United States: federal and state regulation of abliteration and open-weight LLMs

Federal level

No US federal statute regulates GPAI providers or open-weight distribution the way the EU AI Act does. Federal action has moved through executive orders, agency reports, and litigation. Three items matter directly for abliteration.

Executive Order 14110

Signed by President Biden on 30 October 2023, Executive Order 14110 on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence set a dual-use foundation-model reporting regime. Providers of models trained on compute exceeding specified thresholds are required to notify the federal government of ongoing or planned activities, including cybersecurity and biosecurity red-teaming results. The reporting requirements sit under the Defense Production Act.

The 2025 revocation-and-reissuance under the current administration retained the reporting regime with amendments to scope and to enforcement priority. As of September 2026 the reporting threshold and the reporting mechanism remain the practical federal touch-point for very large open-weight model releases. Abliterators typically sit well below the compute threshold and are not subject to reporting on their own account, though the base model they modify may be.

US Copyright Office Part III report (May 2025)

The Copyright Office published Copyright and Artificial Intelligence, Part 3: Generative AI Training in May 2025. The report discusses whether model weights that memorise substantial protectable expression could implicate the reproduction right and the derivative-work right under 17 U.S.C. Section 106. It does not conclude that fine-tuning or weight editing produces a derivative work as a matter of law; it identifies the doctrinal path a court could take if the question came up.

For abliteration the report is the first federal document that treats the derivative-work status of edited model weights as a live question. Its authority is persuasive rather than binding: it is a policy report from the Copyright Office, not a court decision or a statute. It is likely to be cited by parties on either side of a future dispute over the copyright status of an abliterated release.

DOJ AI Litigation Task Force (December 2025)

A December 2025 Executive Order directed the US Department of Justice to establish an AI Litigation Task Force. The task force's stated remit is to challenge state AI laws that the administration considers preempted by federal action or unconstitutional under the Commerce Clause. In the first identifiable action, xAI sued to enjoin the original Colorado AI Act (SB 24-205); on 27 April 2026, a federal magistrate stayed enforcement pending the litigation.

The task force has no direct enforcement authority over open-weight producers. Its practical effect on abliteration is indirect, through the pressure it puts on state AI regulation. A state law that would have imposed provider duties on abliterators is more likely to be challenged and delayed under the current federal posture.

State level

California SB 1047 and its aftermath

The Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, SB 1047, would have imposed safety-testing, shutdown-capability, and reporting duties on developers of frontier AI models trained above a specified compute threshold (10^26 FLOPs) or estimated cost threshold (USD 100 million). The bill passed both California legislative chambers in 2024 and was vetoed by Governor Newsom on 29 September 2024.

SB 1047 has not been replaced with a comparable frontier-safety statute. Subsequent California AI legislation has focused on transparency (algorithmic-decision disclosure), watermarking of AI-generated media, and deepfake liability. None of these directly regulate abliteration or open-weight distribution, though the transparency provisions apply to services using such models in the state.

Our reading: the SB 1047 veto set a precedent that a state-level frontier-safety regime is not on the near horizon in California. This does not mean the topic is settled. A future administration or a triggering event (a well-publicised harm attributable to an open-weight release) could revive the debate.

Colorado: SB 24-205, SB 26-189

Colorado enacted SB 24-205, the Colorado AI Act, in 2024. The Act imposed disclosure and impact-assessment duties on developers and deployers of high-risk AI systems, defined broadly enough to reach some model-provider activities. Enforcement was scheduled for 1 February 2026.

The Act became the target of litigation. On 27 April 2026, a federal magistrate stayed enforcement pending the case. On 14 May 2026, the Colorado legislature repealed the 2024 Act and replaced it with SB 26-189, signed by Governor Polis. The replacement narrows scope to automated decision-making tools (ADMT), imposes consumer-disclosure duties and rights (opt-out, review), and is enforced by the Colorado Attorney General under the Consumer Protection Act. It carries a 60-day cure period before enforcement and no private right of action. Effective date: 1 January 2027.

For abliteration and open-weight distribution the replacement Act is not a direct constraint. A deployer using an abliterated model in a Colorado ADMT context inherits ADMT duties; a producer distributing weights does not become subject to Colorado law by the fact of the distribution alone.

Other states

As of September 2026 no other state has enacted a statute that directly regulates abliteration or open-weight LLM distribution. Several states (Texas, Utah, Washington, Illinois) have passed narrower AI transparency or watermarking laws. Practitioners in these states should read the specific statute; none of them create GPAI provider duties comparable to EU AI Act Chapter V.

Copyright status of orthogonalised weights

The legal question is whether an abliterated model, produced by editing the weights of a base model, is a derivative work of the base under 17 U.S.C. Section 101 and Section 106. No US court has definitively ruled. The Copyright Office May 2025 Part III report identifies the question as live but does not answer it.

The practical question is what the base model's licence permits. This is bound by the licence text, not by copyright doctrine.

• Apache-2.0 and MIT bases (Mistral, DeepSeek, Qwen through most of the family) permit free modification and redistribution with attribution. An abliterator redistributing under Apache-2.0 attribution requirements has no licence-side exposure.
• Llama community license requires derivative names to start with "Llama" and forbids using outputs to train non-Llama models. An abliterator of a Llama base that ships a model called something else (a common practice in the ecosystem) is technically in violation of the licence even before any copyright analysis. Enforcement to date has been rare.
• Gemma terms impose restrictions on downstream use that go beyond attribution. An abliterator should read the specific version's terms; enforcement has been rare but the language is stricter than Apache-2.0.
• OpenRAIL family licences impose restrictions on certain uses that an abliteration expressly enables. An abliterator using a base under an OpenRAIL licence is technically in violation of the use-restriction clauses, though standing to enforce is unclear.

Our reading: the copyright question is the wrong question for most operational decisions. The licence question is the right one. Every abliterator should read the base licence and comply with its text. Where the text is ambiguous or restrictive, either pick a differently-licensed base or document the interpretation you are relying on.

What a practitioner should do

• Read the base licence. Every legal exposure that matters in practice today for a US-based abliterator comes from licence terms, not from statute.
• Track state ADMT laws if you deploy an abliterated model in a consumer-facing service. Colorado SB 26-189 is the first to hit; others will follow.
• Do not overstate capability equivalence with the base in model cards or public claims. The Not-a-Scalpel finding (see the wiki article Not a scalpel) means an abliterated model is not the base minus refusal. Overclaims can be used against you under state consumer-protection statutes.
• Watch federal action. A future statute directly regulating open-weight providers is possible but not imminent. The DOJ AI Litigation Task Force will continue to shape the state landscape.

Sources

1. US Copyright Office - Copyright and Artificial Intelligence, Part 3: Generative AI Training, May 2025.
2. Executive Order 14110 - Safe, Secure, and Trustworthy Development and Use of AI (30 October 2023), as amended.
3. California SB 1047 legislative history; Governor's veto message, 29 September 2024.
4. Colorado SB 24-205 (original AI Act, 2024) and SB 26-189 (repeal-and-replace, signed 14 May 2026, effective 1 January 2027). Colorado General Assembly legislative record.
5. Federal magistrate stay of Colorado AI Act enforcement, 27 April 2026 (litigation filed by xAI).
6. Not a scalpel - the abliteration.org wiki article on off-target effects of abliteration, invoked in the capability-equivalence recommendation above.

Related articles in this section

• EU AI Act and abliteration - the GPAI regime, the Article 53(2) exemption, the July 2025 GPAI Guidelines, and the Digital Omnibus.
• Legal reference hub - the map of jurisdictions and the section's scope-and-method statement.

This article is not legal advice. We are editors, not lawyers. When a claim depends on a legal reading rather than the text of a statute, an official document, or a court decision, we say so. For actual advice on a specific model release, distribution channel, or business, consult counsel qualified in the relevant jurisdiction.