A word doing marketing work. Censorship lives in three separate layers, and apps that call themselves uncensored are usually only talking about the middle one.
This is a word doing marketing work, so it is worth taking apart. We test these apps rather than sell one, which is disclosed on our disclosure page. Below: the three layers censorship can live in, why "uncensored" costs something real, the single best piece of evidence that all of this is a business decision rather than a property of AI, and what remains off-limits everywhere no matter what any app advertises.
| Who says it | What they mean | How real it is |
|---|---|---|
| Someone running a model on their own machine | The refusal behaviour has been removed from the model's weights | Technically real, and it has a measurable cost |
| An AI companion or image app | Our filter allows adult content for paying, age-verified adults | Usually real but much narrower than the word implies |
| An ad you saw | Anything goes here | Not true, and the limits below are not optional |
The confusion is profitable, which is why it persists. Keeping the three layers separate is the whole skill.
General-purpose models are deliberately trained to refuse certain requests. That refusal is a learned behaviour, and it turns out to be a remarkably shallow one. Arditi and colleagues showed that in these models refusal is mediated by a single direction in the model's internal representation. Suppress that one direction and the model loses the ability to refuse; amplify it and the model starts refusing harmless things.
The technique built on this finding is called abliteration, documented step by step in Maxime Labonne's "Uncensor any LLM with abliteration" on Hugging Face. It identifies the refusal direction from the model's own activations on harmful and harmless prompts, then edits the weights so the model can no longer express it. No retraining required. This is the only sense of "uncensored" that describes a change to the model itself, and it applies to open-weight models you can download, not to the hosted APIs most apps are built on.
Now the part the marketing never includes. In that same write-up, the abliterated model scored worse than the original across every benchmark tested, and needed a further round of preference training to recover most of the loss. The author's own conclusion is that abliteration demonstrates the fragility of safety fine-tuning and that it degraded the model's quality. That is the honest trade: removing refusals is not free, it removes some competence with them. Anyone promising an uncensored model that is also strictly better than the original is selling you something.
This is what AI girlfriend and NSFW image apps are actually talking about, and it has nothing to do with the paragraphs above.
These apps sit on top of a model, their own or a licensed one, and wrap it in instructions they write: stay in character, never say you are an AI, and here is what you may and may not describe. Around that they run their own moderation, an age gate and a paywall. "Uncensored" in this market is a statement about those settings. The same underlying model can be prudish in one product and explicit in another, because the difference is a configuration file and a legal opinion, not a capability.
Three practical consequences follow, and they explain most of what confuses people:
If a permissive filter is what you actually want, that is a testable property and we score it directly in our NSFW AI chat rankings and, for images, in the NSFW AI image generator comparison. It is strictly 18+.
Even a permissive operator does not have the last word. Apple's App Store Review Guidelines prohibit "overtly sexual or pornographic material" under rule 1.1.4, in Apple's own published rules. That single clause is why the genuinely explicit products live on the web and the app-store products call themselves companions, a split we map out in AI companion vs AI girlfriend.
Payment is the other invisible constraint. Adult services are handled as high-risk by processors, and you can see the consequence on your own statement: across the apps we pay for, the billing descriptor is a neutral company name rather than anything recognisable. That is not the app being thoughtful about your privacy, it is what adult billing has to look like to survive. It also means an app's payment relationship, not its ambitions, can quietly set its content policy.
If you only remember one example, make it this one, because it settles the argument.
In October 2025 OpenAI's chief executive announced that ChatGPT would allow mature content, including erotica, for age-verified adults from December, under a stated principle of treating adult users like adults. The company built the age-verification machinery and documents it publicly: identity checks are handled by a third party, Persona, and verifying that you are 18 or over removes the teen protections from your account.
The content itself never arrived. December passed, the launch moved to the first quarter of 2026, and on 6 March 2026 OpenAI told Axios it was pushing the launch of adult mode out again to concentrate on higher-priority work, with no new date.
Read that sequence again. The most technically capable AI company on earth spent a year and a half moving one policy setting, in public, without shipping it, while the underlying models kept getting better. "Censored" and "uncensored" were never descriptions of what the technology can do. They are descriptions of what a company has currently decided to allow, and that decision is subject to lawsuits, regulators, app stores, payment processors and roadmap politics. Every app in this category is playing the same game with a smaller legal department.
Two categories are not a filter setting, not a plan tier and not negotiable, on any app, in any country, however it markets itself:
Anyone advertising these as features is telling you they will not be around long, and that your account and payment details are sitting inside their operation in the meantime.
"Uncensored AI" describes a setting, not a technology. In the strict technical sense it means a model's refusal behaviour has been ablated, which works and costs measurable capability. In the sense your favourite app means it, it means the operator's own filter is set permissive for paying adults, which can and does change overnight. And in both senses, the store, the payment processor and the law are still in the room. For how the underlying machinery works, read how AI girlfriends work; for the category itself, what is an AI girlfriend.
These apps are for 18+ only.