The Next Three Phases of AI Politics

Abstract

US AI politics in 2026–2027 will move through three phases: attempts at federal preemption that Democrats have little incentive to accept before the midterms, a brief legislative window immediately after the new Congress is sworn in, and intensifying politicisation once presidential primary season begins. Because the post-midterm window is short and driven by a sense of needing to act rather than by policy quality, the productive work is preparing the best versions of politically likely ideas rather than designing ideal frontier policy in isolation.

Framing

• 2025 ended with AI politics unfinished: an executive order promising legislative action that did not come, and rising political salience with no policy vehicle to attach it to.
• 2026 is set to be heavy in politics and light in policy.
• The moment is hard to read, with countervailing factors: super-PACs and political funding ramping up on both sides but not showing effects until after the midterms; rising salience that could fizzle without popular policy ideas to latch onto.
• One can equally paint a picture of an anti-AI techlash or of a Congress paralysed by fear of unprecedented industry political spending — which is why many factions all believe things can only improve for them.
• The useful analytical move is looking at how the timing of these effects lines up: when politics are worst, when best, when to play defence and when offence.
• Core recommendation: prepare for the window right after the midterms, and get something through before the primaries close it again.

Phase 1: One last preemption push

• 2026 begins with an accelerationist coalition compelled to cash a cheque David Sacks wrote in the form of EO 14365, the latest attempt at federal preemption of state frontier AI laws, controversial with the public and within the GOP.
• The EO followed failed congressional preemption attempts and the signing of SB-53 and a weakened RAISE Act.
• It is not a complete policy victory. Experts debate the legality of its provisions; drawn-out court battles could remove its teeth and leave only a vague chilling effect. Even supporters frame it as a stopgap preceding a national legislative framework — Congress and the public alike feel AI should ultimately be addressed legislatively.
• OSTP Director Kratsios has reaffirmed intent to produce a proposal this year, but Congress is unlikely to simply pass a White House proposal as a standalone law.
• Vehicles are scarce. A last-minute NDAA preemption attempt failed in the fall, and no comparable must-pass bills remain. Attaching contentious preemption to a continuing resolution risks a shutdown fight; a controversial AI rider could jeopardise priorities like permitting reform. The NDAA episode suggests such attempts are easily sunk by single influential legislators, and become less likely each time last-minute attempts burn goodwill.
• The most promising accelerationist play is attaching the framework to a standalone bill on a specific AI harm, most probably child safety — twice attractive because it is easiest to rally the GOP behind and hardest for Democrats to refuse. Internal preemption opponents would like credit for a child safety law, and AI companies recognise the associated PR risks. Several child safety bills are in committee and could serve as vehicles.

Why Democrats may not move

• Congressional Democrats are well positioned to play for time: anything obtainable now is probably obtainable next Congress on better terms and with more credit to them. Previous accelerationist pushes have left them sceptical of good-faith offers. Meanwhile they can keep running anti-tech and anti-AI campaigns.
• Three potential sources of leverage over them:
  • Democrats would rather not be at odds with tech-affiliated donors and their super-PACs, especially the accelerationist “Leading the Future”; leadership was already somewhat sympathetic to a preemption deal in the fall.
  • It is hard for Democrats to argue publicly why they don’t want to move fast — the object-level reason (they expect nothing beyond SB-53 and RAISE) can’t be stated without admitting they’re content to let Gavin Newsom and Kathy Hochul run national AI policy.
  • Accelerationists might offer something too good to refuse — a substantively strong bill convincing child safety advocates would be hard to kill going into an election year with a hostile PAC in the field.
• The author’s expectation: pro-regulation super-PACs and rising salience mean Democrats may not fear the Leading the Future threat much, and accelerationist messaging is unlikely to produce a package good enough to make opposition look unreasonable. Best guess is deadlock and litigation of the EO, with little else until the midterms.

Phase 2: A post-midterm window

• The midterms change things in two ways: they remove the barrier to compromise created by wanting to campaign on an issue, and they supply much more data on the politics of AI policy through polling and candidate performance.
• That data will be less clarifying than hoped. The NY-12 case is instructive: RAISE Act sponsor Alex Bores faces a crowded primary, and while there has been safetyist triumphalism about Leading the Future buying him free media by targeting him publicly, on base rates he is likely to lose in a crowded field — in which case the super-PAC still gets a visible win. More generally, because AI is not yet election-deciding, much of the signal will be drowned in broader political noise.

The salience story

• Perceived political salience of AI will keep increasing as yesterday’s frontier systems enter mainstream application for both impressive and harmful purposes, generating attention, reporting, and policymaker interest.
• Even if this happens less dramatically than boosters expect, the meme has reached escape velocity — everyone already “knows” AI will be politically big, which can become self-fulfilling.
• Against the critics’ issue-polling objection: just as “as soon as it works, we don’t call it AI anymore,” so “once it’s salient, we don’t call it AI anymore.” AI will register within big-ticket issues — tech oligarchy, economic equality, job prospects, environmental harms — that consistently poll as important.
• Once that shows up around the midterms, paired with clear policy-level polling that voters want legislation, Congress will identify AI as an issue it could touch but has stayed silent on. No self-respecting lawmaker passes on putting their name to a bill regulating something they and the electorate deem important.

The window itself

• A policy window for substantial congressional action may open shortly after the new Congress is sworn in. Far fewer lawmakers will be satisfied doing nothing, removing the greatest current barrier — too many people content to wait.
• Any resulting law would come from triangulation and negotiation, with provisions giving every important voice a win: broader preemption, national security provisions, narrow substantive rules on current harms, federal codification of SB-53, and others.
• Presidential signature seems likely: vetoing a rare congressional consensus on AI would be too much of a political liability for VP and presidential hopeful Vance.
• Whether the law would be good is less clear. The dynamic is driven by a sense of having to do something rather than by the quality of any proposal, and rounds of haggling over language make it easy to lose legislative nuance. The result could pair broad preemption without substantive stipulations against equally myopic provisions addressing only incidental current harms.

How to use it

• The task is not devising a clever one-size-fits-all framework but identifying the best versions of bad ideas. Policymakers’ political priorities will remain fairly immutable, so ask what the best legislation based on each priority would be: what labour stipulation doesn’t just sound good, what child safety measure actually addresses the underlying scalable problems of deception and sycophancy.
• The answer cannot be “do these things and also do very clever frontier policy on top” — that risks frontier policy being thrown out when it conflicts with more politically salient goals.
• To insulate good policy, the political driver and the policy merit must be closely interwoven. Isolated frontier safety policy is rarely effective; specific areas of frontier policy must be linked to specific areas of near-term public salience — e.g. in child safety, arguing against age gates and for evals-based solutions targeting sycophantic and deceptive tendencies; in labour, defining appropriations mechanisms for expandable safety nets early to make policy scalable.
• Two attractors to avoid: piecemeal solutions that don’t advance mid-term policy goals (dismissive of how rare good windows are), and purely horsetrade-based policy that doesn’t anchor frontier policy in politically important issues (dismissive of the true political drivers).

Phase 3: Politics at last

• AI politics will intensify once presidential primary season begins — usually expected around late 2027, when candidates start building national profiles, testing messages, and considering policies.
• Incentives for politicisation will be sharper than in the midterms, partly from rising salience and partly because primaries reward candidates who carve out a niche distinguishing themselves from the party mainstream.
• Republican side: frontrunner JD Vance will hold the bag on the Trump administration’s AI record, leaving a gap for tech-sceptical voices to attack him if the administration stays accelerationist, or for technooptimists to contest his Silicon Valley support if he pivots. Josh Hawley and Ron DeSantis are already lining up for the tech-sceptical angle.
• Democratic side: AI looks like a likely point of contention between moderates and left-wing populists, the latter already making headlines with deeply anti-AI views intersecting with distaste for big tech, billionaires, environmental harms, and labour disruption. Leading moderate Gavin Newsom cannot fully pivot to an anti-AI position without upsetting his California donor and supporter base.
• Even setting aside the salience story, these dynamics make AI politicking attractive as a wedge in two crowded primaries. Once that happens, legislative windows close again: campaigning on an issue creates incentives not to make policy beforehand, since compromise reduces the issue’s profile. With thin majorities, a few electorally motivated defectors can sink legislative attempts.

What follows, by faction

• Accelerationists need to think hard about using borrowed time. Winning legislation before the midterms requires a better lever to move Democrats; the author’s preferred path remains compromising around deep and narrow regulation in exchange for broader preemption. If a law this year is unlikely, it may be better to stop posturing and retain capital and good faith for negotiations in the next Congress.
• Safety advocates and regulation proponents face a comfortable few months of playing defence, but must use them to prepare. The political drivers of any coming push will not align perfectly with reasonable advocates’ priorities, so translation and groundwork are needed to harness those drivers for good policy. Squandering the opportunity risks ending up supporting AI laws that do little while confirming a reputation as unabashedly pro-any-regulation.
• General lesson: the most politically likely times for policy action rarely guarantee good policy. That opens two avenues — making good policy more likely in unlikely moments, or making bad policy better in likely moments. The mistake is optimising for policy quality when politics are prohibitive, or for political viability when politics are already favourable.