AI regulation in 2026 — the EU AI Act, US executive orders, and what's next

AI regulation has moved from theoretical to operational since 2024. The EU AI Act is being enforced. The US has gone through multiple executive-order revisions. Frontier labs have voluntary commitments, mandatory model evaluations, and increasing transparency requirements. The political coverage has been wildly polarized. Here's the actual regulatory landscape in 2026.

EU AI Act

Passed in 2024, enforcement phasing in across 2025-2027. It's the world's most comprehensive AI law and it works on a *risk-tier* model:

- Unacceptable risk: Banned outright. Social-scoring systems, real-time biometric surveillance in public spaces (with narrow exceptions), emotion-recognition in workplaces/schools. - High risk: Allowed with strict compliance. Healthcare diagnostics, hiring algorithms, education-admissions algorithms, critical infrastructure. Requires risk assessments, data governance, human oversight, transparency, accuracy testing. - Limited risk: Allowed with transparency obligations. Chatbots must disclose they're AI. Synthetic media must be labeled. - Minimal risk: Allowed freely. Spam filters, video-game AI.

Frontier models (the "GPAI" — General Purpose AI — tier) have additional obligations: model documentation, copyright compliance, energy disclosures, and serious-incident reporting. The most capable frontier models — the "GPAI with systemic risk" tier — have evaluation and red-teaming requirements before deployment.

Fines for non-compliance scale up to €35M or 7% of global revenue.

US approach

The US has no comprehensive federal AI law. Regulation has happened through three mechanisms:

1. Executive orders (Biden 2023; revised in 2025): voluntary commitments from frontier labs, NIST guidance, agency-specific safety requirements (especially DOD, DHS, HHS). 2. State laws: California's SB-1047 (failed 2024 version, revised in 2025), Colorado's AI Act, NYC bias audit law for hiring algorithms. Patchy and inconsistent. 3. Existing law enforcement: FTC enforcement on deceptive AI claims, copyright litigation against scraping, sectoral regulation (FDA for medical AI, FAA for drones, etc.).

The federal landscape is in flux as of early 2026 and varies significantly by administration position.

What's actually in scope

For ordinary AI tools you use: - Chatbots / assistants: Must disclose AI nature in EU. Generally no specific US federal requirement, though FTC consumer-protection law applies. - Generated images and video: Must be labeled as AI in EU. Voluntary in most US contexts. - AI-assisted hiring: NYC requires bias audit; EU requires high-risk-tier compliance; rest of US largely unregulated. - Medical AI: FDA approval required in US; EU MDR + AI Act stack. - Frontier models: Significant compliance burden in EU; lighter and more voluntary in US.

How to read AI regulation coverage

The political fight over AI regulation has crystallized around three positions:

1. Pro-regulation safety-focused: Frontier risk is real, voluntary commitments insufficient, EU model is roughly right, US needs federal action. 2. Pro-regulation civil-rights-focused: Bias, surveillance, and discrimination harms are immediate and concrete; need enforcement on existing tools more than frontier theorizing. 3. Pro-innovation skeptics: Regulation will favor incumbents, slow US progress vs China, and impose costs that exceed benefits.

All three positions have serious advocates and serious arguments. Articles that demolish straw-man versions of any of them are not informing you. Articles that conflate the safety-focused and civil-rights-focused positions miss that they often disagree with each other almost as much as they disagree with skeptics.

Prism Ask the AIs lets you ask the same regulation question to GPT-5.2, Claude Sonnet 4.5, and Gemini 3 Pro and see where they agree and disagree. (Yes, asking the regulated AIs about AI regulation is a bit recursive. Their disagreements are illuminating.)

Related: How AI is changing news bias · What is media bias.