Isometric workshop island where a glowing unlocked AI cube sits inside a fenced test area while a small inspector checks it, with a shield, key and checklist floating nearby

Quick answer

Uncensored AI for business means using an AI model that has been modified to refuse far fewer requests than mainstream assistants such as Claude or ChatGPT. In a new video, Mark Kashef says he tested 14 uncensored models with 15 identical requests, and found DeepSeek 4.1 Flash on Venice the most willing, fastest and cheapest. Used carefully, these models can cut frustrating false refusals and can run privately on your own hardware. But removing the guardrails also removes a layer of protection, so the legal, ethical, brand and security responsibility sits entirely with you. For most small businesses, the better first step is fixing false refusals with mainstream tools and writing clear AI rules for the team.

Key takeaways

  • Uncensored (also called abliterated or jailbroken) models are open models altered to say “no” less often.
  • Mark reports that 10 of the 14 models he tested would do most of his tasks, with DeepSeek 4.1 Flash on Venice his pick.
  • His demos include a lookalike of the Grok Bot desktop app, tagging family photos by face and moving home camera recordings to his own server.
  • Fewer refusals is not the same as fewer rules. Copyright, data protection, product terms and device safety still apply.
  • Mark himself warns these models hallucinate more and their answers need double-checking.
  • The safest wins come from local, private use on low-risk work, with written rules and human review.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

If you have ever asked an AI assistant for help with something perfectly normal and got a lecture instead, you are not alone. False refusals are one of the most common complaints about mainstream AI tools, and they are driving interest in uncensored AI for business.

In his latest video, Mark Kashef says he tested 14 uncensored AI models to find one that would help with projects his usual tools kept blocking. This article explains what he tested, what he built, and, more importantly for a business owner, where the real risks sit and how to stay on the right side of them. For the basics of how these models are made, see our earlier explainer, Uncensored AI Models: What They Are, Legitimate Uses and the Real Risks.

⚠️ Warning: This article is for understanding, not a how-to. We do not cover how to remove or get around AI safety features, and nothing here is legal advice. If a project could affect other people’s rights, data or devices, get proper advice first.

Watch: the source video

What this video covers: Mark Kashef explains what uncensored AI models are, how he narrowed thousands of Hugging Face variants down to 14 core models, and why he chose DeepSeek 4.1 Flash through Venice. He then demos a Grok Bot lookalike desktop app built with the model, a family photo-tagging task that Claude refused, and rerouting home security camera recordings to his own server.

Video: Mark Kashef (@Mark_Kashef). This summary and the opinions in it are our own. Mark says he is not affiliated with the model providers he names. The video description also links to his community, a free kit and an app-security tool. We have not tested the models or tools shown, and we do not repeat or link to the requests he says he could not show on YouTube.

What is an uncensored AI model?

An uncensored AI model is a version of an existing open model that someone has modified so it is much less likely to refuse requests. You will also see the terms “abliterated” and “jailbroken”. The original model’s knowledge is mostly the same. What changes is how readily it says yes.

Mark makes three points worth keeping in mind:

  • They are easy to find. A search on Hugging Face, the main site where people share AI models, turns up hundreds or thousands of listings, many of them variations of the same few models.
  • They are not all equal. The same base model can behave differently depending on who modified it and which provider hosts it.
  • They make more mistakes. Mark says these modified models are more prone to hallucinations, so you need to “double and triple check” their answers.

💡 Tip: Downloading any model file from a public site carries its own risks. Our guide on checking a downloaded AI model before you run it covers trusted publishers, safe file formats and licences.

What did Mark Kashef test, and which model came out on top?

Mark says he sent the same 15 requests to 14 core uncensored models and picked DeepSeek 4.1 Flash, hosted on Venice, as the most obedient, best performing and most cost-effective. Here is what the video reports:

What he reportsDetail
Models tested14 core models, narrowed down from thousands of variants
Requests15 identical prompts, from everyday tasks that mainstream tools refused to ones he calls “risky”
Providers usedVenice, abliteration.ai and OrcaRouter (he says he is not affiliated with any of them)
Willing models10 of the 14 did most of the tasks
His pickDeepSeek 4.1 Flash on Venice: fast, cheap and the most willing
How he used itConnected to an open-source DeepSeek “harness”, a coding-agent workspace similar to a desktop coding app

He also notes that Venice did not ask for identity checks (known as KYC, or “know your customer”). He presents this as a convenience. For a business, it cuts both ways: less friction to sign up, but less accountability and no reassurance about how your data is handled.

Mark says he could not show some of the requests on YouTube, so he summarised the results instead. We have not seen or tested those requests, and we have no interest in repeating them.

What did he build with it?

He shows three projects. We describe them so you can judge the risks, not as instructions to copy.

1. A lookalike of the Grok Bot desktop app

Mark used the model with a reverse-engineering tool to build “Gratis”, a close copy of the Grok Bot desktop app. It routes chats through his own Claude and Codex subscriptions and gives each bot its own browser computer. In the demo, a bot searches Google Flights and heads towards checkout. He says the build took about an hour and 15 minutes and cost about $5 in model usage.

One genuinely useful habit from this section: because the modified model can hallucinate, he had it build the copy in a separate folder, away from the original, and test itself against the real app. Keeping experiments away from anything that matters is good practice for any AI agent.

2. Tagging family photos by face

Mark says Claude refused to identify his family in around 10,000 photos, even with reference images. The uncensored model did it in under 15 minutes. His process was sensible: reference photos for each person, work on a copy, list every file and who was tagged, quarantine anything uncertain for his approval, and test on a small sample first.

3. Moving home security camera recordings

Finally, he says the model helped him reroute recordings from his own home cameras, which normally go to a paid cloud plan, to his own server, possibly by changing the camera software. He is clear he is “not advocating” this and tells viewers to do their own due diligence.

Is there a legitimate case for uncensored AI for business?

Yes, in narrow cases, though most of them are about fewer false refusals and more privacy rather than doing anything a mainstream tool would never allow.

Legitimate reasonExampleA safer first option
Fewer false refusalsA request blocked because it mentions security, health or a famous brand, even though your purpose is fineRephrase with honest context, or use a purpose-built tool. See what to do when AI refuses a legitimate request.
Private, local useWorking on sensitive notes that should never leave your machineA standard open-weight model run locally, with networking switched off
Defensive security testingTesting your own app, on your own staging copy, before attackers doA licensed security tool, or the steps in our hack-test guide
Research and moderationUnderstanding harmful content in order to filter itSpecialist moderation tools with proper oversight
Cost and controlAvoiding usage caps and keeping your own copy of a modelOpen models from well-known publishers

Notice that most of these needs can be met without an uncensored model at all. A standard open-weight model run on your own computer already gives you privacy and control. Our guide on how much memory you need to run an open-weight AI model is a good place to start.

What are the risks of uncensored AI for business?

The main risks are legal, ethical, brand, security and quality risks, and an uncensored model removes the one layer that sometimes warned you about them.

RiskWhat it looks like in practice
Legal: copyright and termsCopying another company’s app “pixel for pixel”, or recreating a game, can infringe copyright, design rights or the software’s terms. Running paid AI subscriptions through a home-made app may also break the provider’s terms.
Legal: personal dataFace recognition creates biometric data. Using it on staff or customer photos in the UK is special category data under UK GDPR and needs a lawful basis, a specific condition and usually a data protection impact assessment.
Devices and contractsChanging camera firmware can void warranties, break security updates and breach the service terms. Never touch devices you do not own.
SecurityUnvetted models, providers and harnesses can carry malicious code or quietly send data elsewhere. Some providers have no identity checks and unclear data policies.
QualityBy Mark’s own account, these models hallucinate more. Fewer refusals also means fewer “are you sure?” moments.
Brand and trustIf customers or staff learn you used an unfiltered tool on their data, or for a copycat product, the reputational damage can outlast any time saved.
EthicsJust because a model will do something does not mean you should. The guardrails exist partly to protect people who are not in the room.

⚠️ Warning: Mark mentions reports of account bans for requests that mainstream providers treat as off-limits. Repeatedly trying to push a tool past its rules can lose you access to the accounts your business depends on.

What does the video not prove?

🔎 Unconfirmed / rumour: The results come from one creator’s own tests. We do not know the exact requests, how “willing” was scored, which model versions were used, or whether results would be the same today. The $5 build cost and the claim that the camera recordings improved in quality are his own reports and have not been checked independently. Model names and provider behaviour change quickly.

The video also does not show any checks on where your data goes when you use these providers, or whether the clone app, the photo tagging or the camera change would be lawful in a business setting. Those are exactly the questions a business needs answered first.

How can a business use uncensored AI safely?

Treat an uncensored model like a powerful, unpredictable contractor: give it a narrow job, keep it away from anything sensitive, and check everything it hands back.

  1. Write the rule before the experiment. Decide who may use uncensored or unvetted models, for what, and with what data. Our guide on writing an AI acceptable use policy for a small team gives you a one-page template.
  2. Try the mainstream route first. Many refusals disappear with honest context or the right tool.
  3. Check the law and the terms. Copyright, UK GDPR, device terms and the AI provider’s own rules all still apply. Our checklist on whether new AI rules affect your business helps here.
  4. Prefer local and isolated. Run models on a separate machine or account, with no access to your main files or passwords. See how to give an AI agent its own computer and how to stop an AI agent sending your data online.
  5. Use public or dummy data only with any hosted provider whose data terms you have not read.
  6. Work on copies, test small, review by hand. Mark’s photo process is a good pattern: sample first, quarantine anything uncertain, and approve before scaling.
  7. Keep a log. Note the model, provider, task and outcome, so you can explain any decision later. Our AI decision log template works well.

Should your business use one at all?

For most small businesses, no, not as an everyday tool. The everyday wins, such as better drafts, faster research and tidier data, come from mainstream tools used well.

It may be worth a controlled look if you have a technical person, a clear and lawful use case that mainstream tools keep blocking, and the discipline to run it locally and in isolation with written rules.

Stay away if the task involves other people’s personal data, another company’s product, devices you do not own, or anything you would not be comfortable explaining to a customer.

Frequently asked questions

What is an uncensored AI model?

It is an open AI model that someone has modified so it refuses far fewer requests. You may also see it called abliterated or jailbroken. Its knowledge is mostly unchanged, but its willingness to say yes is much higher.

Is uncensored AI legal in the UK?

Downloading or running an open model is generally lawful, but what you do with it is covered by the same laws as anything else, including copyright, data protection and computer misuse laws, plus the terms of the services you use. Take advice for anything borderline.

Which uncensored model did Mark Kashef choose?

He chose DeepSeek 4.1 Flash, accessed through Venice, as the most willing, fastest and most cost-effective of the 14 he tested. This is his own test, and results can change as models and providers update.

Are uncensored models less accurate?

Often, yes. Mark says modified models are more prone to hallucinations, so their answers should be double-checked. Removing refusals can also remove some of the caution that catches mistakes.

Can I run an uncensored model privately?

Yes. Open models, including uncensored ones, can run on your own computer or a rented server, which keeps your data off third-party services. Check the model file and its licence first, and keep it isolated from your main work.

Why does Claude refuse to tag faces in my photos?

Face recognition creates sensitive biometric data, so many AI providers block it by default. Built-in features in Apple Photos or Google Photos already group faces for personal libraries, which is usually the simplest legitimate route.

Should my team use uncensored AI at work?

Only with written permission, a clear lawful purpose, isolated set-ups and no personal or client data. Most teams are better served by a clear AI policy and mainstream tools.

The bottom line

Mark Kashef’s 14-model test shows how easy it now is to find an AI model that will say yes to almost anything. That is useful when a mainstream tool blocks a reasonable request, and local, private use has real appeal. But an uncensored model does not change what is legal, fair or safe. It just stops reminding you. For a small business, the smart order is simple: fix false refusals honestly, write your team’s AI rules, and only then consider a tightly controlled experiment. Your first step today: draft your one-page AI policy.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.
2 thoughts on “Uncensored AI for Business: What Mark Kashef’s 14-Model Test Shows”

Leave a Reply