AI Regulation: the White House reverses course on AI safety  May 21st 2026

Watch & Listen First

Hard Fork — "A.I. Safety Is So Back + Mythos Mayhem with Nikesh Arora" · May 15, 2026 · Apple Podcasts
-> The NYT tech show on why the Trump administration dropped its "doomer fear-mongering" framing — with Palo Alto Networks CEO Nikesh Arora on securing the internet as cyber-capable models like Mythos arrive.

Key Takeaways

• Washington reversed itself. The administration that rescinded Biden's AI order and challenged state AI laws is now drafting a federal frontier-model safety and cybersecurity order — triggered by cyber-capable models like Anthropic's Mythos.
• It's voluntary, not licensing. The draft asks labs to share models roughly 90 days pre-launch; there is no mandatory testing mandate. Treat it as an emerging baseline expectation, not a binding regime.
• The state patchwork is widening, not narrowing. Colorado gutted its own risk-based AI Act the same month Connecticut passed a comprehensive one. Multistate deployers now manage divergent regimes.
• The deepfake clock is running. The TAKE IT DOWN Act's takedown duty is live as of May 19, and the NO FAKES Act is back — any product touching likeness or voice needs a removal workflow now.
• The EU gave you runway and a map. Draft high-risk classification guidelines (comment by June 23) plus the omnibus deferral of Annex III duties to December 2027 — use the time; the goalposts still move.

The Big Story

Trump prepares an executive order putting frontier AI models under voluntary federal pre-launch review · May 20, 2026 · CNN
-> The draft order splits into a cybersecurity track and a "covered frontier models" track, asking labs to share advanced models with the government roughly 90 days before public release — a voluntary clearinghouse, not a licensing regime, and a sharp reversal for an administration that spent a year dismantling AI rules. For frontier developers (OpenAI and Anthropic are already at the table), the near-term ask is process, not paperwork: a pre-launch review window, security disclosures, and infrastructure-sharing expectations that will harden into de facto baselines even though nothing here is statutorily binding. Everyone else should read it as a signal — federal AI policy has pivoted from "innovation only" to "innovation plus security," and the next order may not be voluntary.

Also This Week

Connecticut sends the AIRT Act, a comprehensive AEDT and chatbot AI law, to its governor · May 14, 2026 · Inside Global Tech
-> Employers and HR-tech vendors in Connecticut face new duties around automated employment decision technology, chatbot disclosure, layoff reporting, and synthetic-content labeling — build an AEDT inventory before the phased effective dates land.

Colorado repeals and replaces its landmark AI Act with a lighter disclosure-based ADMT law · May 14, 2026 · Colorado General Assembly
-> Governor Polis signed SB 189 weeks before the original Colorado AI Act would have taken effect, stripping risk and impact assessments and the algorithmic-discrimination duty of care — retire compliance built for the old law and rebuild for ADMT notice-and-appeal rules effective January 1, 2027.

King's Speech bundles the UK's first statutory AI framework into a "Regulating for Growth" bill · May 14, 2026 · SecurityBrief UK
-> Britain skips a single AI super-regulator, instead giving Ofcom, the ICO, the CMA, the FCA and the MHRA statutory AI remits and the AI Safety Institute a legal footing — multinationals get coordination, but five regulators to satisfy instead of one.

The European Commission opens consultation on draft guidelines for classifying high-risk AI systems · May 19, 2026 · European Commission
-> Largely overlooked amid the executive-order noise, these Article 6(5) guidelines and worked examples are the clearest signal yet of which products the EU will treat as high-risk — pressure-test your own classification calls against the examples before the June 23 comment deadline.

Revised NO FAKES Act returns to Congress to create a federal likeness-and-voice right · May 20, 2026 · Digital Music News
-> Backed by a bipartisan Senate group plus SAG-AFTRA, Amazon and OpenAI, the bill would create a transferable, post-mortem digital-replication right and a takedown duty — generative-media products would need licensing and removal workflows surviving 70 years past a person's death.

Deadlines & Compliance

• Now in effect (May 19, 2026): The federal TAKE IT DOWN Act's platform obligation is live — services hosting user content must honor 48-hour notice-and-removal requests for non-consensual intimate imagery, including AI-generated deepfakes.
• June 23, 2026: Deadline to submit feedback on the European Commission's draft high-risk AI classification guidelines.
• June 30, 2026: The original Colorado AI Act's effective date lapses — repealed by SB 189; the replacement ADMT regime takes effect January 1, 2027. Deployers should confirm they are not still building to the defunct law.
• August 2, 2026: The EU AI Office gains enforcement powers over general-purpose AI model providers — requests for information, model evaluations, and fines up to €15M or 3% of global turnover.

Worth Reading

• What the EU AI Omnibus Deal Changes for the AI Act — and What Lies Ahead — The clearest map of which deadlines moved, which prohibitions were added, and what still needs formal adoption before July.
• Colorado Hits Reset on AI Regulation: What Developers and Deployers Need to Know — A precise side-by-side of what SB 189 deletes and what survives — essential if you scoped a program to the old risk-based Act.
• The U.S. Has 1,200 AI Bills and No Good Test for Any of Them — Why the volume of state activity is itself the compliance risk, and why a federal preemption fight is now unavoidable.

Washington spent a year tearing AI rules down; this week it picked up a hammer to build — just don't expect the blueprint to hold still.