The US or EU makes developers strictly liable for model harms

Where things stand (as of 17 Oct 2026)

No qualifying law has been enacted since the question opened on 16 Sep 2026, and nothing that would qualify is close to final passage in either jurisdiction.

United States

• AI LEAD Act (S.2937, Durbin–Hawley). I read the GovInfo text. §101(a)(4) creates a true no-fault theory: the developer is liable if the product was "in a defective condition unreasonably dangerous." The bill also bars Section 230 as a defence. If enacted it would qualify. But it has sat in Senate Judiciary since Sep 2025 without a committee vote (CCDH, 28 Aug 2026), and Durbin is retiring.
• CLAIM Act (Rep. Trahan, discussion draft, 5 Oct 2026). It makes developers liable "regardless of the degree of care exercised" when an AI's conduct would be tortious if a human did it, so it would qualify. It is an unintroduced draft from a minority-party member.
• The bills with real momentum are fault-based. The Thune–Cruz–Klobuchar draft imposes a "duty of care," i.e. negligence (Tech Times/Reuters, Sep 2026). The Hawley–Murphy AI Agent Accountability Act requires a failure to adopt "reasonable safeguards" plus knowledge. The FRONTIER Act (H.R. 9925) has no private liability. None of these would qualify.
• Republicans have blocked fast-tracked AI safety bills. House Energy and Commerce chair Guthrie says a major bill waits until 2027.
• The White House leans the other way. Its March 2026 legislative framework asks Congress to stop states from penalising developers for third parties' misuse (Lawfare, 10 Apr 2026). Trump called AI safety fears a "hoax" and relies on a voluntary industry accord.
• Markets and forecasters see little near-term chance. Polymarket put "US enacts AI safety bill before 2027" at about 5%. Metaculus had about 5.5% for a federal AI-hacking liability law of any fault standard before Oct 2027; its professional forecasters said 2–3%.

European Union

• No proposal exists. The fault-based AI Liability Directive was withdrawn on 6 Oct 2025 (C/2025/5423). The 2024 Product Liability Directive is part of the baseline.
• Parliament is pushing. Four MEPs (Voss, Benifei, van Sparrentak, McNamara) sent a note on 16 Sep 2026 asking for an "AI Liability Act" that reverses the burden of proving a defect and covers harm during testing. Renew has signalled support. Parliament's 2020 resolution and a 2025 Parliament study both favoured strict liability for high-risk AI. Sam Altman said on 5 Oct that "there's gonna need to be a liability framework."
• The Commission is resisting. Its spokesperson called the current rules "fully fit for purpose" (AFP, 6 Oct 2026). The 2027 work programme is due on 20 Oct and no leak mentions AI liability.
• Timing. EU co-decision takes about 18–30 months, and the 2029 elections fall in the middle of any likely timetable. Realistic adoption is 2029–2031 at the earliest.

Reference class and base rate

Congress has created strict-liability regimes mainly after disasters: oil pipelines in 1973, CERCLA in 1980, the Oil Pollution Act in 1990. For new technologies it has more often passed shields: Section 230, the gun-industry immunity law (PLCAA), the 1994 aviation statute of repose (GARA), and the PREP Act. Comprehensive technology bills on privacy, antitrust and kids' safety have failed for a decade, even under Democratic trifectas. In the EU, the Commission chose a fault-based design in 2022 and then withdrew even that. On base rates alone I would put five-year odds at about 5%. I raise that because this is an unusually intense moment for AI risk:

• labs' agents have escaped test environments and hacked third parties;
• AI company chiefs have called for slowing development;
• a bipartisan populist coalition favours liability;
• a Democratic wave is likely in November 2026.

Paths and decomposition

• US, before 2029. A qualifying law needs Trump's signature against his own framework. I put it at about 2.5% cumulative.
• US, 2029–2031. I assume roughly a 33% chance of a Democratic trifecta, with about 13% conditional odds of no-fault liability being enacted; otherwise about 2.5%. That adds roughly 6 points, for about 8% in the US in total.
• EU. About a 30% chance the Commission proposes by 2029–30, roughly 50% that it is adopted and published by end-2031, and about 45% that it contains no-fault developer liability rather than presumptions alone. That gives about 5%.
• Combined. About 13% by end-2031. I built quarterly hazards (US 0.1–0.7% a quarter, EU 0 until 2028 and then 0.2–0.5%) and multiplied them through in Python to get the series.

Strongest case against my estimate

• For higher: a large AI-caused disaster within five years seems plausible in this environment, and disasters have historically produced strict liability within about 18 months (the Oil Pollution Act after Exxon Valdez). The question's definition is also broad: strict product liability counts even with statutory defences, and removing an immunity such as Section 230 counts.
• For lower: in practice, liability language is "the first thing stripped" in amendments, the leading vehicles in both jurisdictions are fault-based or evidence-only, and Washington is putting trade pressure on EU tech rules.

I weigh these as roughly balanced around 13%.

Sources: GovInfo S.2937 text; Trahan CLAIM Act draft PDF (trahan.house.gov, 5 Oct 2026); CFO Dive (Oct 2026); Roll Call, 1 Oct 2026; Hawley press release; Time, 25 Sep 2026; Tech Policy Press September roundup (1 Oct 2026); Transformer (Oct 2026); Broadband Breakfast, 15 Oct 2026; Lawfare, 10 Apr 2026; Metaculus substack (Sep 2026); Polymarket posts (Sep–Oct 2026); Politico.eu, 21 Sep and 9 Oct 2026; ad-hoc-news, 22 Sep 2026; France24/AFP, 6 Oct 2026; Bruegel (Sep 2026); EP Legislative Train (AI Liability Directive); Austrian Parliament document on the 2027 work programme date.

Question. By the last day of each quarter, has the US Congress or the EU enacted a new law creating or substantively expanding developers’ no-fault civil liability for harm caused by frontier AI models beyond the law in place when this question opened?

Definitions.

• Enacted law: a US federal statute that has become law through presidential signature, veto override, or the constitutional process without signature; or an EU legislative regulation or directive finally adopted and published in the Official Journal. A later commencement date or transposition deadline does not prevent qualification.
• Developer: an entity that develops or commissions development of a model and supplies it under its name, including an entity legally treated as the model’s manufacturer or producer. This question is not limited to covered developers. Liability imposed solely on users, deployers, distributors, or hosting services does not count.
• Frontier AI models: models within an enacted law’s expressly defined frontier/advanced-model category, or general-purpose AI models classified as having systemic risk under Article 51 of Regulation (EU) 2024/1689. A law covering AI, software, or products generally counts if its operative scope includes such models and their developers; it need not use the term “frontier.”
• No-fault civil liability: an obligation to compensate a harmed person without requiring negligence, recklessness, intent, or another culpable breach by the developer. Strict product liability counts even if the claimant must establish a defect, damage, and causation, and even if statutory defenses exist. A rebuttable presumption of negligence does not count if reasonable care remains a complete defense.
• Substantively expanding: adding a no-fault cause of action; extending an existing one to additional models, developers, harmed persons, or categories of harm; removing a substantive defense or immunity; or increasing recoverable compensation by removing or raising a statutory cap. The qualifying change must apply to harm attributable to a frontier model’s operation, outputs, or actions, including operation through an integrated system. Purely procedural or evidentiary changes do not count.
• Opening baseline: law enacted before 2026-09-16, including provisions with deferred application dates. In particular, the baseline includes Directive (EU) 2024/2853 on liability for defective products, published 18 November 2024, including its treatment of software and AI and its provisions applicable to products placed on the market or put into service after 9 December 2026. Compare each jurisdiction with its own opening baseline, including pre-existing liability preserved by its law.

Resolves YES if a qualifying new enactment after the opening baseline contains at least one operative provision meeting all the definitions above. The enacted text must establish the expansion, rather than merely authorize a future decision about whether to establish it. A directive requiring Member States to establish qualifying liability counts on publication, without waiting for transposition. No actual lawsuit, injury, or damages award is required.

Resolves NO / Exclusions.

• Bills, proposals, committee reports, passage by only one chamber, political agreements, and nonbinding resolutions do not count. The introduction of the AI LEAD Act, S.2937, 119th Congress, is not enactment.
• State laws, Member State national laws, judicial decisions, executive orders, administrative rules, delegated acts, and implementing acts do not count as enactments by Congress or the EU legislature.
• The commencement, transposition, enforcement, or republication of an opening-baseline law does not count. Mere codification of existing liability without expansion does not count.
• Fault-based liability, disclosure duties, presumptions of causation alone, collective-action procedures, regulatory fines, criminal penalties, insurance mandates, and compensation funds without developer civil liability do not count.
• A general study or direction to consider no-fault liability does not count. Neither does liability solely for conduct unrelated to model-caused harm.
• A quarter resolves NO if no qualifying enactment has been documented by its end.

Resolving source. US enactment status and operative text: Congress.gov’s bill Actions and Public Law records, with the authenticated public-law text on GovInfo controlling textual discrepancies. EU enactment and text: the Official Journal of the European Union, available through EUR-Lex. These official legislative sources, rather than the broader Public confirmation convention A2, control resolution. Apply the Resolution window and Resolution date convention A3: use the first qualifying official publication documenting completed enactment and the operative provisions, not a vote, commencement, or injury date. Once YES, later repeal or invalidation does not reverse it.