How the EU, US, China and Singapore Regulate AI Harms

AI regulation was taking four different shapes by September 20, 2026. The EU had enacted a broad AI Act but was phasing in its requirements. The US had no comprehensive federal AI law, while states continued to legislate. China regulated generative-AI services through targeted measures, and Singapore combined existing rules with AI governance frameworks.7111416 Comparing them requires one distinction above all: a law in force, a future compliance date, a proposal and voluntary guidance are not the same thing.

European Union: one AI Act, several deadlines

The EU AI Act entered into force in August 2024, with provisions becoming applicable in stages.7 Its risk-based approach distinguishes prohibited uses from high-risk applications and systems subject to other requirements. The European Commission identifies harmful manipulation or deception, exploitation of vulnerabilities, social scoring and untargeted scraping to build facial-recognition databases among the prohibited practices.3 High-risk areas identified by the Commission include biometrics, critical infrastructure, education, employment and border control.6

The timetable matters as much as the categories. Following 2026 simplification measures, the European Parliament announced December 2, 2027 for obligations on stand-alone high-risk systems and August 2, 2028 for certain systems embedded as safety components.5 It also announced approval of a measure banning AI systems that generate child sexual abuse material or specified non-consensual intimate content.5 Parliamentary approval alone should not be mistaken for proof that every part of that measure was already enforceable on September 20.

The provided material identifies the EU AI Office and member-state authorities as enforcement bodies, but the supporting Commission FAQ is dated September 23, 2026—after this article’s cutoff. It offers retrospective corroboration, not evidence that was published by September 20.1 The available extracts also do not establish the complete four-tier classification, every labelling duty or the detailed technical requirements for high-risk systems.

United States: agency oversight and state-by-state laws

The US had no unified federal AI law. Instead, federal and state initiatives operated alongside oversight by existing agencies, including the Federal Trade Commission.1116 The White House favored lighter national standards, but that policy preference was not itself a comprehensive statute.11

State legislation shows both the reach and the limits of this approach. Colorado’s 2024 law addressed algorithmic discrimination in consequential decisions involving high-risk AI. A law signed in May 2026 repealed and replaced those provisions with requirements focused on automated decision-making technology; the available extracts do not confirm the replacement rules’ commencement date.913 Meanwhile, a September 2026 legal review reports that California laws covering companion chatbots, AI training-data transparency and frontier-AI transparency took effect on January 1, 2026.12

Claims about a nationwide mandatory AI “kill switch” should be treated cautiously: the provided evidence does not identify an enacted federal requirement, its scope or its effective date.

China: targeted rules for generative-AI services

China’s Interim AI Measures are identified as a specific administrative regulation for generative-AI services. A September 20 comparison describes a wider approach overseen by the Cyberspace Administration, with attention to lawful data use, content labelling and protection of minors.1611

That evidence supports a more targeted regulatory model than the EU’s single overarching AI Act, but not a complete inventory of China’s obligations. In particular, the available extracts do not establish the precise registration triggers, training-data consent rules or commencement dates for proposed AI-companion safeguards. Break reminders, age checks and restrictions on virtual relationships should not be presented as enforceable requirements without confirming their legal status.

Singapore: existing rules plus evolving guidance

Singapore had not adopted a single AI-specific statute according to the regulatory tracker supplied for this comparison; its approach drew on existing regulation and AI governance frameworks.1611 In 2026, the Infocomm Media Development Authority published a Model AI Governance Framework for Agentic AI, while the Personal Data Protection Commission issued proposed advisory guidelines on using personal data in generative AI.14 A proposal is not a binding rule, and a governance framework should not be treated as one without a separate legal basis.

The available extracts do not establish whether particular recommendations—such as human approval for high-risk agent actions or notifying consumers about AI training uses of their data—were mandatory on September 20. For a practical comparison, the question is therefore not merely which safeguard is discussed, but who must follow it, under what authority and from what date.