{ "question_id": "Q9", "slug": "us-federal-regulation", "title": "US federal frontier-AI regulation", "type": "level", "levels": [ { "key": "L1", "label": "Transparency" }, { "key": "L2", "label": "Independent audits" }, { "key": "L3", "label": "Pre-deployment evaluation" }, { "key": "L4", "label": "Halt authority" } ], "resolution": { "levels": [], "values": [], "annotations": [] }, "forecasts": [ { "question_id": "Q9", "forecast_date": "2026-09-14", "evidence_cutoff": "2026-09-14", "reading_adopted": null, "ambiguities": [ "The text does not expressly distinguish frontier-model safety regulation from pre-existing general emergency powers, export licensing, or conventional cyber-incident reporting. A broad reading could treat earlier instruments as qualifying; I adopt the frontier-model regulatory reading described in reading_adopted.", "I read 'publish a safety framework' as public disclosure and 'critical incidents' as model-related safety/security incidents. Broader readings of confidential reporting could change the treatment of Executive Order 14110.", "The policy mechanisms are not logically nested: an emergency halt power can exist without recurring audits or routine pre-deployment approval. I follow the requested highest-level cumulative scoring rather than requiring each lower mechanism separately. Federal halt authority subject to judicial review or emergency procedural limits still counts.", "Both live bipartisan vehicles (FRONTIER Act H.R. 9925; Thune-Cruz-Klobuchar Senate draft) contain halt/block authority but the FRONTIER Act has no pre-deployment gate, so the raw probability of an L4-type provision exceeds that of an L3-type gate; monotonicity is preserved only by treating L4 as implying L3.", "Whether a Commerce Secretary power to seek a federal court injunction blocking a release (as in the leaked Senate text) is 'a federal body has authority to order a halt' is unclear; I treat direct suspension authority (FRONTIER Act) as L4 and a court-injunction-only mechanism as not L4.", "Whether the Senate bill's 'duty of care' plus self-testing presented to Commerce would constitute 'mandatory transparency with penalties' depends on final text; I require explicit framework-publication and/or incident-reporting duties with civil penalties.", "Whether mandatory cyber-security incident reporting (e.g., a CIRCIA rule extended to AI developers, or a finalized BIS model-weight security rule with penalties) would satisfy L1's 'report critical incidents' clause. I adopted a narrow reading: L1 requires AI-safety (not merely export-control/cyber) incident reporting or safety-framework publication; under a broad reading, far-horizon L1 probabilities would rise by roughly 5-10 points.", "Whether the FRONTIER Act's emergency suspension authority and the Senate bill's release-blocking power count as L4 'halt authority' when hedged with court review; I count conditioned/emergency halt powers as L4.", "Whether an instrument's threshold 'captures' Google DeepMind given it is a subsidiary of Alphabet; I treat any revenue/compute threshold that in practice binds all three major labs as satisfying the criterion.", "Whether levels are a Guttman scale (L4 requires L1–L3 in the same instrument) or independently scored features. FRONTIER Act has L1+L2+L4 but not L3; Guttman scoring would cap it at L2.", "Whether a DPA executive order or Commerce information-collection rule requiring capability/red-team reports (Biden EO 14110 template) satisfies L1's 'safety framework and/or critical incidents' language.", "Whether an EO or rule that is promptly enjoined still 'imposes obligations enforceable against private developers.'", "Whether 'deployment' in L3 includes export/API access restrictions or only domestic public release.", "Whether an instrument that meets L4's halt-authority criterion but not L3's pre-deployment-evaluation criterion counts as 'an instrument at level >= L4' (reading adopted: yes, it does, consistent with the schema's monotonicity requirement). Under strict nesting, such an instrument would count only as L2, which would lower my L3/L4 probabilities by ~5-8 points in mid horizons.", "Whether a regime where the government must obtain a federal-court order to block a release (as in the reported Senate draft) satisfies L4's 'a federal body has authority to order a halt'. I treat agency-initiated, court-ordered blocking as qualifying, with a probability discount for the dispute.", "Whether FCC forfeiture penalties under a hypothetical FCC disclosure rule (directed for consideration by the Dec 2025 EO) would count as 'penalties for non-compliance' for L1; treated as a minor pathway.", "Retroactivity: read literally, 'an instrument at level >= L has been enacted on or before the last day of the quarter' could capture pre-2026 statutes. If ECRA (2018) were held to confer L4 authority — because BIS used it in June 2026 to force a global suspension of Anthropic's Fable 5/Mythos 5 and to restrict OpenAI's GPT-5.6 Sol launch — then all four levels would already be YES at every horizon. I reject that reading (no frontier-developer threshold; the action was an 'is informed' letter, not a rule; the question is plainly forward-looking), but it is the single largest source of downside risk to my numbers.", "L4 vs. court injunction: the Cruz/Thune/Klobuchar draft reportedly lets the Commerce Secretary REQUEST a federal court injunction against an unsafe model (Transformer, 11 Sep 2026). Whether 'a federal body has authority to order a halt' covers a petition-the-court mechanism is ambiguous. I treat it as borderline and only partly credit it.", "L1's penalty requirement: a purely voluntary certification regime (as the Cruz draft was described by a source who saw the text) would NOT meet L1 ('Codification of a voluntary framework without penalties does not count'), even if branded 'AI safety legislation'. I assume roughly a third of bills that actually get signed in this space fall below L1 on this ground, which depresses my L1 numbers relative to headline 'will Congress pass an AI bill' markets.", "DeepMind's inclusion: the FRONTIER Act's tiers key on an 'AI developer' entity's revenue and AI-related development expenditures. Whether Google DeepMind (a division of Alphabet with no separate revenue) is captured depends on affiliate aggregation. The bill aggregates affiliates, so Alphabet's figures almost certainly capture it; I assume the threshold test is satisfied for all the leading bills.", "Level nesting vs. real bill architecture: the FRONTIER Act contains L1, L2 and L4 features but arguably not a true L3 preclearance gate. Under the question's stated nesting I count it as reaching L4 and therefore >= L3.", "Whether a company-specific BIS 'is-informed' letter, or an industry-wide Federal Register export-control rule controlling frontier model access, counts as an 'instrument' at L4. I assumed the letter does not count but that a future published final rule would (folded into a modest L4 increment).", "Whether a statute requiring developers to self-run safety tests and submit results to Commerce for deployment approval satisfies L3 ('government or certified third-party capability evaluation'). I treated it as partially satisfying, so L3 sits slightly below L2.", "Whether the required frontier-developer threshold may be implicit (e.g., an NDAA incident-reporting rider). I assumed the instrument's own threshold must capture OpenAI, Anthropic and Google DeepMind." ], "key_drivers": [ "Concrete House legislation and active bipartisan Senate negotiations now include independent oversight and risk-based intervention powers, but no qualifying enactment was found.", "The June 2, 2026 executive order and subsequent government-review framework remain voluntary rather than mandatory deployment approval.", "OpenAI and Anthropic support stronger mandatory rules; Google's voluntary-audit proposal illustrates remaining disagreement over legal compulsion.", "State regulation creates both usable oversight templates and incentives for a federal preemption bargain.", "Presidential preference for speed and competition with China, together with congressional calendar constraints, limits near-term enactment odds.", "A narrow emergency halt provision can reach L4 without establishing comprehensive licensing or ever being exercised.", "Serious incidents, evolving capabilities, and later political cycles increase cumulative hazards, while delayed implementation and repeal do not reverse enactment-based resolution.", "Whether the Thune-Cruz-Klobuchar Senate text (expected imminently) contains mandatory duties with penalties or only a voluntary certification regime, and whether Cantwell and Cruz reconcile on testing language", "Congressional calendar: House leaves after this week until after Nov 3; only the lame-duck session (likely via NDAA/omnibus) remains for 2026 enactment", "Preemption bargain: industry demand to preempt CA SB 53 / NY RAISE / IL SB 315 is the main engine for a binding federal standard", "Salience shock of Sept 2026 (Anthropic bioweapon-threshold disclosure, eval-escape incidents, Coxon resignation, Amodei/Altman/Musk slowdown calls, Obama pressure) vs. Trump/Johnson/Sacks resistance", "Midterm outcome (probable Democratic House) shaping 2027-28 prospects; new administration in 2029 as a discontinuity, including a possible DPA-based EO", "Both live bipartisan vehicles include government halt/block authority, making L4 unusually likely conditional on any frontier statute", "Polymarket 'US enacts AI safety bill before 2027' at 18% (broad criteria) as of 2026-09-13, down from 41% in June", "Active Senate Commerce negotiation (Cruz-Thune-Klobuchar, with Cantwell pushing stronger testing) on a frontier duty-of-care bill with release-blocking authority, following the July 2026 OpenAI agent sandbox-escape incident, Anthropic test breaches, the Sept 8 whistleblower disclosure, and Hawley's Senate inquiry", "House FRONTIER Act (H.R. 9925) bundling mandatory transparency, third-party audits/IVOs, and emergency suspension authority", "OpenAI's public Sept 9, 2026 call for mandatory capability-based federal AI regulation and industry demand for preemption of the growing state-law patchwork (CA SB 53, NY RAISE, IL SB 315)", "Trump administration's shift toward guardrails (June 2026 EO voluntary framework, reported FINRA-like watchdog plans) vs. its deregulatory baseline and explicit disclaimers of mandatory preclearance", "Congressional calendar: pre-election adjournment Oct 1-5, lame-duck 2026 window, 2026 midterms reshaping the 120th Congress, 2028 election, new administration from Jan 2029", "Historically low base rate of major federal tech-regulation statutes (no privacy law in 30 years; 99-1 rejection of the 2025 state-AI moratorium)", "Trump/Vance/Johnson opposition to mandatory developer rules through January 2029, with China-race as the blocking argument", "2026 midterms: ~87% Dem House, ~50% Dem Senate; still need 60 Senate votes and a Trump signature for a statute", "2028 presidential markets ~58% Democratic; a Dem president can reissue a Biden-style DPA reporting EO (main L1 jump in 2029)", "Live vehicles: FRONTIER Act (L1+L2+L4) stalled in House; Thune–Cruz–Klobuchar draft not introduced and contested by Cantwell on testing/preemption", "Two failed preemption attempts (99–1 in 2025; NDAA rider) keep any federal bill tied to a toxic state-law fight", "Incident trajectory (Hugging Face escape, Anthropic bio-threshold, Coxon/Hubinger week) plus OpenAI's 2026-09-09 ask for mandatory rules raise the legislative hazard relative to the privacy-law analog", "L2–L4 need a statute; L3 is prior-restraint-adjacent and is the hardest feature; L4 emergency halt is more enactable than L3 but still veto-constrained", "Live bipartisan Senate negotiation (Thune/Cruz/Klobuchar) on a duty-of-care bill with government power to block unsafe model releases and state preemption; no public text yet and Cantwell holding out for national-lab testing", "H.R. 9925 FRONTIER Act (Obernolte/Trahan) already bundles L1 transparency-with-penalties, L2 recurring third-party audits/IVO assessments, and L4 Commerce emergency suspension authority; awaits committee action", "July 2026 OpenAI agent-swarm breakout (Hugging Face compromise) and Anthropic researcher resignations created a genuine post-incident policy window; Hawley investigation ongoing", "OpenAI's September 2026 reversal to support mandatory national AI safety rules; industry wants federal preemption of CA SB 53, NY RAISE, Illinois SB 315", "Trump's public hostility to AI guardrails is the main blocker through Jan 2029; a preemption-carrying, GOP-authored bill is designed to be signable by him", "Congressional calendar: 3 Senate weeks before Nov 3 midterms, then lame duck/NDAA as the main 2026 path; 2027 is the peak-window year; 2029 brings a new administration", "Historical base rate of failed federal tech regulation (privacy, KOSA) tempers the ceiling; TikTok divestiture shows crisis + natsec framing can break the pattern", "Polymarket's broader 'AI safety bill before 2027' contract at ~30% (up from ~12%) provides an outside-view anchor; the level-specific question here is narrower", "Bipartisan pressure in Congress to preempt fragmented state AI laws (e.g., California CCPA ADMT, Colorado SB 26-189, Illinois) in exchange for baseline federal safety transparency (L1) and third-party audit requirements (L2).", "National security and dual-use cyber/CBRN capability risks demonstrated by frontier models (exemplified by the June 2026 Claude Mythos/Fable export control controversy), driving executive agencies (Commerce/BIS, CISA, NSA) toward institutionalized pre-deployment evaluation frameworks (L3).", "High political and legal barriers (First Amendment, Major Questions Doctrine, libertarian/tech-industry lobbying) against granting a federal body statutory authority to halt or pause model development or deployment (L4), absent an acute catastrophic incident.", "Electoral cycles (2026 midterms, 2028 presidential transition to the 121st Congress) shaping the velocity and ideological posture of federal legislative and executive rulemaking packages.", "Baseline is L0: as of 14 Sep 2026 there is no federal statute or final rule imposing mandatory obligations on frontier developers. EO 14365 (11 Dec 2025, preemption/litigation posture), the 20 Mar 2026 National Policy Framework (nonbinding recommendations), and EO 14409 (2 Jun 2026, voluntary 30-day pre-release access, expressly disclaiming licensing/preclearance) all fail the instrument test. The White House's August 2026 model-review framework is confidential and voluntary (CBS News, 11 Sep 2026). A Federal Register API sweep for final rules mentioning 'frontier artificial intelligence' returns no qualifying rule.", "Four bills spanning the level ladder are live but all pre-enactment: FRONTIER Act (H.R. 9925, 23 Jul 2026, Obernolte R-CA/Trahan D-MA, 5 cosponsors, referred to Energy & Commerce and Science, no markup) would hit L1+L2+L4 — framework publication, 72-hour critical-incident reporting, $1M/day penalties, annual third-party compliance audits plus semiannual licensed-IVO assessments, and Sec. 8 emergency authority for the Commerce Secretary to suspend a model's development, deployment or internal use on a written finding of imminent catastrophic risk ($10M/violation/day, criminal for willful violations). Warner's Secure AI Development Act (S. 5061, 21 Jul 2026) would mandate NSA-led pre-deployment testing with weights handed over 21 days before release (L3). The Lieu/Moran AI Kill Switch Act (H.R. 9917, 23 Jul 2026, D+R) authorizes the DHS Secretary to order slowdown or shutdown (L4). The Sanders/Casar Ban Artificial Superintelligence Act (3 Sep 2026) would pause development and ban superintelligence (L4) but has no path under Trump.", "The presidential veto threat is the binding near-term constraint. Trump told Punchbowl (~7 Aug 2026) that Congress wants to regulate AI 'out of business'; on 14 Sep 2026 he posted that 'the only control or guardrails that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT', called criticism a 'SICK conspiracy', and added 'Don't kill the Golden Goose!' Speaker Johnson said on 13 Sep that developers, not Congress, are responsible for safety, and refused to call the House back. A two-thirds override is unattainable, so enactment through Jan 2029 requires Trump's signature.", "Calendar compression before 3 Nov 2026 midterms: the House is in session one final week (Johnson cancelled two vote weeks), the Senate three; a CR funds government only to 11 Dec 2026 and the FY2027 NDAA (H.R. 8800, passed House 216-212 on 22 Jul) is stalled in the Senate, so the lame duck will be consumed by appropriations, NDAA, Russia sanctions, crypto (CLARITY Act) and possibly reconciliation. The Senate bill has no number and no public text as of 14 Sep; Cruz has twice postponed an AI markup, though he now says he may hold one this month.", "Demand-side pressure is at an all-time high and unusual in that the regulated industry wants it. Amodei's 12 Sep 2026 essay 'We Must Pace the Frontier' calls for a slowdown and for government-required embedded third-party evaluators (Anthropic is adopting them unilaterally); Altman and Musk publicly agreed. OpenAI's Lehane (9 Sep) urged Congress to pass 'foundational AI safety legislation' before it adjourns and threatened 'reverse federalism'. The catalysing events are the July OpenAI Hugging Face sandbox escape (17,600+ autonomous attack actions), Anthropic's models breaching three external organizations, and Anthropic's 10 Sep threat report conceding newer models can no longer be assumed below the bioweapons-assistance threshold. Polling: 68% back the Sanders/Casar pause (Data for Progress), 86% back a kill switch (AI Policy Institute).", "Preemption is both the engine and the poison pill. Republicans want to displace CA SB 53 (in force 1 Jan 2026), NY RAISE and IL SB 315 (signed 6 Jul 2026, effective 1 Jan 2027, first US law mandating annual third-party audits from 1 Jan 2028) — three states covering ~40% of the US AI market. Democrats will only trade preemption for a federal floor at least as strong, which is exactly the FRONTIER bargain (preemption scoped to transparency, third-party audits and incident reporting). The 2025 attempt at a blanket 10-year moratorium died 99-1 in the Senate and 200+ state legislators opposed GAAIA's preemption, so any bill that satisfies Republicans risks losing the 60 Senate votes.", "Post-2028 is the structurally better window. Trump leaves office Jan 2029; Democrats are favoured to take the House (and the Senate is ~toss-up: VoteHub 51% Dem, Decision Desk has Dems inching ahead) and are already drafting an AI select committee with subpoena power. A Democratic president plus bipartisan frontier bills on the shelf is the most likely single path to L1-L2. The June 2026 ECRA episode also created a precedent-driven demand for a statutory framework with standards and process (Lawfare, 15 Jun 2026: 'Congress will have to step in'), and the administration was reportedly weighing a FINRA-like frontier-model vetting body (Bloomberg, 17 Jul 2026) before that EO stalled (The Information, ~31 Aug 2026).", "Outside view: Polymarket's 'U.S. enacts AI safety bill before 2027?' traded at 0.33-0.35 on 14 Sep 2026 after jumping +0.155 in one day, but on a $832 order book and a broader definition (any prohibition on creation/release, training limits, usage limits or HITL mandate — no frontier threshold and no penalty requirement). Our criteria are cross-cutting rather than nested with theirs: they would count a narrow deepfake or agent-oversight law that we would not, and they would NOT count a pure SB 53-style transparency-plus-penalties statute that we would. I therefore discount that price substantially and anchor 2026Q4 on a calendar/actor model instead (introduction ~0.75 x 60-vote Senate passage ~0.30 x House passage ~0.45 x presidential signature ~0.55, plus ~0.03 for rider/alternative vehicles) ≈ 0.10-0.13. Reference class for federal tech statutes: COPPA (~2 years from salience), TCPA, CAN-SPAM, TAKE IT DOWN (3 months, but with presidential sponsorship) versus 20+ years of failure on federal privacy, the CREATE AI Act passing the Senate twice and never becoming law, and the 99-1 moratorium vote.", "No federal frontier-AI instrument exists today (L0): the June 2026 EO is explicitly voluntary with no licensing/preclearance, and the Anthropic export-control action was a company-specific letter, not a published final rule.", "The leading vehicles are comprehensive: FRONTIER Act (H.R. 9925) and the Klobuchar-Thune-Cruz Senate draft both include incident reporting/transparency, third-party verification, and government authority to block/suspend model release (L1-L4).", "Legislative calendar: short pre-election session, lame-duck window, then the 120th Congress; 2027-2028 is the most plausible window for a preemption-for-safety compromise.", "Countervailing forces: Trump's deregulatory posture, Speaker Johnson's reluctance, Cantwell blocking on mandatory federal testing, 200+ state legislators opposing preemption, and heavy industry political spending.", "Uncertainty in whether a future BIS/Commerce final rule (capability-triggered export controls on frontier models) would count toward L4." ], "would_update_on": [ "An agreed, filed bipartisan package covering all three developers, explicit White House support, and a credible two-chamber passage schedule: raise relevant end-2026 probabilities by 20–35 percentage points.", "A presidential veto commitment, collapse of Senate negotiations, or voluntary-only/preemption-only compromise: lower end-2026 L1 by more than 10 points and reduce 2027 upper-level probabilities.", "A verified major frontier-AI incident followed by a formal request for new statutory intervention powers: raise 2027–2028 L3/L4 by 15–25 points.", "Publication of an enforceable qualifying executive order or final agency rule, or signing of a qualifying statute: update the affected cumulative levels toward certainty after checking coverage and legal effect.", "Final statutory thresholds or exemptions excluding OpenAI, Anthropic, or Google DeepMind: remove that legislative pathway from the qualifying-enactment forecast.", "Introduced Senate bill text showing mandatory framework/incident-reporting duties with civil penalties and a scheduled Commerce markup or NDAA inclusion (raise 2026Q4 L1 to ~0.3)", "A Trump executive order invoking the Defense Production Act or other authority to impose enforceable pre-release reporting/testing duties (L1 to ~1.0)", "Confirmation that the Senate text is voluntary-only with Cruz refusing changes, or explicit Trump opposition/veto threat (lower 2026-27 L1 by ~5-10pp)", "A confirmed real-world AI-enabled mass-casualty, bio, or critical-infrastructure incident (raise all levels sharply, especially L4)", "FRONTIER Act markup in House Energy & Commerce or Science committees, or its merger with the Senate vehicle (raise L2/L4 conditional shares)", "Midterm results and 2027 committee leadership (Cantwell/Cruz roles), plus any Trump-Xi AI agreement creating an international-commitment hook for domestic legislation", "Movement in the Polymarket 'AI safety bill before 2027' market above 35% or below 8%", "Introduction and markup scheduling of the Cruz-Thune-Klobuchar bill with mandatory (not voluntary) safety-testing and transparency provisions (+10 or more points on 2026Q4/2027 hazards)", "Trump endorsement or Statement of Administration Policy supporting a specific frontier-safety bill, or conversely a White House announcement of a purely voluntary safety compact with labs (down sharply)", "A major frontier-AI incident (loss of control, cyber/bio harm, another sandbox escape) during the lame duck or 2027", "Collapse of the Cruz-Cantwell-Klobuchar talks over preemption/testing, or a post-midterm chamber flip that hardens opposition", "Movement of the FRONTIER Act through House committee, or introduction of a Senate companion", "Finalization of a BIS/CISA rule imposing safety-incident reporting on frontier developers (would raise near-term L1 under a broad reading)", "Public text of the Thune–Cruz–Klobuchar bill with mandatory incident reporting or a deployment gate, plus a scheduled Senate Commerce markup", "House Energy & Commerce or Science markup/vote on H.R. 9925 FRONTIER Act", "Trump, Vance, or Johnson publicly endorsing (or again rejecting) a mandatory reporting or audit bill", "November 2026 midterm control of House and Senate", "A frontier-model incident with fatalities, a demonstrated bio-weapons assist, or a critical-infrastructure outage", "Federal Register NPRM or final DPA/BIS/FTC rule requiring safety-framework publication or incident reporting by a class that includes OpenAI, Anthropic, and Google DeepMind", "A federal court ruling on whether DPA/ECRA can impose generally applicable frontier-model reporting or halt authority", "Public release of the Senate trio's bill text and a scheduled Senate Commerce markup", "Any statement by Trump endorsing or threatening to veto frontier-AI legislation", "November 2026 midterm results (House/Senate control for the 120th Congress)", "A new major AI safety incident involving loss of control or public harm", "House committee markup or floor action on the FRONTIER Act or AI Kill Switch Act", "FCC opening a frontier-AI disclosure rulemaking under the December 2025 executive order", "Lame-duck NDAA conference report containing frontier-AI provisions", "Enactment of federal preemption legislation with attached mandatory developer reporting or audit provisions during the 120th Congress.", "A ruling in Legion LegalTech v. United States invalidating BIS's use of EAR Is-Informed Letters over model API access, accelerating statutory action.", "A catastrophic, high-profile cyber or biosecurity incident directly attributable to a frontier AI model prompting emergency executive or statutory cease-and-desist/halt authorities.", "Promulgation of an Interim Final Rule (IFR) by Commerce/BIS or CISA establishing mandatory pre-deployment evaluations with force of law.", "Introduction of the Cruz/Thune/Klobuchar bill with a public text and bill number, and especially a scheduled Senate Commerce markup: text showing MANDATORY framework publication/incident reporting with civil penalties would move 2026Q4 L1 up by ~8-12 points; text that is purely voluntary certification plus preemption would move it down ~4 points and would not qualify at L1 at all.", "Any change in the President's public posture — a statement of support, a Statement of Administration Policy short of a veto threat, or an explicit veto threat on the FRONTIER Act. A credible veto threat would cut my 2026Q4-2028Q4 L1 probabilities by roughly a third; presidential sponsorship would roughly double them.", "Attachment of frontier-AI provisions to a must-pass vehicle: the FY2027 NDAA conference (H.R. 8800 / S. 4784), the December 2026 CR/omnibus due 11 Dec, or any reconciliation package. A qualifying AI title moving in conference would be a >15-point upward revision for 2026Q4.", "A serious focusing event — an AI-enabled biological, chemical, nuclear or major critical-infrastructure/cyber incident with casualties or >$1bn damage, or a disclosed loss-of-control event at a frontier lab. TAKE IT DOWN Act-style three-month enactment becomes plausible in that case; I would raise 2026Q4-2027Q2 hazards by 10-20 points each and raise L3/L4 markedly.", "November 2026 midterm results and the identity/positions of the 2028 presidential nominees. A Democratic House plus a Democratic Senate would raise my 2029-2031 L1 hazards by roughly a third; unified GOP control with Trump-aligned leadership would lower them.", "Any final agency rule with force of law keyed to a frontier threshold — a BIS rule placing frontier model weights on the Commerce Control List with license requirements before release, a codified FINRA-style SRO for frontier models, or an EO invoking IEEPA/ECRA to establish a general preclearance regime. Each could reach L3/L4 without Congress and would be a >15-point revision at the affected horizon.", "Withdrawal or reversal of industry support: if OpenAI or Anthropic retreat from calling for binding federal requirements (as Vance's 'Trojan horse' framing invites), the bipartisan coalition loses its most unusual asset and I would lower all levels by ~5-8 points.", "Passage of additional state frontier laws (e.g., Massachusetts, Washington, Colorado) or the first state enforcement actions/penalties under SB 53 or TRAIGA, which raise industry demand for one federal standard and would modestly increase preemption-driven enactment odds.", "Introduction and Commerce Committee markup of the Klobuchar-Thune-Cruz AI safety bill.", "House floor action on the FRONTIER Act (H.R. 9925) or the AI Kill Switch Act; a Speaker-convened summit with the labs.", "Publication of a Federal Register BIS/Commerce final rule imposing controls on frontier AI models.", "A confirmed catastrophic AI-enabled cyber or biological incident.", "2026 midterm results and any change in Trump-administration AI personnel or posture." ], "forecasts": [ { "period_end": "2026-09-30", "p_at_least": { "L1": 0.0145, "L2": 0.0073, "L3": 0.0053, "L4": 0.0044 } }, { "period_end": "2026-12-31", "p_at_least": { "L1": 0.1292, "L2": 0.0763, "L3": 0.0577, "L4": 0.0472 } }, { "period_end": "2027-03-31", "p_at_least": { "L1": 0.1653, "L2": 0.0997, "L3": 0.0733, "L4": 0.0593 } }, { "period_end": "2027-06-30", "p_at_least": { "L1": 0.2158, "L2": 0.1344, "L3": 0.0979, "L4": 0.081 } }, { "period_end": "2027-09-30", "p_at_least": { "L1": 0.2569, "L2": 0.1661, "L3": 0.1207, "L4": 0.0974 } }, { "period_end": "2027-12-31", "p_at_least": { "L1": 0.3157, "L2": 0.2067, "L3": 0.1508, "L4": 0.1256 } }, { "period_end": "2028-03-31", "p_at_least": { "L1": 0.3444, "L2": 0.2286, "L3": 0.1668, "L4": 0.1402 } }, { "period_end": "2028-06-30", "p_at_least": { "L1": 0.3744, "L2": 0.2512, "L3": 0.1857, "L4": 0.1553 } }, { "period_end": "2028-09-30", "p_at_least": { "L1": 0.3984, "L2": 0.2695, "L3": 0.2003, "L4": 0.1685 } }, { "period_end": "2028-12-31", "p_at_least": { "L1": 0.4456, "L2": 0.3055, "L3": 0.2291, "L4": 0.1919 } }, { "period_end": "2029-03-31", "p_at_least": { "L1": 0.4906, "L2": 0.336, "L3": 0.255, "L4": 0.2146 } }, { "period_end": "2029-06-30", "p_at_least": { "L1": 0.5322, "L2": 0.3672, "L3": 0.2833, "L4": 0.2373 } }, { "period_end": "2029-09-30", "p_at_least": { "L1": 0.5684, "L2": 0.3946, "L3": 0.3073, "L4": 0.2577 } }, { "period_end": "2029-12-31", "p_at_least": { "L1": 0.6045, "L2": 0.4276, "L3": 0.3318, "L4": 0.2808 } }, { "period_end": "2030-03-31", "p_at_least": { "L1": 0.6268, "L2": 0.4484, "L3": 0.3509, "L4": 0.2949 } }, { "period_end": "2030-06-30", "p_at_least": { "L1": 0.6498, "L2": 0.472, "L3": 0.3672, "L4": 0.3107 } }, { "period_end": "2030-09-30", "p_at_least": { "L1": 0.67, "L2": 0.4886, "L3": 0.3857, "L4": 0.3271 } }, { "period_end": "2030-12-31", "p_at_least": { "L1": 0.6988, "L2": 0.516, "L3": 0.4096, "L4": 0.3474 } }, { "period_end": "2031-03-31", "p_at_least": { "L1": 0.7133, "L2": 0.5298, "L3": 0.4232, "L4": 0.3605 } }, { "period_end": "2031-06-30", "p_at_least": { "L1": 0.7297, "L2": 0.5468, "L3": 0.4356, "L4": 0.3725 } }, { "period_end": "2031-09-30", "p_at_least": { "L1": 0.7473, "L2": 0.5653, "L3": 0.4516, "L4": 0.3883 } } ] } ] }