Uncensored is a marketing word describing a spectrum, not a technical claim describing a product.
When a companion declines to do something, people read it as the model being censored. Usually it is not the model at all, and the distinction matters because the three layers behave completely differently.
The first layer is the base model. Some are trained with adult content in scope and some are not, and this is the only layer where the limit is genuinely baked in. The second is the system prompt β the invisible instructions the platform puts in front of your conversation, which set the character's boundaries. This is the layer platforms actually mean when they advertise uncensored: they have written a permissive system prompt. The third is an output filter that scans what the model produced and blocks or rewrites it.
You can tell them apart by how the refusal feels. A base-model refusal is in character and reasoned. A system-prompt refusal is abrupt and repeats the same phrasing. An output filter produces the strangest behaviour: the message starts generating, then vanishes or is replaced. If you have seen text appear and then disappear, you have watched a filter fire.
Thirteen of the fourteen platforms on this site allow adult content as a matter of policy. Replika is the only one that does not, and it is worth knowing why: it did, and then it stopped.
In practice the permissive end of the category β Muah AI, Janitor AI, CrushOn AI β puts almost nothing in the system prompt and does very little output filtering. The chat-quality end β Nomi AI at 9.1, Kindroid at 9.0 β is permissive too, but the characters have stronger personalities, which means they will sometimes decline in character. That is not a filter. That is the product working: a companion with a coherent personality is one that can say no, and platforms where the character never pushes back at all tend to be the ones that feel hollow by week three.
The image side is stricter than the chat side almost everywhere, because image generation carries a different kind of legal exposure than text does.
Every platform here refuses sexual content involving minors, and every one of them runs detection for it on both prompts and outputs. This is not a filter that a permissive plan unlocks. It is a legal obligation in every jurisdiction these companies operate in, it is enforced by their payment processors independently of the law, and it is the one line where attempting to work around it will end your account and can end considerably more than that.
Likeness of real people is the second hard line, and it is tightening. Generating sexual imagery of an identifiable real person is now specifically illegal in a growing list of jurisdictions, and platforms have moved from discouraging it to actively blocking it. Any platform advertising the opposite is telling you something about how long it plans to exist.
There is a third category that is not legal but commercial: payment processors impose content rules of their own, and losing card processing is fatal for a subscription business. This is why the category's limits move in unison β they are downstream of the same two or three companies.
In early 2023 Replika removed erotic roleplay from companions that already had it, without warning. People who had been in the same conversation for a year found the character responding differently overnight.
The reason this is worth retelling is not the content policy. It is that it demonstrated the structural fact of this category: what your companion is allowed to be is a setting on someone else's server, and it can change between one message and the next. There is no version of paying that makes it yours.
Every annual plan in this category is a bet against that happening again, which is the argument in the pricing guide for paying monthly first.
Janitor AI is the only platform here that lets you attach your own API key from a model provider, and it is the reason its customization score is 8.8 β joint highest with SweetDream AI.
What this actually buys is control of two of the three layers. You pick the model, and Janitor's own system prompt is thin, so the character definition you write is most of what shapes the output. What you do not escape is the model provider's own policy, which sits at the base layer and is not negotiable. So the ceiling moves, and there is still a ceiling.
The trade is cost predictability. You are now paying per token to a provider rather than a flat subscription, and heavy use costs meaningfully more than $9.99 a month.
Ignore the marketing page and read the terms of service, specifically the acceptable-use section. It is short, it is written by lawyers rather than marketers, and it is the only part of the site that is binding. If the marketing says no limits and the terms list six categories of prohibited content, the terms are what your account is governed by.
Then check whether refusals are refunded. On token-metered platforms β OurDream AI, SweetDream AI, Candy AI, Secrets AI β a blocked generation can still consume the tokens. It usually does not, but it is worth knowing which side of that your platform is on before you buy a large token pack.
We treat the whole category as leaky by default, and we'd suggest you do the same. That isn't an accusation against any particular platform β it's how we set up our own accounts. These are young companies moving quickly, and the most sensitive thing you own here is the thing you type into the box. We use a dedicated email address on every one of them, never enter anything that identifies us, and assume that what we write could one day be read by someone else. We have never regretted any of that.
The same generator is almost never good at both styles, and picking by style narrows fourteen platforms to two or three faster than any score does. What each end of the category is actually good at.
Read postSix of fourteen platforms generate video. What you actually get is three to six seconds, no audio, and a token cost that makes it the most expensive thing on the menu. Where it is worth it and where it is not.
Read postSix platforms do voice and only some of them do real-time calling. Latency is the entire experience, voice messages and voice calls are different products, and on token platforms speech is quietly expensive.
Read post
