Long Yi
← Back to all writing

Where US AI policy is actually being written

June 13, 2026

governancepolicygovernance

In February I mapped global AI governance and concluded that the United States — home to roughly half of frontier AI development — had no binding federal safety regulation and was actively working to prevent it. That was accurate then. It is still accurate now.

For the full political map — factions, Sanders/populism, Rutinel, 2026 races, and preemption wars — see A political map of US AI policy. This article is a focused case study on New York.

What changed in the four months since is that policy is being written anyway, just not in Congress. California and New York passed the first state laws requiring frontier labs to publish safety frameworks, report critical incidents, and protect whistleblowers. Industry spent millions trying to kill them. When that failed at the state level, a super PAC spent $7 million trying to keep one of the bills’ authors out of Congress.

This is the story of where US AI policy actually lives right now: state legislatures, federal preemption battles, and a congressional primary that became a national referendum on whether frontier AI should be regulated at all.

The federal vacuum — still

The baseline hasn’t moved much. Biden’s Executive Order 14110 (October 2023) required safety reporting for large models and pushed NIST to develop standards. Trump revoked it on January 20, 2025, replacing it with an order focused on “removing barriers to American leadership.” NIST’s AI Safety Institute was renamed the Center for AI Standards and Innovation (CAISI), its budget stayed around $10 million, and its mandate narrowed.

Congress has introduced dozens of AI bills. None have passed as comprehensive federal law. The Cruz attempt to attach a 10-year moratorium on state AI enforcement to the 2025 budget bill was stripped 99–1 in July 2025 — a rare bipartisan rejection of broad preemption. But that doesn’t mean preemption is dead. It means the fight moved to narrower, more sophisticated drafts.

The question for 2026 is not “will the US regulate AI?” It is who gets to regulate it, at what level, and with what teeth.

Two states, one template

After California’s SB 1047 — which would have required safety testing and let the state halt dangerous releases — was vetoed by Governor Newsom in September 2024, frontier-safety advocates pivoted. The new strategy: transparency and accountability first, not training bans or compute caps. Get labs to document what they’re doing before asking legislatures to stop them from doing it.

Two sister laws emerged.

California SB 53 (Transparency in Frontier AI Act)

Authored by State Senator Scott Wiener, signed September 2025, in force January 2026. Anthropic publicly supported it.

It applies to large frontier developers — roughly, entities training models above 10²⁶ FLOPs with over $500 million in annual revenue. They must:

  • Publish a Frontier AI framework on their website: how they assess catastrophic risks, secure unreleased weights, use third-party evaluators, and update annually
  • Issue transparency reports before deploying new or substantially modified frontier models
  • Report critical safety incidents to California’s Office of Emergency Services (15 days standard, 24 hours for emergencies)
  • Maintain whistleblower protections and anonymous internal reporting channels

Penalties: up to ~$1 million per violation, enforced by the state Attorney General.

It does not require pausing training, government pre-approval of releases, or compute caps. It makes voluntary safety practices mandatory law.

New York RAISE Act (Responsible AI Safety and Education Act)

Co-authored by Assemblymember Alex Bores and Senator Andrew Gounardes, passed the legislature June 2025 with bipartisan support (~84% NY voter support cited by sponsors), signed by Governor Kathy Hochul December 19, 2025 — but significantly weakened before signing. Effective January 1, 2027 after chapter amendments.

RAISE covers similar ground: large frontier developers ($100M+ training compute threshold) must write and publish safety and security protocols, conduct third-party risk assessments, protect model weights, report critical safety incidents within 72 hours, and maintain whistleblower protections. An Office of Responsible AI Safety and Education within NY’s Department of Financial Services gets oversight authority. AG civil penalties — originally up to $30 million, cut by Hochul toward California-scale ($1 million).

Politico called the passed version “widely regarded as the furthest-reaching in the country.” A Trump executive order reportedly targeted the bill. Tech lobbyists and VCs spent heavily to kill or water it down.

Side by side

CA SB 53NY RAISE
AuthorSen. Scott WienerAsm. Alex Bores + Sen. Andrew Gounardes
SignedSept 2025Dec 2025 (weakened)
EffectiveJan 2026Jan 2027
Incident reporting15 days / 24h emergency72 hours
Penalties~$1M/violation~$1M/violation (after Hochul cuts)
Industry stanceAnthropic endorsedIndustry fought; Trump EO targeted

These are not bans. They are mandatory safety plans + incident reporting + whistleblower protection — the same package the AI safety research community has advocated for years, now encoded in state law in the two states where most frontier labs operate.

Encode (Sneha Revanur’s organization) helped pass both.

The federal counter-move

June 2026 brought three overlapping federal responses — and they disagree on the central question.

Trump EO (June 2)

A voluntary pre-release review framework: frontier developers can choose to submit models to CISA/NSA/ Treasury review for up to 30 days before public release. Benchmarks for “covered frontier models” to be defined within 60 days. Explicitly cannot be used for mandatory licensing or pre-approval. NSA leads benchmark definition; CAISI is advisory only.

Hands-off. No mandatory incident reporting. No penalties for non-participation.

GAAIA discussion draft (June 4)

Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page bipartisan discussion draft of the Great American Artificial Intelligence Act — the most substantive federal frontier governance proposal to date.

For large frontier developers, it would require:

  • Public Frontier AI frameworks (similar to SB 53/RAISE)
  • Semi-annual third-party audits by CAISI-certified Independent Verification Organizations
  • Mandatory reporting of critical safety incidents
  • Penalties up to $100 million per day for violations
  • Codification of CAISI with ~$100M/year authorization

But it also includes 3-year federal preemption: states cannot enact new laws “specifically regulating the development” of AI models. General laws (civil rights, labor, privacy, consumer protection) and post-deployment regulation are preserved. Training data transparency laws like California’s AB 2013 would be preempted. Frontier safety laws in CA, NY, and Illinois would be “federalized” — meaning replaced by the federal standard, not duplicated.

Roll Call and Encode flagged preemption as the flashpoint. Wiener and Encode oppose it. Industry often wants federal rules — but weaker ones that override state laws they already fought.

The fight is not “regulate vs. don’t”

It is:

  1. Who sets benchmarks — public thresholds (GAAIA/SB 53) vs. classified NSA standards (Trump EO)
  2. Mandatory vs. voluntary — incident reporting and audits required, or optional 30-day review
  3. Whether states keep authority — or get frozen for three years while Congress negotiates

If GAAIA passes with preemption intact, SB 53’s development-stage obligations could be superseded. That is the scenario safety advocates are organizing against.

Alex Bores: case study, not hero

Alex Bores is not an AI safety researcher. He is a NY State Assemblymember (District 73, Manhattan East Side) with a Georgia Tech MS in computer science and a Palantir past — facts his opponents used relentlessly in attack ads. He claims to be the only Democrat elected at any level of NY government with a CS degree.

He is relevant because he co-authored RAISE and ran for Congress on a platform to nationalize it — making him the test case for whether a state-level frontier safety legislator could survive industry PAC warfare and reach the House.

His framework

Bores uses a 3-axis taxonomy for AI legislation:

  1. Outlook: pessimistic vs. optimistic about AI
  2. Time scale: near-term vs. long-term
  3. Scope: use-case-specific vs. general

Most state AI bills are pessimistic + near-term + specific (hiring bias, deepfakes, chatbots for kids). Bores spans all quadrants. RAISE is long-term frontier safety. His NY AI Consumer Protection Act (modeled on Colorado) is near-term algorithmic discrimination. Training data transparency (A6578, passed Assembly unanimously) is near-term general disclosure. Strict liability (A8833) is long-term economic incentive alignment — make developers pay for catastrophic harm regardless of care exercised.

This taxonomy explains something attack ads flatten: Bores voted no on several employment and lending AI bills (A3779, A773) while leading RAISE. His stated reasons: bills imposed costly assessments without changing standards, didn’t cover national banks that dominate the market, or froze state AI adoption for ten years. Lasher co-sponsored those bills and campaigned on Bores’s no votes as proof of inconsistency. Bores’s counter: targeted, implementable regulation beats performative guardrails — read the bill, fix the scope, or don’t pass it.

Both can be true. He is genuinely more selective about which regulations he supports. He is also the public face of frontier safety in a state that passed the country’s strongest transparency law. Those facts coexist uncomfortably.

The money war

NY-12 — the safely Democratic Manhattan seat Bores sought to succeed retiring Jerry Nadler — became the highest AI PAC spend of any 2026 primary.

SideActorSpend (approx.)
Anti-regulationLeading the Future super PAC (Palantir/OpenAI-aligned funders per press)$6–7M+ attack ads
Pro-regulationChris Larsen (Ripple co-founder)$3.5M
Pro-regulationVarious safety-aligned donors~$4.6M total pro-Bores

Politico described the strategy: “Beat up on Bores so badly that when the idea of regulating AI development comes up, other politicians run the other direction.”

The paradox: attack ads may have helped — they elevated Bores’s profile, fed an anti-oligarch narrative, and landed him in the New York Times, Vanity Fair, the Ezra Klein Show, and Time 100 AI 2025.

Only ~12% of his campaign contributions came from within NY-12 (February 2026). This was a national fight fought in a Manhattan primary.

The primary

Eight candidates. Early vote June 13–21. Election day June 23.

Bores’s rival Micah Lasher — former Hochul policy director, co-sponsor of RAISE, champion of SAFE for Kids (social media algorithm restrictions for minors) — ran a broader tech accountability frame: labor, lending, data centers, civil rights, plus frontier safety. Hochul endorsed Lasher and weakened RAISE before signing it — signaling support for AI regulation, but not Bores’s specific lane.

Result: Lasher won ~39% to Bores’s ~35%, a ~4,000-vote margin. NBC/CBS/CNN called it election night. Schlossberg finished third at ~11%.

Leading the Future got its stated goal: Bores is not going to Congress. But 35% second place in an eight-way race, after $7M in attack spending, is not a repudiation of frontier regulation. Lasher co-sponsored RAISE. The industry blocked the author, not the policy.

What Bores still has

Bores remains Assemblymember for AD-73. RAISE is still law. He leads A9449 (chapter amendments aligning operative text with the Hochul compromise), A6578 (training data transparency, Assembly passed), and A10379 (unsafe AI chatbots for minors, passed legislature June 2026, awaiting Hochul signature).

His federal 43-point framework — nationalize RAISE, AI dividend, training data transparency, international safety standards — won’t be introduced by him in Congress. If Scott Wiener wins CA-11 in November, Wiener may become the sole frontier-law author in the House.

The parallel race: Wiener

California’s Scott Wiener won his June 2 primary with 41.3% in a crowded field — same policy lane, different outcome. He authored SB 53. He is expected to carry federal SB 53 nationalization if elected to CA-11 in November.

Two frontier transparency authors ran for Congress in 2026. One won the primary. One lost. The policy template survived both races.

What this means if you care about AI safety

For researchers: SB 53 and RAISE are now law (or becoming law). If you work at or evaluate frontier labs operating in California or New York, mandatory safety frameworks and incident reporting are not voluntary best practices — they are legal obligations. The definitions of “catastrophic risk” and “critical safety incident” in these bills are worth reading alongside your eval work.

For governance: The state-lab strategy worked partially. Two major states passed transparency laws industry fought hard to kill. The federal response is not to copy them cleanly — it is to federalize a version and preempt states for three years (GAAIA) or offer voluntary review with classified benchmarks (Trump EO). The Cruz 99–1 defeat shows broad preemption can be beaten. Narrower preemption in a bipartisan bill is harder to stop.

For political economy: NY-12 proved frontier AI regulation is now a national PAC issue. The spend ratio — $7M to block one candidate vs. ~$200–400M global AI safety funding — tells you where power sits. Industry’s goal is chilling effect: make the next state legislator think twice before authoring a RAISE equivalent.

For individuals: As I wrote in how to participate in AI safety, governance lacks public pressure more than policy ideas. 70–86% of Americans support AI regulation. Calling your representative about opposing broad federal preemption of state AI safety laws is one of the highest-leverage actions available — including if you can’t vote.

Open questions

  • Does Hochul sign A10379 (chatbot restrictions for minors) and the data center moratorium (A11560) before November?
  • Does the Senate pass A6578 (training data transparency) in 2026?
  • Does GAAIA’s preemption survive stakeholder feedback, or get stripped like Cruz’s moratorium?
  • If Wiener reaches Congress, does a frontier-transparency coalition (Wiener + Anthropic + Encode) beat an omnibus civil-rights frame (Lasher + labor + AG James)?
  • Does Bores’s loss chill other state legislators — or does RAISE on the books matter more than who carries it to Washington?

The US still has no federal AI safety law. But for the first time, it has state laws that frontier labs must comply with — and a political fight fierce enough to spend seven figures keeping one assemblyman out of Congress. That is not nothing. It is also not enough.

Sources