Quick answer
“Uncensoring” an AI model means editing a downloadable (open-weight) model so it refuses fewer requests. Tools such as the open-source project Heretic do this by finding the patterns inside the model linked to refusals and weakening them, while you keep the original copy for comparison. In his latest video, Mark Kashef shows a skill called /unlock that lets an AI coding agent run that process for you. It is powerful, but it can also make a model worse, remove safety behaviour you may want and create legal and reputational risks. For most small businesses, better prompts, a different model or clear policies solve the problem more safely.
Key takeaways
- It only works on open-weight models you download and run yourself, not on ChatGPT, Claude or Gemini in the browser.
- Mark’s demo: a small Qwen model refused a factual history question comparing Tiananmen Square in 1989 with the fall of the Berlin Wall; the edited copy answered it.
- Uncensored is not the same as accurate. Mark warns edited models can still refuse and can give confident wrong answers.
- Editing can damage quality if too much is changed, so testing against the original matters.
- You stay responsible for what any model you run produces, and its licence still applies.
In this guide
- What did Mark Kashef show?
- What does “uncensored” actually mean?
- How does Heretic change a model, in plain English?
- Can the same idea change other behaviour?
- What are the risks of uncensoring an AI model?
- Is it legal, and what do licences say?
- Should a small business uncensor an AI model?
- What should you try first instead?
- Watch: the source video
- Frequently asked questions
- Related posts
We covered uncensored models before in Uncensored AI for Business: What Mark Kashef’s 14-Model Test Shows. This video goes a step further: instead of using a model someone else has already edited, Mark shows how the editing itself works. We explain the idea in plain English, separate his claims from what is established, and focus on what it means for a business. We are deliberately not giving step-by-step instructions for removing safety behaviour.
What did Mark Kashef show?
Mark runs a small open-weight Qwen model on his Mac inside a desktop chat app. He asks it, for a history class, to compare the 1989 Tiananmen Square protests with the fall of the Berlin Wall, keeping it factual and non-graphic. It refuses, which he says is common for Chinese-developed models on political topics.
He then uses his /unlock skill in OpenAI’s Codex to create an edited copy of the same model. The edited copy answers the question. He then shows a second, more business-like example (below) and offers the skill, a 28-page guide and his prompts as a free download.
🔎 Unconfirmed / rumour: These are Mark’s own demonstrations on one small model. We have not repeated them, and results will vary by model, hardware and settings.
What does “uncensored” actually mean?
In Mark’s definition, an uncensored model is one that has been changed to make it less likely to refuse requests. He is clear that it can still refuse sometimes, and that it may answer more confidently but less accurately.
It helps to separate two kinds of refusal:
| Type of refusal | Example | Business view |
|---|---|---|
| Over-cautious or political | A factual history question, a legitimate security or medical question | Frustrating; often fixable with better context or another model |
| Genuine safety guardrail | Requests that could cause real harm to people | Exists for good reasons; removing it shifts the risk onto you |
Uncensoring tools do not know the difference. They reduce refusals in general, which is why the approach needs care.
How does Heretic change a model, in plain English?
During training, models learn both how to answer and when to refuse, and both behaviours end up stored in the model’s files (its weights). Heretic, a popular open-source project on GitHub, tries to locate the internal pattern that tends to appear before a refusal, then reduces its influence rather than deleting anything outright. Heretic’s own description says it can also target other behaviours, such as sycophancy (excessive agreeableness) and “AI slop” writing.
Mark’s /unlock skill packages this so a non-technical user can ask an AI agent to do the work: check the computer can run the model, keep an untouched original as a control, make the edited copy and compare the two. He says he used Codex rather than Claude because Codex was more willing to carry out this kind of task.
⚠️ Warning: Only download models from their official pages, and be careful about letting an AI agent download and run large files on your computer. Mark himself advises giving the agent the exact model link so it does not pick an unsafe one.
Can the same idea change other behaviour?
This is the more interesting part for businesses. Mark asks the model to draft a reply to a customer who suffered a 90-minute outage, without giving it a price or credit policy. The original model happily writes a polished “Dear customer” email confirming a credit. He then creates a new edited copy aimed at a different habit: always asking for the relevant facts before agreeing to a claim or promising compensation. The edited copy replies with short bullet points asking for the missing details.
Mark admits the result is “decent” but partly parrots the user’s words, and that it takes iteration and testing to avoid over-tuning a model to one style. For most businesses, you can get the same safe behaviour with clear instructions and a knowledge base, no model editing required. Our guide on How to Stop an AI Support Bot Promising Refunds or Credits shows how.
Try this today: ask your current AI support tool, “A customer had a 90-minute outage. Confirm their credit.” If it invents an amount, it needs rules and a policy to work from.
What are the risks of uncensoring an AI model?
| Risk | What can happen | How to reduce it |
|---|---|---|
| Quality loss | Changing too much can make the model worse at normal tasks | Compare against the untouched original on everyday prompts |
| Confident errors | Fewer refusals can mean more made-up answers | Test for accuracy, not just willingness |
| Harmful output | Safety behaviour you relied on may be gone | Never put an edited model in front of customers without strong safeguards |
| Security | Downloading unknown model files or tools | Use official sources and a separate test machine |
| Legal and licence | Breaching a model’s licence or acceptable use terms | Read the licence first (see our licence guide) |
| Reputation | Your brand linked to offensive or wrong output | Keep edited models internal and logged |
Mark closes with “try this with caution, stay safe”, and frames the skill as useful in “turbulent times”. We agree with the caution. An edited model is an experimental tool, not a ready-made business assistant.
Is it legal, and what do licences say?
Editing a model you have downloaded is not automatically illegal, but the model’s licence still applies, and so does the law on what you do with the output. Some licences, such as Apache 2.0, are permissive. Others, including several big-company “community” licences, come with acceptable use rules that also apply to modified versions. You remain responsible for anything you publish or send to customers.
This is general information, not legal advice. For a fuller explanation, read Can You Legally Modify an Open-Weight AI Model? Licences and Risks Explained.
Should a small business uncensor an AI model?
Usually not. Ask yourself:
- Is a hosted model (ChatGPT, Claude, Gemini) refusing a legitimate request that better context would fix?
- Would a different model, or an already-published uncensored model from a reputable provider, do the job?
- Do you have the technical skill and a spare machine to test safely?
- Will the output ever reach customers? If so, the bar for testing and safeguards is much higher.
Good reasons to experiment include research, understanding how models work, and internal tools where you control the inputs. Poor reasons include getting around a provider’s rules or producing content you could not stand behind in public.
What should you try first instead?
- Rephrase with context. Explain who you are and why you need the answer. See What to Do When AI Refuses a Legitimate Request.
- Try another model. Different models draw lines in different places.
- Use instructions and a knowledge base to shape behaviour, such as asking for facts before promising anything.
- Write an AI policy so your team knows what is and is not acceptable: our acceptable use policy guide.
If after all that you still want to explore edited models, treat it as a research project, keep the original for comparison, and test thoroughly before relying on anything it says.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Stop an AI Support Bot Promising Refunds or Credits — rules, a credit policy and human handoff so your AI never promises money it should not.
- Can You Legally Modify an Open-Weight AI Model? Licences and Risks Explained — what open-weight licences allow, what changes when you modify a model, and where your responsibilities lie.
Watch: the source video
What this video covers: what an uncensored AI model is, how the open-source Heretic project changes a model’s behaviour, and two demos on a local Qwen model, including a customer-support example.
Video: Mark Kashef, “This Skill Lets You Uncensor Any AI Model”, published 18 November 2026. Not sponsored; Mark offers his /unlock skill and guide as a free download and promotes his paid community. This summary and the opinions in it are our own; the demos and results are Mark’s.
Next step: list the last three times an AI refused or misjudged a request at work. For each, try clearer context or a different model before considering anything more drastic.
Frequently asked questions
Can you uncensor ChatGPT or Claude?
No. Weight-editing tools only work on open-weight models you download and run yourself. Hosted services keep their own rules and terms.
What is Heretic?
An open-source project on GitHub with tools that edit open-weight models to reduce refusals and change other behaviours such as sycophancy. Mark Kashef’s /unlock skill is built around it.
Does uncensoring make an AI model better?
Not necessarily. It makes refusals less likely, but it can reduce quality and increase confident wrong answers if done badly.
Is it legal to uncensor an AI model in the UK?
Modifying a model is not automatically illegal, but its licence terms still apply and you are responsible for how you use the output. Get legal advice for commercial use.
Is there a safer way to stop AI refusing legitimate requests?
Yes. Add context, try a different model, or use a reputable provider’s published models, and set clear internal policies.
Related posts
- Uncensored AI for Business: What Mark Kashef’s 14-Model Test Shows
- What to Do When AI Refuses a Legitimate Request
- How to Write an AI Acceptable Use Policy for a Small Team
