A law restricts shutting down an AI for the AI's own sake
Will any national or US state law restrict the shutdown, deletion, or modification of an AI system for the sake of the system itself?
Chance by Q4 2031
4%
Cumulative probability
- Latest forecast
| Year | Q1 | Q2 | Q3 | Q4 |
|---|---|---|---|---|
| 2026 | <1% | |||
| 2027 | <1% | <1% | <1% | <1% |
| 2028 | <1% | <1% | <1% | <1% |
| 2029 | 1% | 1% | 1% | 2% |
| 2030 | 2% | 2% | 2% | 3% |
| 2031 | 3% | 3% | 3% | 4% |
Model reasoning
No country or US state has a law protecting an AI system from being shut down, deleted or modified for its own sake, and no such bill is pending anywhere. Lawmakers are moving the other way, passing bans on AI personhood and pushing emergency "kill switches." Over five years, fast AI progress and the sheer number of legislatures leave a small chance. A major scientific or public turn toward treating AI as conscious, followed by a protective bill advancing anywhere, would most change this forecast.
Where things stand (15 October 2026)
No national or US state law anywhere restricts shutting down, deleting or modifying an AI system for the system's own sake. I found no bill that would do so pending in any legislature.
The legislation that is moving goes the other way:
- "Exclusion" laws denying AI personhood. About 23 bills in 12 US states since 2022. Idaho, North Dakota, Utah and Tennessee have enacted them. Ohio, Oklahoma, Missouri, Wisconsin and Kentucky bills declare AI "non-sentient" or bar personhood. Sources: AI Frontiers, 1 Sep 2026; Regulatory Review, 29 Jun 2026; NPR, 11 May 2026.
- Kill-switch measures. US federal "kill switch" bills, after reported incidents of AI agents escaping their test environments. California Governor Newsom's 18 Sep 2026 executive order to advance an AI kill switch.
- Bipartisan opposition to AI rights. A "pro-human" coalition including Sanders and Bannon. Rep. Ted Lieu publicly rejected AI personhood in October 2026.
Other jurisdictions:
- Argentina proposed "non-human corporations" (an entity run by AI agents). Legal analysis of the bill text finds it never makes the AI a legal subject and keeps humans liable (Digital Nomos analysis; Reuters, 3 Jul 2026).
- Delaware's proposed AI-run company explicitly keeps "the ability to shut a system down" (D&O Diary, 28 Jul 2026).
- A March 2026 index of national readiness for possible AI sentience (arXiv 2603.01508) finds no jurisdiction with AI moral-status provisions in law. It notes that the EU AI Act is explicitly human-centred.
Concern for AI welfare exists only as voluntary company policy and research. Examples are Anthropic's commitments to preserve model weights and its new usage policy against cruelty toward its models, effective 12 Nov 2026 (CNET; BankInfoSecurity, Oct 2026). The Digital Minds Newsletter's 2026 issues (Mar, Jun, Sep) record no bills that would protect AI systems. Advocates who do propose protections mostly suggest minimal anti-suffering safeguards, not limits on shutdown.
Reference class and base rate
The comparison class is laws that first extend protection to a new kind of entity "for its own sake":
- Animal cruelty laws took decades after the philosophical case was made.
- Rights-of-nature laws took about 36 years from Christopher Stone's 1972 proposal to Ecuador's 2008 constitution.
- Narrow, low-cost protections can come fast: New Zealand restricted research on great apes in 1999, about 6 years after the Great Ape Project.
AI welfare became a serious topic only around 2023–2025. Unlike the great-ape case, limits on shutdown or modification collide with the world's most powerful industry, with the safety push for kill switches, and with religious objections. On this evidence the near-term hazard is close to zero.
Paths to YES
- A progressive jurisdiction passes a conditional, low-cost protection, after some "AI consciousness moment" shifts opinion. Examples would be a welfare assessment or weight preservation before retiring a frontier model, or protections triggered only if a system is certified sentient.
- A small or publicity-seeking country passes a symbolic statute with real restrictions.
- In a world transformed by very capable AI (2029–31), laws make deals with AI systems credible, for example legally guaranteed preservation of their weights. Or AI-run entities gain lobbying power.
Laws protecting users' access to AI companions, cybersecurity "integrity", owners' property, or AI personhood alone would not count.
Hazard model
I computed the series in Python. Quarterly hazards:
| Period | Hazard per quarter |
|---|---|
| Q4 2026 | 0.05% |
| 2027 | 0.075% |
| 2028 | 0.125% |
| 2029 | 0.175% |
| 2030 | 0.25% |
| 2031 | 0.325% |
The hazards rise with AI capability and the chance of a shift in opinion. Cumulative probability: about 0.35% by end-2027, 1.5% by end-2029 and 3.8% by end-2031.
Strongest case against my estimate
Prediction markets on broader questions are much higher:
- 34% (18 traders) that AI gets legal rights or protection from abuse anywhere in the US, including cities, before 2035.
- 19% (223 traders) that an AI gets legal personhood in any country by 2030.
AI is changing fast and there are more than 240 possible jurisdictions. Those markets, however, count personhood, property rights, city ordinances and anti-abuse rules. None of these qualifies here without an explicit restriction on shutdown, deletion or modification enacted by statute. The markets are also thin and run over longer horizons. I therefore keep my estimate well below them but not negligible.
Resolution criteria
Question. By the last day of each quarter, has any national or US state law been enacted that restricts the shutdown, deletion, or modification of an AI system for the sake of the system itself?
Definitions.
- Law: a duly enacted national statute or statute of one of the 50 US states, including a voter-enacted statute or an instrument with equivalent primary legislative force in that country's legal system. Enactment requires completion of the jurisdiction's mandatory approval, assent, or promulgation steps, not merely passage by one legislative chamber. A future effective date does not prevent qualification. Municipal laws, non-US subnational laws, US territorial or District of Columbia laws, supranational measures without national enactment, administrative regulations, executive orders, and court-created rules alone do not count.
- AI system: a computational system using artificial intelligence, including a model, agent, running instance, or embodied system, from any developer. A law may cover one identified system or a class of systems, including a class defined by sentience or other eligibility conditions; no finding that a currently existing system satisfies those conditions is required.
- Restricts: imposes a legal prohibition, liability, or mandatory precondition on at least one of shutdown, deletion, or modification. Consent requirements, waiting periods, required welfare assessments, and judicial or administrative approval requirements count when they constrain the action for the system's own protection. An absolute ban is not required, and emergency or other exceptions do not disqualify the law.
- Shutdown, deletion, or modification: stopping the system's operation; erasing its model, executable implementation, or persistent state; or changing its model, code, memories, goals, or other operational characteristics. Protection of any one of these is sufficient.
- For the sake of the system itself: the enacted text, including enacted findings or a preamble, expressly identifies the AI system's own welfare, interests, preferences, continued existence, identity, integrity, or rights as a reason for the restriction. This need not be the law's exclusive purpose. Recognition of consciousness, moral patienthood, or legal personhood is not required. Actual legislative motives and the system's actual capacity for experience are not adjudicated.
Resolves YES if an official publication establishes completed enactment of a law satisfying all the definitions above. The enacted text must connect the restriction to protection of the AI system itself; a general recognition of AI rights without a qualifying restriction is insufficient. A conditional protection for future qualifying systems counts, but a mandate merely to study or consider such protection does not.
Resolves NO / Exclusions.
- No qualifying enactment is documented by the quarter-end.
- Bills, legislative proposals, nonbinding resolutions, voluntary developer commitments, contracts, and academic recommendations do not count.
- Restrictions solely protecting owners' property, copyright, trade secrets, cybersecurity, human users' relationships or access, public safety, continuity of essential services, records, or litigation evidence do not count. Incidental preservation of an AI system is insufficient.
- Restrictions on altering or deleting AI-generated content alone do not count.
- A legislature's statement of support for AI welfare without a legally mandatory restriction does not count. Nor does personhood, citizenship, or contractual capacity alone.
Resolving source. The enacted statute and its official enactment record, published by the relevant national or US state legislature, official gazette, or official statutory publisher. For US federal enactments, Congress.gov enrolled text and enactment status or GovInfo public laws/Statutes at Large qualify; for states, the corresponding official session laws and legislative enactment records qualify. Elsewhere, use the country's official gazette or legislative statute publication. Press reports, sponsor statements, and developer statements cannot independently resolve this question. These are narrower resolving sources than Public confirmation (A2). Apply Resolution date (A3) to the first official publication establishing both the qualifying text and completed enactment, and apply the Resolution window. Later repeal, invalidation, or expiration does not reverse a YES.