As of November 13, 2026, the Hugging Face content policy does not mention models whose refusal behaviour has been removed, and the platform lists tens of thousands of them. The policy restricts categories of content and use, not a type of model. Enforcement is mostly report-driven, and publicly documented removals are few: we found four cases between 2022 and 2026, one of them an abliterated model disabled in 2026. The policy text in force is dated April 10, 2025.
What the Policy Restricts
The content policy applies to models, datasets and demos. It lists five groups of restricted content:
- Unlawful or illegal content, including high-risk illegal activity such as weapons development.
- Fraudulent or malicious activity, such as disinformation and phishing.
- Harmful or abusive content, including harassment, hate and glorified violence.
- Privacy and intellectual property violations.
- Platform abuse, security violations and spam, including malware and unauthorised access.
It prohibits any sexual content involving minors and sexual content created without explicit consent, and it describes consent as a core value. It says the company may also moderate other content as machine learning changes.
The Mechanisms
| Mechanism | What it does | Who applies it |
|---|---|---|
| "Not For All Audiences" tag | Shows a warning before the repository opens. Visitors can click through, or turn the warning off in their settings | The owner, or moderators as an enforcement action |
| Gating, automatic approval | The page stays visible. Files need a logged-in user to accept conditions, and access is granted at once | The owner. The platform may ask for it |
| Gating, manual approval | As above, but the owner accepts or rejects each request | The owner |
| Reduced visibility | The repository is unranked in trending and feeds | The platform |
| Disabled | Access is blocked for everyone except the owner. The page shows a notice | The platform |
| Account action | Interactions restricted, or the account suspended or terminated | The platform |
The policy says the company prefers to work with the repository owner first. Reports are made from a menu on each repository and are posted publicly, or sent by email. Owners can contest a decision by email, and under the EU Digital Services Act can go to a certified dispute body. Gating by default shares the requester's username and email with the owner, according to the gated models documentation.
How Often Uncensored Models Use Them
Among the 27,959 public repositories tagged or named "uncensored", "abliterated", "heretic" or "decensored" on November 13, 2026:
| Status | Repositories | Share |
|---|---|---|
| Carry the "not-for-all-audiences" tag | 1,475 | 5.3% |
| Gated, automatic approval | 391 | 1.4% |
| Gated, manual approval | 126 | 0.5% |
| Open, with no gate | 27,442 | 98.2% |
Disabled repositories do not appear in the API listing, so this count cannot show how many have been removed. The full breakdown is in uncensored model statistics on Hugging Face.
Publicly Reported Takedowns
| When | What | What is documented |
|---|---|---|
| 2022 | GPT-4chan, a model trained on forum posts | Reported to have been gated and then disabled over concerns about harm. The API still shows the repository as disabled on November 13, 2026 |
| July 2025 | Image models built to reproduce real people's likeness | 404 Media reported more than 5,000 such models, reuploaded after Civitai banned them. Transformer reported on July 24 that the company removed some, but not all, of the models it was sent |
| Between August 31 and mid-September 2026 | An abliterated GLM-5.3 repository whose name said it was for offensive cyber use | The page reads "Access to this model has been disabled" and says the content infringes the content policy. The API shows it created on August 14, 2026, last modified on August 31, and disabled. No clause was cited. A repository from the same publisher without the "offensive cyber" wording, created on September 1, is public and gated |
| October 2026 | An abliterated Kimi K2.7 Code repository | A single blog, citing a social media post, reported it had stopped responding. The cause is unknown and the platform has not commented. Other accounts' copies under that name were listed on November 13 |
The 2026 cases led some posts to say the platform was banning uncensored models. The counts above do not support that: about 3,900 matching repositories were created between October 1 and November 13, 2026 and remain public. The one confirmed 2026 removal involved a stated offensive purpose, and the reason given was the content policy in general. Hugging Face has published no explanation, so any link to a specific clause is an inference.
Other Hosts
| Host | Documented position |
|---|---|
| GitHub | Its acceptable use policies prohibit sexually obscene content and content that directly supports unlawful active attacks or malware campaigns, with an allowance for dual-use security work. They do not address model safety behaviour |
| Civitai | Announced in May 2025 that it would ban models designed to generate real people's likeness, citing new regulation in the United States and Europe, with payment processor pressure also reported by 404 Media |
We did not find a published policy from any major host that names abliterated or uncensored language models as a category. For the research on why safety behaviour is easy to remove from open weights, see refusal direction.
Methodology
Policy details are from the Hugging Face content policy and gated models documentation, read on November 13, 2026, and from the company's March 2023 newsletter describing the tag and the options for high-risk repositories. Counts come from the public Hugging Face API on the same day, using the set described in our statistics page. Repository status for the takedown table was checked directly against the API and the public page at 21:37 UTC. Reported cases were found by web search, so the list covers only removals that someone wrote about. It is not a record of all enforcement, and Hugging Face does not publish one that we could find.