or against membership. It is an attempt to set out the facts on which that argument could be held.

One caution applies throughout. All three regimes moved substantially during 2026, and two of them (the US and, to a lesser degree, the EU) are still moving. Dates and obligations here are correct as at 17 November 2026; anyone relying on them should check again.

Three approaches at a glance

The three regimes differ less in what they care about than in who decides, and when.

European Union United Kingdom United States
Core instrument AI Act (Regulation 2024/1689), amended by the Digital Omnibus on AI No AI statute; existing regulators apply existing law No federal AI statute; executive orders, agency guidance, state laws
Philosophy Risk-based product safety law, applied across the economy “Pro-innovation”, sector-led, sandboxes Light-touch, “minimally burdensome” national standard
Who enforces National market surveillance authorities; Commission AI Office for general-purpose models FCA, PRA, ICO, CMA, MHRA, Ofcom and others, each in its own remit FTC, sector agencies, state attorneys general; DOJ task force challenging state laws
Bans on uses Yes: listed prohibited practices since Feb 2025, two more from Dec 2026 None AI-specific None federal
High-risk duties Yes: from 2 Dec 2027 (standalone) and 2 Aug 2028 (in products) No general category; sector rules apply Some state laws, narrowing; federal push to preempt them
Automated decisions about people GDPR Article 22 (a general restriction) UK GDPR Articles 22A to 22D (permitted with safeguards, since Feb 2026) State-level, patchy
Top penalty EUR 35m or 7% of worldwide turnover Regulator-specific (e.g. ICO up to GBP 17.5m or 4%) Agency- and state-specific
Reach beyond borders Yes: applies to non-EU firms when their AI, or its output, is used in the EU Limited Limited, but export controls on chips and models

The European Union: one law, delayed but intact

The EU AI Act is the only comprehensive AI statute of the three, and in 2026 the EU moved its dates without removing its substance.

The Act sorts AI by use. A short list of practices is banned outright: social scoring, certain manipulative techniques, untargeted scraping of faces, most real-time biometric identification in public, and others. Those bans have applied since February 2025. Providers of general-purpose AI models have had transparency and copyright duties since August 2025. Below the bans sits a “high-risk” category covering AI used in employment, education, credit and insurance, critical infrastructure, law enforcement and biometrics, plus AI built into products already regulated for safety, such as medical devices, machinery and toys. High-risk systems need risk management, data governance, technical documentation, logging, human oversight and, in many cases, a conformity assessment before sale.

The high-risk duties were due in August 2026. By late 2025 the standards and guidance needed to comply were plainly not ready, and the Commission tabled the Digital Omnibus on AI in November 2025. Parliament and Council reached political agreement on 7 May 2026 and the Council gave final approval on 29 June. The result:

  • Standalone high-risk systems (the Annex III list) now apply from 2 December 2027, a 16-month delay.
  • High-risk AI inside regulated products (Annex I) applies from 2 August 2028, a 12-month delay.
  • Two new prohibitions were added, applying from 2 December 2026, one of them on AI tools that generate non-consensual intimate imagery (“nudification” apps). Both carry the top penalty tier.
  • Watermarking of AI-generated content (Article 50(2)) for existing generative systems applies from 2 December 2026, a short grace period rather than a long one.
  • Relief for smaller firms was widened to cover “small mid-caps”, not just SMEs.

Nothing was repealed. The penalty structure is unchanged: up to EUR 35m or 7% of worldwide turnover for prohibited practices, and up to EUR 15m or 3% for most other breaches, including the transparency duties. The penalty provisions have applied since August 2025, and the Commission’s power to fine providers of general-purpose models since August 2026.

Two other EU instruments matter as much as the AI Act for most organisations. GDPR Article 22 still restricts decisions about people made solely by automated means, and the EU’s separate data Omnibus proposes, but has not yet enacted, allowing “legitimate interest” as a basis for training AI models. The revised Product Liability Directive also brings software, including AI, within strict product liability.

The Act reaches beyond the EU. It applies to any provider, wherever based, that places a system on the EU market, and to a provider or deployer outside the EU where the output of the system is used inside it. A UK firm selling AI-enabled hiring software to an EU employer is caught today, whatever the UK does.

The United Kingdom: no AI Act, by choice

The UK has decided, for now, not to write an AI statute, and to regulate AI through the bodies and laws it already has.

The July 2024 King’s Speech promised a bill for the most powerful models. It never appeared. The May 2026 King’s Speech contained no AI bill; instead it announced a Regulating for Growth Bill, which would put regulatory sandboxes on a statutory footing and strengthen regulators’ duty to consider growth. Lord Holmes’s private member’s AI (Regulation) Bill fell at the end of the session without a second reading, and a private member’s Artificial Superintelligence Security Bill is not expected to progress.

The practical regime therefore has four parts.

  1. Sector regulators applying existing rules. The FCA, PRA, ICO, CMA, Ofcom, MHRA and others each treat AI as a way of doing something they already regulate. Their principles are the 2023 white paper’s five: safety, transparency, fairness, accountability, contestability. They are not statutory duties.
  2. Data protection, now looser than the EU’s. The Data (Use and Access) Act 2025 brought its main changes into force on 5 February 2026. Old Article 22 of UK GDPR, a general prohibition on significant decisions made solely by automated means, has been replaced by Articles 22A to 22D. Automated decisions are now permitted on ordinary personal data, provided the organisation tells people, lets them make representations, offers human review and lets them contest the outcome. The stricter rule survives only for special category data such as health or ethnicity. This is the single largest legal divergence from the EU in this field.
  3. Sandboxes. The government’s AI Growth Labs, the FCA’s AI Lab and the MHRA’s AI Airlock let firms test products under supervision with some rules relaxed.
  4. The AI Security Institute, renamed from the AI Safety Institute, which tests frontier models by agreement with their developers. It has no enforcement powers.

The UK has kept its EU data adequacy despite the divergence. The European Commission renewed both adequacy decisions on 19 December 2025, after reviewing the Data (Use and Access) Act, until 27 December 2031, with a review after four years.

Northern Ireland is now the exception. On 10 November 2026 the EU AI Act was added to the Windsor Framework and parts of it began to apply in Northern Ireland. The application is limited to AI closely integrated into goods already covered by the Framework (products such as toys, lifts and medical devices), not to software a business uses internally. Part of the UK therefore already works under EU AI rules: a small, real preview of the question this essay turns to later.

The United States: a fight over who regulates

The US has no federal AI law, and its federal government is now working to prevent the states from filling the gap.

On 11 December 2025 President Trump signed Executive Order 14365, “Ensuring a national policy framework for artificial intelligence”. It set up an AI Litigation Task Force in the Department of Justice to challenge state AI laws in court, directed Commerce to list state laws it considers burdensome, and tied some federal funding to states’ restraint. It carved out child safety, data-centre zoning and state procurement. On 20 March 2026 the White House followed with a National Policy Framework for AI, asking Congress to pre-empt burdensome state laws, create no new AI regulator and avoid “open-ended liability” for AI firms. It is a legislative blueprint, not law. Two earlier attempts to write pre-emption into must-pass legislation failed, and Congress has not yet passed a framework bill.

The states have not stopped. By one count they enacted 109 AI laws in the first half of 2026. But the most ambitious one has been rolled back. Colorado’s 2024 AI Act, modelled partly on the EU’s high-risk approach, was challenged by xAI in April 2026; the DOJ intervened, the first use of the executive order, and enforcement was stayed. On 14 May 2026 Colorado repealed and replaced it with a narrower law on automated decision-making technology, focused on notice, record-keeping and human review, effective 1 January 2027. Impact assessments and the duty to prevent algorithmic discrimination have gone. California’s frontier-model transparency law and Texas’s governance act, both effective from January 2026, also lean towards disclosure rather than restriction.

Sector regulators matter more in the US than any AI-specific rule. The FTC treats deceptive AI claims as unfair practices. In banking, the Fed, OCC and FDIC replaced the long-standing model risk guidance, SR 11-7, with SR 26-2 on 17 April 2026: shorter, principles-based and organised around materiality. The agencies have said generative and agentic AI will be addressed separately, after a request for information. For now, US banks have no specific supervisory guidance on large language models.

The result is the least predictable of the three regimes. What a firm must do depends on its sector, the states it serves, and the outcome of litigation and legislation that is still in progress.

Sector by sector: what an organisation can actually do

The gap between the regimes is widest where AI makes decisions about people, and narrowest where AI sits inside a regulated product.

Sector European Union United Kingdom United States
Financial services (credit, insurance, advice) Credit scoring and life and health insurance pricing are high-risk from Dec 2027: documentation, bias testing, human oversight. GDPR Art. 22 limits fully automated decisions No AI-specific rules. The FCA’s Mills Review (July 2026) recommended keeping the principles-based regime; Consumer Duty and SM&CR carry the weight. Automated decisions allowed with safeguards SR 26-2 for model risk; fair-lending law requires reasons for adverse decisions. No specific guidance yet on generative or agentic AI
Health and medical devices AI medical devices need MDR certification now, plus AI Act high-risk duties from Aug 2028 MHRA under UK MDR 2002, with the AI Airlock sandbox. Government accepted all 44 recommendations of the National Commission on 6 Oct 2026; draft guidance on adaptive AI due Dec 2026 FDA device clearance, with an established route for pre-approved model changes
Employment and HR Recruitment, promotion and monitoring tools are high-risk from Dec 2027; workers must be told. Emotion recognition at work is banned Equality Act and UK GDPR Arts. 22A to 22D; the ICO has investigated automated recruitment City and state rules (e.g. New York City bias audits); Colorado’s narrower ADMT law from Jan 2027
Generative content and marketing Machine-readable marking of AI output from 2 Dec 2026; deepfakes must be disclosed; nudification tools banned Online Safety Act duties enforced by Ofcom; no general labelling duty; AI and copyright still unresolved Federal law against non-consensual intimate deepfakes; state disclosure laws; no general labelling duty
Products with embedded AI (machinery, toys, vehicles) Existing product law plus AI Act duties from Aug 2028 Great Britain: existing product law only. Northern Ireland: parts of the AI Act since 10 Nov 2026 Existing product safety law

Read across the rows and a pattern appears. For embedded products and medical devices, all three regimes rely on safety regulators who already certify products, and the EU’s extra layer is real but incremental. For decisions about people (credit, jobs, insurance) the EU imposes a full compliance regime, the UK relies on outcomes-based conduct rules and a looser data law, and the US relies on older anti-discrimination law while its newer state laws shrink.

In practice, a UK firm has the most freedom to deploy AI in decisions about its UK customers and staff. It loses much of that freedom for any EU customer or employee, because the AI Act and GDPR follow the person affected, not the firm.

If the UK had to follow EU rules

Following EU rules on AI would change less for UK firms that already sell into Europe than for those that do not, and the biggest single change would be in data protection, not the AI Act.

There are two quite different ways the UK could end up following EU rules, and the difference matters.

  • Membership. The AI Act is a regulation, so it would apply directly from accession, with no transposition and no opt-out. The UK would also gain votes in the Council, MEPs, a commissioner and a seat in the AI Board, and so a say in future amendments. Accession would take years; by the time it happened the Act’s high-risk regime would be fully in force and, very likely, amended again.
  • Alignment without membership. The UK is already doing this elsewhere. Under the planned SPS agreement on food and plant health, agreed in principle in early November, Great Britain will follow EU rules dynamically, including future changes, in return for fewer border checks. The same model applied to AI would bring the rules without the vote. Northern Ireland, since 10 November, is a small working example of exactly that.

Whichever route, the main effects would fall in seven places.

Area Great Britain today If EU rules applied
Prohibited practices None AI-specific The EU list applies, including emotion recognition at work and in education, social scoring and nudification tools
High-risk uses (credit, insurance, hiring, education, infrastructure) Sector conduct rules, outcomes-based Conformity assessment, documentation, logging, human oversight, registration in the EU database
Automated decisions about people Permitted on ordinary data with safeguards (Arts. 22A to 22D) Reverts to the GDPR Art. 22 restriction; recent UK flexibilities on legitimate interests and research would need review
Who enforces Each sector regulator in its own remit The UK would designate market surveillance authorities, probably existing regulators such as the ICO, FCA and MHRA; the Commission’s AI Office for general-purpose models; final say with the Court of Justice
Sandboxes Growth Labs may relax rules under supervision Sandboxes are required in every member state, but cannot disapply the Act itself
Generative content No general labelling duty Marking of AI output and deepfake disclosure
Data flows with the EU Rely on adequacy, renewable in 2031 No adequacy decision needed

The costs and benefits fall unevenly.

  1. Firms already serving EU customers largely comply today, because the Act follows their output into Europe. For them a single rulebook would remove the cost of running two, and end the Northern Ireland anomaly.
  2. Firms serving only UK customers would face new compliance work, mainly if they use AI in hiring, lending, insurance or education. Smaller firms would get the Act’s lighter treatment for SMEs and small mid-caps, but not exemption.
  3. UK regulators would move from setting their own approach to applying a common one. The FCA’s Mills Review conclusion, no AI-specific rules, would sit alongside a statute that has some.
  4. Influence is the trade-off. Inside the EU the UK would help write the next amendments; aligned from outside, it would take them as written.
  5. Relations with the US would be harder to manage. The current US administration treats EU-style AI regulation as a barrier, and the UK’s AI and compute investment depends heavily on US firms.

The contrast is less stark than it looks, for two reasons. The EU is itself loosening and delaying parts of the Act, so the gap is narrower in 2026 than it looked in 2024. And the UK has not yet diverged as far as its rhetoric suggests: much of what UK regulators expect of AI, such as fairness, explainability, human accountability and testing, resembles what the Act requires, minus the paperwork and the fines.

What boards can take from this

For most UK boards the practical question is not which regime is better but which ones already apply to them, and the honest answer is often more than one.

  • Map where the output lands, not where the firm sits. Any AI system with output that reaches an EU customer, applicant or employee brings the AI Act with it today.
  • Treat the EU’s delay as time to prepare, not a reprieve. The bans and the general-purpose model duties already apply; marking of AI-generated content applies from 2 December 2026; the high-risk duties arrive in December 2027.
  • Do not read UK flexibility as UK permission. The FCA, ICO and others will still judge AI by outcomes: unfair treatment, poor explanations, weak accountability. A firm can breach Consumer Duty or the Equality Act with an AI tool that no AI law mentions.
  • Build once, to the higher standard, where it is cheap to do so. Documentation, testing and human oversight built to EU expectations will satisfy UK regulators too; the reverse is not true.
  • Watch the US for volatility, not direction. Federal policy is clear in intent but unsettled in law, and state obligations can appear, shrink or vanish within a year, as Colorado showed.

Whether the UK should move closer to EU rules is a political question. On the facts set out here, the choice is narrower than it is usually presented: much of the EU’s regime already reaches UK firms, part of the UK already works under it, and both blocs are adjusting as they go.

Richard Anderson
Founder, AndersonRisk