Why AI got regulated

Regulation arrived because three failure modes stopped being hypothetical. Bias and discrimination: automated hiring, credit and benefit systems encode and amplify the biases in their training data, with documented harms in employment and welfare contexts. Opacity: systems affecting people's lives made consequential decisions with no explanation and no appeal. Scale and authenticity: generative AI made synthetic text, images, audio and video cheap, destabilising everything from elections to fraud detection. Different governments weighted these differently — the EU legislated horizontally across all AI uses, the US leaned on existing sectoral law, China regulated services one by one — but every major jurisdiction now has binding AI rules or active bills. For anyone building or deploying AI, the question is no longer whether rules apply, but which ones and when.

The EU AI Act: the reference framework

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and phases in over years. Its architecture is risk-based. Prohibited practices (applicable since February 2025) ban a defined set: social scoring by public authorities, manipulative techniques that cause harm, untargeted facial-image scraping, emotion inference in workplaces and schools, and similar. High-risk systems — AI in employment and worker management, education, credit scoring, insurance pricing, law enforcement, migration, critical infrastructure and as safety components of regulated products — carry the Act's heavy obligations: risk-management systems, data governance, technical documentation, logging, human oversight, accuracy and cybersecurity requirements, plus registration and conformity assessment. Transparency obligations require that people know when they interact with AI systems and that AI-generated content is machine-detectable and disclosed. And general-purpose AI (GPAI) models — the foundation models — owe technical documentation, copyright-related policies and training-data summaries, with extra duties (evaluations, adversarial testing, incident reporting, systemic-risk mitigation) for models designated as posing systemic risk.

The timeline is where readers need the most care. GPAI obligations have applied since 2 August 2025, and their enforcement phase began on schedule in August 2026. But the Act's costliest chapter moved: the Digital Omnibus on AI (Regulation (EU) 2026/1744), in force since 27 July 2026, deferred high-risk obligations to 2 December 2027 for the main Annex III use cases and to 2 August 2028 for high-risk AI embedded in regulated products, citing unready standards and oversight structures. Prohibitions, transparency duties and the GPAI regime continue as planned. Our newsroom covered the deferral decision in detail — see EU AI Act high-risk deadlines deferred to December 2027 — and the EU's own implementation page (linked above) is the authoritative tracker.

The United States: a patchwork, not a statute

The US regulates AI through three overlapping layers. Existing law, actively enforced: consumer-protection, anti-discrimination and financial rules apply to AI systems — the FTC has pursued deceptive AI claims and unfair practices, and civil-rights statutes reach hiring and lending algorithms. Voluntary frameworks: NIST's AI Risk Management Framework has become the de-facto structure for corporate AI governance — inventory, risk assessment, measurement, mitigation — even without legal force. State legislation: a growing set of state laws, with Colorado's AI Act (2024) the template for requiring impact assessments and notices for high-risk automated decisions, plus deepfake, election-content and companion-chatbot rules elsewhere. Federal comprehensive bills have been proposed repeatedly without passage, and executive-order priorities have swung with administrations. The practical consequence: US-facing organisations manage a patchwork, and multi-state operations often end up building toward the strictest applicable standard anyway.

China, the UK and international efforts

China took the vertical route: rules on recommendation algorithms (2022), deep synthesis (2023) and generative AI services (2023) impose security assessments, content labelling, real-name registration and content-control duties on service providers, with a national labelling standard effective September 2025 covering AI-generated content. Enforcement is direct and service-specific. The UK chose principles over a single law: its pro-innovation framework assigns regulators to apply safety, transparency and fairness principles within their sectors, supported by an AI Safety Institute that evaluates frontier models. Internationally, the Council of Europe's AI convention (2024) opened for signature as the first treaty-level framework; the G7's Hiroshima Process set code-of-conduct expectations for advanced AI; and successive summit cycles — Bletchley (2023), Seoul (2024), Paris (2025) — built the political rhythm of frontier-safety pledges. The result is a genuinely multi-polar landscape where the EU writes the template, the US enforces by sector, China regulates by service, and treaty-level work proceeds slowly.

What organisations actually do to comply

Across jurisdictions, serious compliance programmes converge on the same moves. Inventory AI use — every model in products and operations, including vendor-provided ones, because deployers carry duties too. Classify by risk against the frameworks you fall under (the EU tiers, NIST's profiles, applicable state laws). Build documentation muscle: records of training data sources, evaluation results, intended use and limitations — the GPAI duties already require this of model providers, and high-risk obligations will demand far more. Implement transparency: AI-disclosure notices, content labelling, user-facing explanations. Keep humans meaningfully in the loop for consequential decisions, with oversight that can actually intervene. Prepare for December 2027 deliberately rather than at the deadline: the deferred obligations still arrive, and conformity work (standards, testing, registration) takes longer than organisations expect. Our machine-learning trends coverage tracks the technical practices — evaluation, documentation, monitoring — that make this work real.

How to follow AI policy without the noise

AI regulation coverage over-promises constantly — rules "banning AI", deadlines that move, penalties quoted out of context. Three habits keep you accurate: anchor every claim to the instrument (regulation number, official text, agency page — the sources above are the ones we use); distinguish what is in force from what is deferred or pending; and read enforcement as the real signal, because agencies reveal priorities through their cases. We track the news flow in our AI security & privacy section, hold policy reporting to the standards in our editorial policy, and update this guide when the legal picture materially changes.

Frequently asked questions

What is the EU AI Act?
The world's first comprehensive AI law. Adopted in 2024 and in force since 1 August 2024, it regulates AI by risk tier rather than by technology: a short list of prohibited uses, strict obligations for “high-risk” systems, transparency duties for chatbots and generated content, and wide obligations for general-purpose AI models. It applies to providers and deployers whose systems are used in the EU, regardless of where the company is based.
What are the AI Act's risk tiers?
Four. Unacceptable-risk practices are banned outright — including social scoring and manipulative systems. High-risk systems (for example in employment, education, credit, law enforcement and critical infrastructure) must meet requirements on risk management, data quality, documentation, human oversight and accuracy before market entry. Limited-risk systems carry transparency duties — users must know they are talking to an AI, and AI-generated content must be labelled. Minimal-risk systems face no new obligations.
What changed with the 2026 Digital Omnibus?
The Digital Omnibus on AI (Regulation (EU) 2026/1744), in force since 27 July 2026, restructured the Act's application timeline. The headline change: obligations for high-risk systems were deferred — to 2 December 2027 for Annex III uses (employment, education, credit and similar) and to 2 August 2028 for high-risk AI embedded in regulated products. General-purpose AI obligations and enforcement were not deferred, prohibitions and transparency duties remain, and some clarifications were added. The official reason: readiness — standards and oversight structures were not ready on the original schedule.
Does the AI Act apply to my organisation?
It applies if an AI system you provide or deploy is used in the EU and falls in scope — including as an employer using AI in hiring, a lender using AI in credit decisions, a school using AI in admissions, or a company building or fine-tuning general-purpose models. Even outside those high-risk categories, transparency duties apply if your product interacts with consumers as an AI or generates content. Compliance preparation is a legal question that depends on your exact use — the risk-tier summaries here are orientation, not legal advice.
What does the United States do about AI regulation?
There is no comprehensive federal AI law. The US approach combines sectoral enforcement (existing consumer-protection, civil-rights and financial rules applied by agencies such as the FTC), voluntary frameworks — notably NIST's AI Risk Management Framework — executive actions that shift with administrations, and a growing layer of state laws, such as Colorado's AI Act focused on high-risk automated decisions. Companies operating nationally end up managing a patchwork rather than one statute.
How does China regulate AI?
Through targeted rules issued since 2021: regulations on recommendation algorithms, deep synthesis (synthetic media) and generative AI services that require content labelling, security assessments and alignment with core values, plus a national AI labelling standard taking effect in 2025. The model is vertical — service-by-service rules with strong enforcement — rather than the EU's horizontal risk-tier law.