⚡ Quick Summary
Published by the Future of Life Institute, this model bill is drafted for enactment by the Legislature of the State of [STATE]. Its stated purposes are to designate frontier AI models as products under state product liability law, to designate frontier AI developers as manufacturers except where a third party has substantially altered the model, to authorise joint and several liability across a model's value chain, and to specify factors informing whether a frontier AI model is defective.
The bill runs to ten sections. It defines artificial intelligence model, deployer, distributor, frontier AI developer, frontier AI model, manufacturer, person, substantial alteration and value chain; designates frontier AI models as products and their harmful outputs as the product; and makes the frontier AI developer the manufacturer, subject to a substantial-alteration exception the developer must prove by a preponderance of the evidence.
Liability is joint and several across developers, distributors and deployers who are a proximate cause of harm, with apportionment based on causal contribution, access to information about model behaviour and temporal proximity. Defectiveness is judged under existing doctrines of manufacturing defect, design defect and failure to warn, supplemented by factors covering control mechanisms, safety and security protocols, and pre-deployment testing, plus a six-year record-preservation duty carrying an adverse inference.
🧩 What’s Covered
- Findings and purpose (Section 2): the legislature finds that a small number of developers control the most powerful models, that existing product liability frameworks produce inconsistent results for software and AI, that developers hold exclusive access to design choices, training data and evaluation results, and that B2B contracts shift liability to downstream parties with limited ability to test or safeguard; the stated purposes follow from these findings.
- Definitions (Section 3): artificial intelligence model, deployer, distributor, frontier AI developer (a training run exceeding [10^26] floating-point operations, or developing, maintaining or controlling a frontier AI model), frontier AI model (cross-domain capability in at least two domains of human cognitive activity, plus capability to cause death, substantial physical injury or illness, severe psychological harm or significant disruption of critical infrastructure), manufacturer, person, substantial alteration and value chain.
- Product designation (Section 4): frontier AI models are products; intangibility, capacity to learn or adapt, delivery as a service or provision through a software interface do not exclude them; harmful outputs are treated as the product regardless of foreseeability; a model is not a legal person and autonomy or unexpected behaviour is not a defence.
- Manufacturer designation (Section 5): the frontier AI developer is the manufacturer, except where a third party has substantially altered the model and that alteration is the proximate cause; the burden lies with the developer; multiple developers, including base-model developers and those releasing weights, are each manufacturers.
- Joint and several liability (Section 6): developers, distributors and deployers found to be a proximate cause are jointly and severally liable, with apportionment factors and contribution rights; indemnification clauses and compliance with regulation or voluntary standards do not defeat liability.
- Defectiveness factors (Section 7): control mechanisms (per-action approval, real-time monitoring, design-phase specification), safety and security protocols (risk management, information security, access controls, misuse safeguards, monitoring and incident response, documentation and disclosure), pre-deployment testing (capability evaluations, adversarial testing and red-teaming, domain-specific safety testing, evaluation of control mechanisms under realistic conditions, documentation and retention, re-testing), failure to warn, and six-year record preservation with adverse inference.
- Nexus and preservation of other law (Sections 8–9): harm to residents, deployment or availability in the state, or regular business conduct including more than $25,000,000 in annual revenue from AI products or services to persons in the state; the Act supplements other claims and does not preempt state agencies; forum and governing-law waivers are void.
- Effective date and severability (Section 10): the Act takes effect 90 days after enactment and applies to harm occurring on or after that date, regardless of when the model was developed.
💡 Why it matters?
Product liability is one of the few routes by which people harmed by AI systems can seek redress, and the bill's central move — treating a frontier model as a product and its developer as the manufacturer — determines who can be sued and for what. For developers and their advisers, the defectiveness factors read as a de facto checklist: control mechanisms that are practicable under realistic operational conditions, security measures protecting weights and training data, misuse safeguards robust to jailbreaking and fine-tuning, post-deployment monitoring, and pre-deployment testing including independent red-teaming. Voiding indemnification clauses and regulation-compliance defences changes the risk calculus in contracts across the value chain.
❓ What’s Missing
No publication year or version is printed, and the text is a model bill with placeholders — [STATE] and [10^26] floating-point operations — left for an adopting legislature to fill. The term 'harm' is used throughout but not defined. The bill does not say who enforces it, create any regulator or reporting duty, address insurance, or state its relationship to federal law beyond preserving state agency authority. How 'state of the art' is to be established is left to courts, and no drafting notes or commentary explain the thresholds or definitions chosen.
👥 Best For
State legislators, legislative counsel and policy staff drafting or amending AI liability statutes, and legal, compliance and risk teams at frontier model developers, distributors and deployers assessing exposure, contract terms and indemnity positions. Also relevant to insurers and practitioners mapping how defectiveness, control mechanisms and record-keeping duties may be argued.
📄 Source Details
Frontier AI Product Liability Act: A Model Bill, published by the Future of Life Institute and attributed to Alex Tsalidis, with the contact address alexandra@futureoflife.org. No publication year, version or edition is printed. The document is ten pages and in English; Section 1 gives the short title as 'Frontier AI Products Liability Act.' A printed line reads 'View this document online: futureoflife.org/product-liability'. The input was a text extraction of all ten pages; no page images were reviewed.
