A political map of US AI policy

In February I mapped global AI governance: institutions, summits, the EU AI Act, industry self-regulation, and the 600:1 safety-to-capability spending gap. That map was about what exists on paper.

This one is about US domestic politics — who is fighting over AI, with what tools, at what level of government, and with how much money. The headline from February still holds: the United States has no comprehensive federal AI safety law, and the executive branch has been working to prevent state laws from filling the gap. But the four months since have been among the most consequential in American AI politics: two states passed frontier transparency laws, a bipartisan federal draft emerged with a preemption clause that could undo them, and congressional primaries became national referendums on whether frontier AI should be regulated at all.

If you only read one thing before engaging with US AI policy — as a researcher, advocate, or voter — understand that the fight is not Democrats vs. Republicans. It is at least seven factions, often aligned on specific votes and opposed on others, fighting over three different regulatory templates at three levels of government.

The one-page mental map

                    PUBLIC (70–86% want regulation; AI not yet top-5 election issue)
                                        │
        ┌───────────────────────────────┼───────────────────────────────┐
        │                               │                               │
   STATE LAWS                    FEDERAL FIGHT                   2026 RACES
   (where policy is              (who overrides whom)            (where money is)
    actually passing)
        │                               │                               │
   ┌────┴────┐                    ┌─────┴─────┐                   ┌─────┴─────┐
   │Frontier │                    │ Trump EO  │                   │ Wiener    │
   │transparency│                 │ voluntary │                   │ CA-11     │
   │SB 53    │                    │ 30-day    │                   │ Bores ✗   │
   │RAISE    │                    │ review    │                   │ NY-12     │
   ├─────────┤                    ├───────────┤                   │ Rutinel   │
   │Consumer │                    │ GAAIA     │                   │ CO-8      │
   │/ labor  │                    │ mandatory │                   └───────────┘
   │CO SB205 │                    │ audits +  │
   │NY A768  │                    │ 3yr       │
   ├─────────┤                    │ preemption│
   │Infra    │                    └───────────┘
   │DC pause │
   │Sanders  │
   │S.4214   │
   └─────────┘
        │
   SEVEN FACTIONS (not two parties):
   1 Frontier transparency (Wiener, Bores, Encode, Anthropic on SB 53)
   2 Omnibus Dem accountability (Lasher, Jeffries, Warren — civil rights + labor + kids)
   3 Near-term consumer AI (Rutinel, Colorado algorithmic discrimination)
   4 Progressive infrastructure pause (Sanders, AOC, env/labor coalitions)
   5 Populist right anti-oligarch (Bannon, parts of MAGA)
   6 Accelerationist / industry (Sacks, Leading the Future, parts of GOP leadership)
   7 Bipartisan federal framework (Obernolte-Trahan GAAIA — transparency + preempt states)

The central tension: States are writing laws. Industry wants one weak federal standard to replace them. Progressives want a federal pause on infrastructure. Populists on left and right distrust Big Tech oligarchs but disagree on solutions. And 2026 congressional primaries are where these abstractions turn into seven-figure ad buys.

What Americans actually think

Before the factions: the electorate. 70–86% of Americans support AI regulation in general (KPMG global, Eurobarometer, US anti-AI populism polling). Specific policies poll even higher and bipartisan:

Policy	Support	Source
Ban deepfakes in political ads	84% (R 83%, D 86%)	PPC/UMD 2024
Government testing before AI deployment	81% (R 76%, D 88%)	PPC/UMD 2024
Mandatory deepfake labeling	83–91%	PPC/UMD, Carnegie
Six-month pause on large-scale AI development	69%	YouGov 2023

Democrats are somewhat more pro-regulation on testing and oversight. Republicans match or exceed Democrats on deepfake bans and child safety. This is not a culture-war issue in polling — it is a high-consensus, low-salience issue. AI rarely ranks in voters’ top five priorities; healthcare, the economy, and immigration consistently dominate.

The pattern that matters politically: people express concern in surveys but keep using AI daily (the “say vs. do” gap). That makes it hard to build a mass movement against AI itself — but easy to build coalitions around specific harms: deepfakes, chatbots harming kids, algorithmic discrimination, data center electricity bills.

Sanders’s data center moratorium went from “fringe” (December 2025) to 100+ local moratoria and 12 state proposals (March 2026, per his office) — not because voters ranked “AI” as issue #1, but because electricity prices and local land use are tangible.

Implication: Politicians will frame AI through jobs, kids, bills, and oligarchs — not through alignment theory or P(doom). The safety research community’s concepts (catastrophic risk, deceptive alignment, RSI) enter politics only when translated into those frames.

Layer 1: State legislatures — the laboratory

After California’s SB 1047 was vetoed by Newsom in September 2024, the state-level strategy shifted: transparency and accountability first, not training bans. Get labs to document safety practices before asking legislatures to halt capability development.

Three templates emerged.

Template A: Frontier transparency (California + New York)

California SB 53 (Sen. Scott Wiener, signed September 2025, effective January 2026). Anthropic endorsed it. Applies to large frontier developers (~10²⁶ FLOPs training, $500M+ revenue). Requires published Frontier AI frameworks, transparency reports before deployment, critical safety incident reporting to Cal OES (15 days / 24h emergency), whistleblower protections. Penalties ~$1M/violation via state AG.

New York RAISE Act (Asm. Alex Bores + Sen. Andrew Gounardes, signed December 2025, weakened by Gov. Kathy Hochul before signing, effective January 2027). Similar obligations: safety protocols, third-party evaluation, 72-hour incident reporting, whistleblower protection, DFS oversight office. Original penalties up to $30M; Hochul cut toward California scale.

Politico called RAISE “widely regarded as the furthest-reaching in the country.” Industry fought both bills. Encode helped pass both. These laws do not pause training or require government pre-approval. They make voluntary safety practices mandatory.

Template B: Consumer / algorithmic discrimination (Colorado + spreading)

Colorado SB 24-205 (2024) — first comprehensive state AI consumer protection law, requiring developers and deployers of “high-risk” AI systems to use reasonable care against algorithmic discrimination. Effective February 2026. Lead sponsor in the 2025 session expansion: Rep. Manny Rutinel.

This is the near-term, optimistic-general quadrant: AI affecting hiring, credit, housing, healthcare decisions today, not catastrophic risk tomorrow. Industry often prefers fighting this to fighting frontier transparency — but consumer groups and civil rights organizations prioritize it.

Illinois, NYC (Local Law 144 on hiring algorithm audits), and New York’s pending A768 (Bores) extend this lane.

Template C: Infrastructure pause (Sanders + states + local)

Separate from frontier model law: data center moratoria. Sanders’s S.4214 (March 2025, with AOC) would pause new/upgraded AI data centers federally until comprehensive AI legislation passes covering safety, worker protection, and community/environmental standards. New York’s A11560 (passed legislature June 2026, awaiting Hochul) imposes a one-year state moratorium on large data center permits with prevailing wage and ratepayer protection requirements.

This regulates the physical substrate of capability — compute, power, water — not model weights directly.

State preemption skirmishes

Governors matter. Newsom vetoed SB 1047 but signed SB 53. Hochul signed RAISE but weakened it and endorsed Bores’s primary opponent. Both signal: transparency yes, hard pause no, and we pick which Democrats get credit.

Layer 2: Federal politics — three competing visions

The vacuum (still)

Biden’s EO 14110 (2023) → Trump revoked (January 2025). CAISI (~$10M budget) survives in narrowed form. No comprehensive statute.

Vision 1: Voluntary federal review (Trump EO, June 2, 2026)

Frontier developers may choose up to 30 days pre-release government review. CISA/NSA/Treasury define covered models within 60 days. Cannot be used for mandatory licensing. NSA leads benchmark definition; CAISI advisory only. No mandatory incident reporting. Hands-off.

Vision 2: Mandatory federal framework + state preemption (GAAIA, June 4, 2026)

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

For large frontier developers: public safety frameworks, semi-annual third-party audits (CAISI-certified IVOs), mandatory critical incident reporting, penalties up to $100M/day, CAISI codified (~$100M/year FY27–29).

Plus: 3-year preemption — states cannot enact new laws “specifically regulating the development” of AI models. General laws (civil rights, labor, privacy) and post-deployment rules preserved. California training-data transparency (AB 2013) would be preempted. Frontier laws in CA, NY, IL “federalized.”

Encode, Wiener, and Public Knowledge oppose preemption. Industry often wants federal rules — if they override state laws already fought.

Vision 3: Infrastructure moratorium + comprehensive legislation (Sanders S.4214)

Federal pause on AI data center construction until Congress passes laws ensuring: (A) AI product safety before release, (B) wealth sharing / anti-displacement, (C) community/environmental standards. Export ban on AI compute infrastructure to countries without equivalent safeguards.

Almost certainly will not pass the 119th Congress. Value is agenda-setting and linking to 100+ local moratoria.

The preemption wars — and strange bedfellows

Critical distinction: two different “moratoriums” get conflated in media:

Term	Who pushes it	What it pauses
State-law preemption moratorium	Cruz / Trump / Sacks	States cannot enforce AI laws for ~10 years
Data center moratorium	Sanders / AOC	New data centers until federal comprehensive AI law

Sanders wants to slow infrastructure; he opposed Cruz’s state preemption (aligned with Bannon and Blackburn on that single vote). The Senate killed Cruz’s broad preemption 99–1 in July 2025 — one of the few issues where progressive state-law advocates, Republican governors (DeSantis: “denying the people the ability to channel these technologies via self-government”), and MAGA populists (Bannon) overlapped.

GAAIA’s narrower 3-year preemption is harder to kill because it comes wrapped in mandatory audits and bipartisan sponsors.

GOP is not monolithic

GOP faction	AI stance	Example
Accelerationist / tech-right	De-regulate, win China, federal preemption	Sacks, parts of Trump WH
Federalist / culture-war	Block broad preemption; keep state child safety, deepfake, musician protection	Blackburn (killed Cruz; pushes TRUMP AMERICA AI Act with KOSA/NO FAKES carve-outs), DeSantis
National-security hawk	Export controls, visibility into RSI/internal models, CAISI in testing	Banks (AI OVERWATCH Act, June 2026 letter on recursive self-improvement)
Bipartisan framework	One national standard	Obernolte GAAIA

Sen. Jim Banks (R-IN) is not a pause advocate — but his June 2026 letter to the Trump cabinet urged treating recursive self-improvement and unreleased internal models as national security issues, and pushed CAISI into testing (weakening NSA-only control). Same week, Anthropic published on collective slowdown; OpenAI released an RSI policy paper. The fight is over who sets standards and who sees models, not regulation vs. no regulation.

Layer 3: Democrats — three lanes, not one party

It is tempting to say Democrats are “more pro-AI regulation.” Elites in the party are more likely to sponsor AI bills. But Democrats are fighting each other over which template wins.

Lane 1: Frontier transparency

Wiener (SB 53), Bores (RAISE), Encode, Anthropic (endorsed SB 53). Frame: catastrophic risk, mandatory safety plans, incident reporting, whistleblowers. Federal goal: nationalize SB 53/RAISE.

Lane 2: Omnibus accountability

Lasher (won NY-12 primary over Bores), Jeffries (House Dem AI Commission), Warren, AG Tish James. Frame: AI civil rights, hiring/lending bias (A3779, A773), kids (SAFE for Kids, chatbot bill A10379), data centers, deepfakes — frontier safety as one piece, not the whole brand.

Lasher co-sponsored RAISE and attacked Bores for voting no on employment/lending bills. Hochul signed weakened RAISE and endorsed Lasher — establishment signal: regulate AI, but not only through the frontier-only CS-legislator lane.

Lane 3: Near-term consumer protection

Rutinel (Colorado SB 24-205 lead, running for CO-8). Frame: algorithmic discrimination, reasonable floor for businesses, protect workers and consumers now. Not x-risk primary.

Lane 4 (overlapping): Progressive pause

Sanders, AOC. Frame: oligarchs, jobs, electricity bills, democracy catching up. Cites Hinton and FLI pause letter but master frame is economic justice, not EA alignment. Leadership (Jeffries) sympathetic but does not co-sponsor moratorium.

Net: No unified “Democratic AI policy.” NY-12 was Lasher’s omnibus lane beating Bores’s frontier lane. CA-11 is Wiener’s frontier lane winning. CO-8 (primary June 30) is Rutinel’s consumer lane vs. Shannon Bird’s innovation-friendly moderation.

Layer 4: Populism — left, right, and misaligned alliances

Populist left (Sanders / AOC / env-labor)

Anti-oligarch master frame: “a handful of billionaire Big Tech oligarchs” deciding humanity’s future without democratic debate
Data center moratorium, wealth sharing, pre-release product review
200+ organizations (Food & Water Watch, nurses unions, etc.) backed federal moratorium letter

Populist right (Bannon / parts of MAGA)

Names Altman, Musk, Amodei, Hassabis, Sacks as not acting in US interests
Demands pre-deployment vetting; fears employment and deepfake/truth crises
Helped kill Cruz preemption 99–1 — not aligned with Sanders on infrastructure pause

	Shared	Not shared
Left + right populists	Distrust Big Tech; want vetting; anti-oligarch rhetoric; killed broad preemption	Sanders wants wealth redistribution; Bannon doesn’t. Sanders pauses data centers; Bannon doesn’t necessarily.
Populists + frontier transparency advocates	Want oversight before harm	Populists frame jobs/democracy; frontier advocates frame catastrophic risk and evals
Populists + industry	Nothing	Everything

Foreign Affairs, DNYUZ, and Tech Policy Press documented anti-AI populist backlash building across both parties in early 2026. It is real but fragmented — not yet a coherent third force, more a set of overlapping grievances industry lobbyists exploit when convenient.

Layer 5: Money — where AI politics becomes expensive

2026 is the first cycle where AI regulation is a seven-figure primary issue.

Race	Candidate	AI policy lane	Notable outside spend
NY-12 (Jun 23)	Alex Bores (lost ~35%) vs Micah Lasher (won ~39%)	Frontier vs omnibus	Leading the Future $7M+ anti-Bores; Larsen $3.5M pro-Bores
CA-11 (Jun 2)	Scott Wiener (won 41.3%)	Frontier (SB 53)	High spend; Wiener first in top-two primary
CO-8 (Jun 30)	Manny Rutinel vs Shannon Bird	Consumer vs innovation	$5.5M+ pro-Rutinel per Axios; Anthropic/Google/Meta/OpenAI employees $265K direct

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

Outcome in NY-12: Industry got its stated goal — Bores blocked from Congress. But 35% second in an eight-way race after $7M in attacks is not voter repudiation of regulation. Lasher co-sponsored RAISE.

Compare to global AI safety funding: ~$200–400M/year worldwide for safety and governance vs. $7M in one primary to stop one assemblyman. The political economy of frontier regulation is dominated by industry offense, not safety-community defense.

Pro-regulation money exists (Larsen, Anthropic employees, Encode ecosystem) but concentrates on candidates who can win, not movement infrastructure.

Layer 6: Industry and safety-adjacent actors

Actor	Role in US AI politics
Anthropic	Endorsed SB 53; employees donate to Rutinel; RSP as corporate model; Dario Amodei 2026 shift toward FAA-style mandatory testing
OpenAI	Mixed; superalignment team dissolved 2024; RSI policy paper June 2026; Altman meets White House
Leading the Future	Super PAC; ~$7M anti-Bores; Palantir/OpenAI-aligned funders per press
Encode	Youth org; passed SB 53 and RAISE; opposes GAAIA preemption
Data Center Coalition	Opposes Sanders moratorium
CAIS / METR / GovAI	Research; inform policy but tiny budgets vs. industry lobbying

Industry self-regulation (Anthropic RSP, etc.) remains more detailed than most government policies — but competitive pressure and political spending are the actual governance mechanisms in the US, not RSP compliance alone.

How the layers interact — three scenarios for 2027

Scenario A: State lab wins → federal floor follows

SB 53 and RAISE enforce through 2026–27. Wiener (and possibly Rutinel on consumer provisions) reach Congress. GAAIA preemption stripped or expires. Federal law codifies state template with teeth. Encode/Wiener playbook: states first, then nationalize.

Scenario B: GAAIA preemption wins

3-year freeze on state development-stage rules. SB 53/RAISE development obligations federally superseded. Industry gets one weaker standard. Chilling effect on next state author confirmed post-Bores.

Scenario C: Populist infrastructure politics escalates

Data center moratoria spread; electricity prices become local election issue; Sanders frame goes mainstream without S.4214 passing. Frontier transparency and infrastructure pause decouple — different coalitions, different bills.

All three can partially coexist. The mistake is assuming one faction “wins” AI politics outright.

What this means for AI safety researchers

Your eval work connects to law. SB 53 and RAISE define “catastrophic risk” and “critical safety incident” in statutory language. Read them alongside METR/Apollo outputs.
Federal preemption is the highest-leverage political object — more than another eval benchmark. Cruz 99–1 proved broad preemption can die. GAAIA’s narrower version is the live threat.
Framing matters. Sanders enters politics through jobs and oligarchs; Bores through CS credentials and nuclear analogies; Wiener through Encode partnership; Lasher through kids and civil rights. Same underlying demand for oversight, four different voter interfaces.
Don’t map Dem=safe, GOP=unsafe. Blackburn saved state AI laws from Cruz. Obernolte wrote GAAIA. Banks pushes RSI visibility. Sacks pushes preemption.
Salience will lag consensus. 81% want pre-deployment testing; AI still isn’t top-five in voter priority. Mobilization requires local, tangible hooks — data center bills, wrongful denial of benefits, teen chatbot harm.

For practical entry points: how to participate in AI safety. Calling representatives to oppose broad federal preemption of state AI safety laws remains high-leverage even if you cannot vote.

Open questions (June 2026)

Does Hochul sign A10379 (unsafe AI chatbots for minors) and A11560 (data center moratorium)?
Does CO-8 nominate Rutinel (June 30 primary)?
Does Wiener win CA-11 general (November)?
Does GAAIA preemption survive stakeholder feedback?
Does Lasher in Congress nationalize any RAISE elements or only omnibus civil-rights frame?
Does Bores’s loss chill other state frontier-safety authors — or does RAISE on the books matter more?
Does a trigger event (mass layoffs attributed to AI, election deepfake, data center rate hike) move AI into top-tier salience before 2028?

How this fits the other maps

Article	Scope
Global AI governance landscape	International institutions, EU/China/US vacuum, industry self-regulation, funding gap
This article	US domestic factions, state templates, federal preemption, 2026 races, populism, money
US state policy & Bores case study	Deep dive on NY RAISE, NY-12 primary, Bores vs Lasher

The US story in one sentence: policy is being written in state capitols, fought over in primaries, and threatened by federal preemption — while the public wants regulation but hasn’t yet made AI a top-tier election issue.