⚡ Quick Summary
Published by the Future of Life Institute, this model bill—the AI Responsibility for Catastrophic Harm Act, or ARCH Act—would impose strict liability on frontier AI developers for catastrophic harms caused by frontier AI models, designating the development, training and release of such models an abnormally dangerous activity. Its findings state that frontier model development carries a high degree of risk of grave harm that reasonable care cannot eliminate, that AI models cannot bear legal responsibility, and that developers use business-to-business contracts to shift liability to downstream providers.
The core mechanism is strict liability attaching "without regard to fault, negligence, recklessness, or intent", applied jointly and severally to every developer in the chain of development, with contribution based on proportionate causal contribution. Catastrophic harm is defined by thresholds: death of one hundred or more persons, substantial injury or illness to five hundred or more, significant disruption of critical infrastructure, or property damage exceeding $1 billion over twelve months. A frontier AI developer is defined partly by training runs exceeding [10²⁶] floating-point operations.
The bill exempts unaffiliated downstream parties, creates a private right of action and Attorney General enforcement with civil penalties of up to $5 million per violation, voids arbitration, liability-limitation and indemnification clauses, and sets a six-year limitations period.
🧩 What’s Covered
The bill is set out in nine numbered sections.
- Short title and enactment formula (§1): names the Act and opens with the standard "Be it enacted by the Legislature of the State of [STATE]" formula, marking it as template legislation for state introduction.
- Findings and purpose (§2): six findings on the ultra-hazardous characteristics of frontier model development, the inability of AI models to bear responsibility, contractual risk-shifting, and the state's police powers over health, safety and welfare; three stated purposes, including a remedy without proving negligence, defect or absence of human control.
- Definitions (§3): defines AI model, catastrophic harm with its four thresholds, the chain of development (foundation-model developers, capability-enhancing fine-tuners, and releasers of model weights), frontier AI developer by compute threshold and day-to-day managerial control, and frontier AI model by capability across at least two domains of human cognitive activity.
- No responsibility for models (§4): states that an AI model is not a legal person and bars defences resting on autonomous, unexpected or apparently intelligent behaviour.
- Strict liability (§5): declares the activity ultra-hazardous, sets the fault-free standard, provides joint and several liability with apportionment factors, extends liability for open-weight releases to reasonably foreseeable harm within five years, and defines the nexus to the state.
- Exemptions (§6): exempts deployers, integrators, distributors, infrastructure providers, end users, tooling providers and accredited higher-education institutions that meet independence conditions, while preserving their liability under other law.
- Enforcement (§7): private actions and class actions, Attorney General suits with penalties capped at $5 million per violation except for critical-infrastructure harm, non-waiver and anti-indemnification rules, and a six-year discovery-based limitations period.
- Preservation of other law and final provisions (§8–9): preserves other claims and state agency authority, sets a [90-day] effective date and includes a severability clause.
💡 Why it matters?
For developers, deployers and their counsel, the bill is a concrete drafting proposal for shifting catastrophic-risk exposure onto model developers rather than the downstream organisations that host or integrate models. It supplies the operative thresholds—how many deaths or injuries, what value of property damage, which compute level—that a legal, risk or compliance team would need in order to test exposure, insurance cover and contract terms. It also forces review of standard contractual protections, since arbitration, liability caps and indemnities covering frontier developers would be void in the enacting state's courts.
❓ What’s Missing
The text carries unfilled bracketed parameters—[STATE], [10²⁶] floating-point operations and [90 days] before taking effect—so key thresholds and scope are left to the adopting legislature. The definitions are numbered (1), (2), (3), (5) and (6), with no (4), and the definition of a frontier AI model closes by referring to the undefined "high-capability AI model", so the drafting is not internally consistent. There is no testing or disclosure mechanism, no insurance or financial-responsibility requirement, no evidentiary presumption, and no discussion of interaction with federal law, other states' statutes or non-US regimes. No date, version or revision history is printed.
👥 Best For
State legislators and their staff drafting or comparing AI liability proposals; legal, policy and risk teams tracking how strict-liability arguments are being framed for frontier models; and compliance or insurance advisers assessing where catastrophic-harm exposure would sit along the model supply chain under a fault-free standard.
📄 Source Details
The document is AI Responsibility for Catastrophic Harm Act ("ARCH Act"), subtitled A Model Bill, published by the Future of Life Institute and credited to Alex Tsalidis (alexandra@futureoflife.org). No publication date, version or edition is printed. The extraction covers all seven pages of the file, comprising the enactment formula and Sections 1 to 9. The document prints "View this document online: futureoflife.org/absolute-liability"; no http(s) URL is printed.
