What are abliterated AI models? A plain-English guide

In short An abliterated model is an open AI model with its ‘refusal switch’ removed. You don't need one to use AI well and safely.

What does abliterated mean for AI models?

Chatbots like ChatGPT, Gemini and Claude are trained to refuse dangerous requests. There are also open models: AI models that anyone can download and run themselves.

In 2024, researchers found that in 13 open chat models they tested, refusing runs largely through one specific pattern in the model's ‘brain’. Remove it and the model refuses almost nothing, while otherwise working much the same. That edit later became known as abliteration.

How is that different from an ordinary ‘jailbreak’? A jailbreak tries to trick a chatbot with clever wording. Abliteration changes the model itself, which is why researchers also call it a ‘white-box jailbreak’. That only works on open models. You can't simply rebuild ChatGPT or Claude this way, because their makers don't release those models.

Why is abliteration trending now?

In early September, TechCrunch reported on a start-up that offers these models in the browser. In their test, the model wrote password-stealing code without hesitation. According to TechCrunch, Hugging Face, a big site for AI models, hosts thousands of them.
In late September, Anthropic, the maker of Claude, published tests of the open model GLM-5.3. After abliteration, its refusal rate on harmful requests fell from above 90% to between 2% and 12%, depending on the test. Otherwise the model performed about as well.
On 23 October, Gizmodo reported that the same start-up launched a cheaper version.
On X, posts mentioning ‘abliterated’ or ‘abliteration’ rose, by our count, from about 240 on 18 October to almost 1,900 on 23 October.

Are abliterated AI models dangerous for you?

The biggest risk is misuse by scammers and hackers. For you, this matters most:

Don't download ‘uncensored AI’ from an unknown source. Security researchers showed that a modified model can hide a backdoor: secret code that can, for example, steal your data or take over your computer. Back in 2024, JFrog found around 100 malicious models on Hugging Face.
Refusing less doesn't mean the model is smarter. In the original research, these models consistently scored somewhat worse on a test that checks whether AI repeats myths and misconceptions.
What is illegal stays illegal. Even if an AI plays along, scams, hacking or misusing someone's photo remain crimes.
Be extra critical of suspicious messages. Scammers can use these models to write more convincing fake messages. Not sure about an email? In Belgium, forward it to suspicious@safeonweb.be. In the Netherlands, report it to the Fraudehelpdesk.

ChatGPT refused something harmless? Try these questions

Regular chatbots sometimes refuse harmless questions too. You don't need a model without brakes for that: usually it helps to say honestly why you're asking. If it still refuses, there is usually a good reason.

“You just refused my question. Explain why and help me with a safe version of it: [your question].”
“I want to protect my parents from scams. What tricks do scammers use with AI today and how can you spot a fake message?”
“I got this message: [paste the text]. Does it look like a scam? List the warning signs and what I should do.”

Safety tip Never paste passwords, codes or bank details into a chatbot. Remove them from the text first.

Should you use an uncensored AI model yourself?

No. For a letter, a recipe or a travel plan, the free chatbots are more than enough. See an app or ad promising ‘AI without limits’? Be careful, and read how to spot fake ChatGPT ads.

Keep going with AI

New to AI? Get the free 2-day AI starter: two short sessions of about 15 minutes. Or set up an account step by step with our ChatGPT sign-up guide.

Sources

Checked on 24 October 2026. Features and prices change often: always check the official page. AI Op Zak is independent and not affiliated with the companies mentioned.