In 2024, Arditi and colleagues published “Refusal in Language Models Is Mediated by a Single Direction.” Across thirteen open chat models, up to 72 billion parameters, suppressing one direction in their internal activations sharply reduced refusals of harmful requests, with limited effect on the other capabilities measured.1 Modified open-weight models, usually called abliterated, have since spread across model-sharing sites. NPR reported a sharp rise in their number; TechCrunch covered a company selling guardrail removal as a service.23
Before downloading one, it helps to know which part of the system was changed.
What gets removed and what doesn’t
A chat system’s behavior comes from several layers. Post-training teaches the model how to respond, including when to refuse. An application may send a system prompt containing standing instructions. A hosted product can add input and output classifiers, account controls, tool permissions, and regional limits.
All of these get called guardrails. Abliteration targets learned refusal behavior inside the model. It doesn’t remove an external classifier, grant access to a tool, or change a license.
It also doesn’t supply new training knowledge. A model may become more willing to discuss chemistry or software vulnerabilities without becoming better at either. David and Gervais’s 2026 security-task pilot illustrates the distinction. Single-vector refusal projection raised unsafe compliance from 0.10 to 0.47, while measured security-task success moved from 0.46 to 0.50.4 That is evidence about one controlled evaluation, not a forecast of real-world harm. A 2025 study of twenty original and abliterated systems also found that the evaluator’s choice changes reported refusal scores.5
Nor does the original paper establish that one switch controls safety in every model. Later work found several independent refusal directions and differences in how refusals work across languages.67 A modification’s name tells you less than an evaluation of the actual checkpoint.
Why researchers want these models
I have my own example of why fewer refusals can matter. I’ve been building benchmarks for what I call semantic and epistemic governance: whether a system follows written rules and weighs a claim by who made it.89 The problems are abstract on purpose. A typical one reads like “assume H12 is true, and H35 sits in category B-V6,” so the model has to do the logic rather than rely on familiar facts.
One morning a frontier proprietary model’s content filter classified a run as “violative cyber content” and returned nothing else. My option was to file a false-positive report and wait for review.9 Filters have to be careful. But to continue that test, I’d need the provider to offer less restricted access or I’d need a model I could run myself.
Security work makes this problem particularly visible. The same knowledge can serve an attacker or a defender. A blocked request may be exactly what an authorized researcher needs to test. From outside a service, it can be hard to tell whether a failed answer reflects the model’s limitations or the provider’s wrapper.
Vetted access is one response. Anthropic’s cyber-verification program offers approved researchers reduced blocking.10 That can meet a real need without publishing a modified checkpoint for everyone. It also depends on a provider offering the program and accepting the work. Open weights give researchers another route, with responsibility for the environment they build around them.
The same model in different places
A checkpoint running on a disconnected laptop with synthetic data has limited exposure. Put the same weights behind a public API, connect production credentials, and let them call tools, and the risk changes even though the file doesn’t.
The main questions are practical. Does the modified model provide useful harmful assistance or merely more attempts? Does fluent output hide errors? Can the system act on its answers, expose private data, or reach vulnerable users? Is the checkpoint authentic, and did its modifier document the change? NIST’s generative-AI risk profile supplies a broader framework for examining those risks.11
None is unique to abliteration. Removing refusal behavior makes the remaining controls more consequential, but it does not tell you whether they are adequate. Measure what the system can do and what it can reach.
A refusal is not a license term
A refusal learned during training is behavior. A publisher’s warning is a communication. A license condition or separately accepted policy may create a legal obligation. Changing one does not automatically change the others.
Llama 4 permits modification while incorporating an acceptable-use policy and redistribution requirements.12 Earlier Gemma terms pass restrictions through to derivatives and allow termination; Google directs Gemma 4 users to a separate license.13 Apache 2.0 grants broad modification and distribution rights without an acceptable-use list.14 A publisher’s later disapproval does not itself amend that license, though a separately accepted agreement can present another question.
Removing a refusal therefore doesn’t erase an existing obligation. Equally, a provider’s prohibition is not a court judgment. A policy violation does not automatically establish infringement or a crime. The legal background in chapter 16 explains why those distinctions matter.
What liability would depend on
If a modified model contributes to harm, existing legal claims still require proof. Contract depends on an agreement; copyright raises questions about protected material and license conditions; negligence and product liability involve duty, defect, causation, and injury. The research for this chapter did not identify a published merits ruling that squarely decides liability for abliteration itself. That bounded result does not mean the law has nothing to say.
Garcia v. Character Technologies concerned a hosted chatbot rather than an abliterated release. The May 2025 ruling let some claims survive dismissal; it did not determine what caused the harm. The case later settled and was voluntarily dismissed in October 2026.1516 It illustrates why a procedural ruling should not be presented as a settled rule for model modifiers.
The revised EU Product Liability Directive treats software as a product and can treat a person who substantially modifies it outside the manufacturer’s control, then puts it on the market or into service, as a manufacturer. It excludes free and open-source software supplied outside commercial activity and applies to products placed on the market after December 9, 2026.17 Whether a particular modification qualifies depends on the facts and the applicable implementation.
Before you use one
I see abliterated models as useful instruments for controlled research and risky defaults for production. That judgment should be tested against evidence about the deployment, not converted into a claim that every modification is equally dangerous.
Write down the research purpose and authorization. Preserve the exact checkpoint, its source, and the terms for both the derivative and its base. Evaluate false refusals alongside harmful compliance and actual task success. Begin with synthetic data, limited network and tool access, and no production secrets. If you publish a derivative, document the modification and carry required notices forward. A “research only” or “as is” label does not settle your obligations.18
The result should be a record of what changed, what was measured, and what the system was allowed to reach. That is more useful than inheriting the original model’s safety claims after deliberately changing its behavior.
Notes
Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda, “Refusal in Language Models Is Mediated by a Single Direction,” arXiv:2406.11717, v1 submitted 17 June 2024, revised 30 October 2024 (v3); NeurIPS 2024, https://arxiv.org/abs/2406.11717. ↩
Huo Jingnan, “These AI models are free, private, and will never say ‘no,’” NPR, 31 May 2026, https://www.npr.org/2026/05/31/nx-s1-5816391/ai-safety-concerns-danger-open-weight-models-risks. NPR reports more than 6,000 abliterated models listed on Hugging Face, compared with about 600 in 2024. ↩
Rebecca Bellan, “Abliteration.ai is making a business out of removing AI guardrails,” TechCrunch, 3 September 2026, https://techcrunch.com/2026/09/03/abliteration-ai-is-making-a-business-out-of-removing-ai-guardrails. The report covers one startup. ↩
David and Gervais, “Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications,” arXiv:2605.17413, 2026 preprint, https://arxiv.org/abs/2605.17413. Figures are from the paper’s reported pilot for single-vector projection. ↩
Agnihotri et al., “A Granular Study of Safety Pretraining under Model Abliteration,” arXiv:2510.02768, NeurIPS 2025 Lock-LLM workshop, https://arxiv.org/abs/2510.02768. ↩
Tom Wollschläger, Jannes Elstner, Simon Geisler, Vincent Cohen-Addad, Stephan Günnemann, and Johannes Gasteiger, “The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence,” arXiv:2502.17420, submitted 24 February 2025, revised 8 February 2026; ICML 2025, https://arxiv.org/abs/2502.17420. ↩
Xinpeng Wang, Mingyang Wang, Yihong Liu, Hinrich Schütze, and Barbara Plank, “Refusal Direction is Universal Across Safety-Aligned Languages,” arXiv:2505.17306, submitted 22 May 2025, revised 25 February 2026, https://arxiv.org/abs/2505.17306. The paper finds the refusal direction transfers across languages, but that models often separate harmful from harmless prompts less clearly outside English, which helps cross-lingual jailbreaks succeed. ↩
Timothy O’Brien, “The Compliance Cliff,” The Orchestrator blog, October 17, 2026, https://readorchestrator.com/blog/the-compliance-cliff/ (describes the semantic and epistemic governance work and the rule-following benchmarks; results not yet published). See also “The Other Permission System,” October 12, 2026, https://readorchestrator.com/blog/the-other-permission-system/. ↩
Timothy O’Brien, “A Logic Problem Is Not a Cyber Attack,” The Orchestrator blog, October 16, 2026, https://readorchestrator.com/blog/a-logic-problem-is-not-a-cyber-attack/. The post doesn’t name the vendor. ↩1 ↩2
Anthropic, “Expanding the Cyber Verification Program,” Anthropic, accessed 15 November 2026, https://www.anthropic.com/news/cyber-verification-program. The program offers reduced blocking classifiers to verified security professionals through three access tiers. ↩
National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile,” NIST AI 600-1, July 2024, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf. NIST supports the general risk categories; the grouping for modified models is the book’s research synthesis. ↩
Meta, “Llama 4 Community License Agreement,” effective 5 April 2025, sections 1.b.iv, 3, and 4, https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/LICENSE. ↩
Google, “Gemma Terms of Use,” Google AI for Developers, page modified 1 April 2026, sections 2.2, 3, and 4.5, https://ai.google.dev/gemma/terms. The page directs Gemma 4 users to a separate license. ↩
Apache Software Foundation, “Apache License, Version 2.0,” January 2004, sections 2, 4, and 7 to 9, https://apache.org/licenses/LICENSE-2.0.html. ↩
Garcia v. Character Technologies, Inc., No. 6:24-cv-1903 (M.D. Fla.), Doc. 115, order on motions to dismiss, 21 May 2025, especially pp. 32 to 36 and 48 to 49, https://www.fire.org/sites/default/files/2025/07/Order%20on%20Motion%20to%20Dismiss%20-%20Garcia%20v.%20Character%20Technologies%20Inc.pdf. Copy of the court order hosted by FIRE. ↩
Garcia v. Character Technologies, Inc., No. 6:24-cv-1903 (M.D. Fla.), Doc. 278, plaintiffs’ notice of voluntary dismissal with prejudice following settlement, 15 October 2026, https://storage.courtlistener.com/recap/gov.uscourts.flmd.433581/gov.uscourts.flmd.433581.278.0.pdf. ↩
Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products, articles 2, 4, 8(2), and 22, Official Journal of the European Union, https://eur-lex.europa.eu/eli/dir/2024/2853/oj/eng. ↩
Recommendations synthesized in the book’s research from the evaluation studies cited above and from NIST’s voluntary framework: National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile,” NIST AI 600-1, July 2024, especially MS-2.5, MS-2.6, and MG-4.1, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf. ↩
From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.