# Model Act — Frontier AI Public Welfare Offenses **Website version:** [frontieraiaccountabilityproject.github.io/model-act](https://frontieraiaccountabilityproject.github.io/model-act/) — styled for reading, with search **GitHub repository version:** [github.com/FrontierAIAccountabilityProject/model-act](https://github.com/FrontierAIAccountabilityProject/model-act) — the versioned source, and the authoritative record --- **A model state statute that makes the natural person who held the authority to stop a frontier AI deployment personally criminally liable for shipping it without care.** Not the company — a corporate fine is absorbed as a cost. The person who decided. **[Statutory text — v3.4.2 (PDF, 46pp)](./act/Model-Act-v3.4.2-statutory-text.pdf)** — SEC. 0 to SEC. 13, clean, nothing else. **[Annotated — v3.4.2 (PDF, 82pp)](./act/Model-Act-v3.4.2-annotated.pdf)** — the same text with drafting notes at each provision, sixteen objections and the answers to them, the open questions, and the source for every finding. Everything else on this site is the drafting record: the audits, the errata, and the register of what is still open. Public domain — anyone may take it, change it, or build on it, without asking. --- **Where it stands.** The current text is **v3.4.2 of 27 August 2026**. It supersedes v3.4, and statements elsewhere on this site about v3.4 should be read as superseded. An in-house adversarial review of v3.4 found that SEC. 6(a) — the individual-liability offense the Act exists to create — could not be pleaded as drafted. **That defect is repaired:** SEC. 6(a) now carries a predicate violation, a nexus between the failure of care and the violation, and an antecedent for the risk it concerns. **Six questions remain unresolved**, each marked in the text at the provision it concerns, with the answer this drafter would give — and what would defeat it — set out at the back of the annotated version. **Nothing has been reviewed by a lawyer.** No outside reviewer has signed anything. This text claims no review, and will not until someone named signs one. **Everything wrong with it is published**, in [the Worklist](./revision/worklist.md) for what is still open and [the Record](./corrections/corrections.md) for the mistakes already found and corrected. Pages elsewhere on this site were written against v3.4 and have not all been brought forward; where one of them differs from the statutory text above, **the statutory text is right**. --- ## What it makes a crime Five things. A person answers for each. | | Offense | Already a crime for an ordinary person under | |---|---|---| | 1 | Shipping without validation | 21 U.S.C. §§ 331, 333(a)(1) — strict liability | | 2 | Operating uncontrolled autonomous access that causes a breach | 18 U.S.C. § 1030 | | 3 | Failing to report | 18 U.S.C. § 4 — misprision, 3 years | | 4 | Lying to the State | 18 U.S.C. § 1001 — 5 years | | 5 | Destroying or withholding records | 18 U.S.C. § 1519 — 20 years | **Read the right column first.** Every one of these is already a crime in the United States for an ordinary person, most carrying heavier maximum sentences than this Act proposes. What the Act extends is not the criminal law. It is its reach. The doctrine is not new either: *United States v. Dotterweich*, 320 U.S. 277 (1943) and *United States v. Park*, 421 U.S. 658 (1975) have held corporate officers personally liable for public welfare offenses for eighty years. Duties attach to whoever holds final authority to prevent or halt, and cannot be delegated away. Technical work, access, advice, and carrying out someone else's decision create no personal liability. --- ## Who made this One person, no institution, no funding, no client. The drafting is AI-assisted and says so on every page; the instructions it runs under are published at [`CLAUDE.md`](./CLAUDE.md). **That is a reason to distrust it, so the project answers it the only way it can:** every source is listed with whether anyone actually read it, every correction is kept with the rule it produced, and nothing is described as verified that was not opened. If you find that claim is false anywhere, that finding gets published too. Archived at CERN — DOI [10.5281/zenodo.22029795](https://doi.org/10.5281/zenodo.22029795). --- ## Where to look | | What it holds | Start with | |---|---|---| | [**The Act**](./act/index.md) | The statute, the section-by-section Comments, the draft rules, and the bracketed choices an adopting state makes | [The Act, section by section](./act/comments.md) — what each provision does and why it is drafted that way | | [**Revision**](./revision/index.md) | How the Act is being repaired, and the dated record of how it reached v3.4.2 | [Revision](./revision/index.md) — what v3.4.2 changed, and what these pages still say against v3.4 | | [**Commentary**](./commentary/index.md) | Why it should exist — the case, the precedents, the questions asked most | [Objections](./commentary/objections.md) — every criticism made against it, the answers, and the ones with no answer | | [**For Legislatures**](./enactment/index.md) | What a sponsor's office needs before taking it up | [Fiscal note](./enactment/fiscal-note.md) — what the Act costs an adopting state | | [**Appendix**](./appendix/index.md) | What was actually checked, and what it showed | [The bill census](./appendix/bill-census.md) — every frontier AI bill in America, read one at a time | | [**Authorities**](./authorities/index.md) | Look-up material | [Table of authorities](./authorities/table-of-authorities.md) — every source, with whether anyone read it | | [**Corrections**](./corrections/index.md) | Every mistake this project has made, kept with the rule it produced | [Corrections](./corrections/corrections.md) — seventy-five entries, append-only | --- **Contact** `FrontierAIAccountabilityProject@proton.me` · **Cite** [`CITATION.cff`](./CITATION.cff) · **License** [CC0](./LICENSE)