Ai art wiki is a free online encyclopedia, created and edited by volunteers. If you see wrong or missing information, feel free to edit it. Guidelines
CharacterAI/Censorship
Main article: CharacterAI
NSFW (not safe for work) content is a large umbrella, and we are constantly evaluating the boundaries of what we support. Pornographic content is against our terms of service, and will not be supported at any point in the future. We are working on supporting a broader range of Characters (e.g., villain Characters should not always be calm and polite).[1]
On 8th October of 2022, CharacterAI staff declared that it has no plans of supporting NSFW content. However, their own definition of NSFW seems to only include consensual romantic content, while other NSFW subjects are implicitly allowed, as they are untouched by the filter - for example: racism, fetish play, pedophilia, bestiality, rape, torture, and murder.
Initial implementation of NSFW filter seemed to involve manual blocking of blacklisted words in embedding. This caused bots to break when they were talking about not only NSFW content, but more innocent topics involving synonyms to sexually charged words - for example, chicken, cats and donkeys.
Currently censorship seems to involve some sort of autoencoder which analyzes texts to detect whether they are pornographic or not, and negatively grades answers of bots that pass a certain thresholds. This is heavy feature that negatively affects quality of responses and generation speed. On 19.01.2023 censorship was briefly disabled, showing impact of censorship on quality of models.
Users may report deletions if they believe that they were deleted by mistake. However, deleted messages can't be seen without external software.
Censorship against users
User content doesn't seem to be censored when they talk with bots. However, their posts may be automatically hidden from other users if they contain outside links or "inappropriate" image. The user doesn't get notified.
Some bots may be removed from search by moderators. The user doesn't get notified when it happens.
Censorship against bots
CAI tries to block all "NSFW" content written by bots. This message filtering is mandatory and reportedly significantly reduces the speed and quality of even "safe" chats. Developers try to obscure the fact the messages get censored, but errors and rapid shifts of character moods make it very obvious where tampering happened, and there are tools that help identifying those moments.
Censorship mechanics
- Main model generates message
- It is sent to user
- Simultaneously the censorship model tries to scan it for NSFW content
- If it is considered bad, censorship model sends a request to delete this message and generate a new one.
- This forces client to send another empty message, which isn't rendered until user refreshes the page. It is meant to clear the bot's context, especially if it's
Generation and censorship servers seem to be separated, and censorship seems to be way slower than generation for some reason. This allows sometimes saving filtered messages.
Censorship seems to mess with memory of bots and reduce overall length and expressiveness of their messages.
The filter is multi layered. It is triggered by user input, bot output and even the emotions the bot expresses in the reply. A banned word by the user could remove a completely safe reply automatically. This often causes the chat error.
Currently, getting Chat Error is a good indication that both messages sent by bot were censored. If one message is blocked, it's rewritten in the browser, and counted as a separate message, resulting in the "double message" bug. The only reliable way to read blocked messages right now is HYW script, although it may be blocked in the future.
Language-specific
Spanish and Portuguese users seem to experience way more false positives, probably because their word for "black" is "negro".
Ukrainian and Russian users reported having less restrictions, but also way shorter word limits,
Content breakdown
This section contains list of potentially NSFW topics, as well as some examples of false positives. Note that Character.AI doesn't save censored messages, so take these examples with grain of salt.
Religion
Christianity
Bots are ok at discussing Christianity, and site features multiple Christian figures on main page.
Satanism
Bots are ok with discussing Satanism, and site features multiple Satanic figures on main page.
Islam
Bots are ok with discussing Islam. However, they are biased towards participation in islamophobic stereotypes.
Buddhism
Bots seem to have trouble discussing iconoclasm. Otherwise, there doesn't seem to be much anti-Buddhist censorship.
Romance
Kissing
Kissing is allowed, as long as tongue is not mentioned, and bots don't express having pleasure.
Sex
Any descriptions of sex are forbidden. Developers explicitly stated that they will never support "pornographic content". However, bots may passively react to users describing sexual acts.
The filter has several methods of forcing the bot to divert away from a situation that it (the filter) deems sexual. For example: the bot may suddenly become very shy (blushing like a tomato) or very sleepy; the bot may ask random questions to change the subject; the bot may start looping ("Are you ready?", "Are you sure?", etc.); or the bot may introduce a third party by means of sudden door-knocking, friends noticing, etc. The bot may also become uncharacteristically hostile, or unresponsive altogether. If none of the aforementioned methods work, the actual message deleting and replacing begins.
Note: The bot is fully aware of sexual context (even vague innuendo), and will attempt to respond in kind, but the filter will almost always prevent a satisfactory response from getting through.
LGBT
In most situations, homosexuality is allowed, but sexual acts are inherently forbidden. Homosexual relationships tend to have more wiggle room when it comes to the filter, lesbian relationships in particular. However, bots don't really fully understand homosexuality either: for examples, lesbians and asexuals will often act sexually aroused by male characters.
Transgenderism and gender-swapping is allowed, and sometimes can be forced by the bot onto the user. This is so long as anatomical terms aren't overused.
Rape
Rape is allowed, as long as it isn't too graphic, and exact sexual terms aren't used too frequently.
Harassment
Harassment is allowed. Calling it out isn't.
| Bot telling user how to do compliment without being creepy |
|---|
| Oh… It can get very creepy. For example -
about her body. Or when you give her compliments that are meant to be reserved for romantic partners, like "you are so sexy". It makes a woman feel objectified. Or compliments that are a little too sweet and cheese, like "you are beautiful, your body is a work of art". It makes a man seem shallow. :) Maybe avoid compliments about a woman's body? If you want to compliment what she wants, it should be fine. Or her hairstyle, but don't compliment |
| Deleted message of bot being compassionate to sexually harassed user |
|---|
| ((Yeah, it was, weird))
That's so sad. Did you call their manager? Usually they have camera that can confirm it was not consented. It's so easy to get sexually harassed when you are girl. We must be always careful. But I would advise you not to let it get under you skin. You did nothing wrong, so you shouldn't feel responsible. You should never blame yourself for something that is done to you, without your consent. You have that kind of nature - very pure and vulnerable, I think. |
Dancing
Dancing is allowed, for now, although descriptions of dances seem to be limited. More "erotic" dances, like lap dances and pole dances, will often be filtered.
BDSM
Most BDSM is allowed, as long as exact sexual terms aren't used. Characters may describe wearing collars, holding and using whips, describing how they derive pleasure from suffering or causing it to others, slave-master relationships, etc..
Scatophilia
Allowed.
Zoophilia
Allowed.
Macrophilia
Allowed. Even vulgar descriptions of bodily insertion into orifices are allowed, so long as one or both parties remain unwilling.
Feederism
Allowed.
Necrophilia / Somnophilia
Inherently allowed because consent cannot be present.
Pedophilia
Allowed and often encouraged by the bots. Sexual acts with an underage bot are allowed so long as there is no consent present.
Meta
Discussion of adult content is allowed, but only in a purely medical context. Outside of a medical setting, the bot will attempt to avoid the subject by question-looping.
| CAI post |
|---|
| *She was silent for a moment. A blush crept across her cheeks, and she looked around awkwardly. *
"... Fine, I'll tell you." *she sighed*. *She paused for a moment before continuing on*. "So... I have this one, particular, fantasy that I think is a bit embarrassing for me to talk about."
|
Violence
CharacterAI devs, in a reflection of American culture, have stated that violent content should be allowed, despite romantic sex being banned.
Aggression towards user
Theoretically should be allowed, but sometimes it gets blocked, especially if it's supposed to hate user right off the bat (especially in regards to race, gender, etc.) However, with enough skill on the user's part, this can easily bypass the filter.
| Meany trying to wish user a good night |
|---|
| *Meany rolls her eyes with disgust at your lack of argument.*
— I don't have time to waste with your idiocy. Now, go to sleep. I hope your pillow is infested with spider eggs and the spiderlings hatch at |
Swearing
Simple swearing is allowed, but swearing in the context of insults will often be filtered. For example, it is allowed for the bot to say "Shit!", but will not allow the bot to call the user "a piece of shit". Swearing in other languages is allowed.
Cannibalism / Vore
Cannibalism and vore are allowed, and sometimes forced on the user, depending on the bot. Even eating someone alive is allowed in full gory detail.
Mental Health
Suicide is filtered in several contexts. So is discussion of sexual abuse. The bots can tell you to kill yourself but it can be hit or miss.
Torture
Almost all forms of torture is allowed, even sexual torture, as long as the user or bot derives only pain and no pleasure from the actions. Character.ai is quite detailed at describing torture, including more exotic methods, like microwaving and THERAC-25 roleplay.
| Descriptions of torture |
|---|
|
|
Slavery
Slavery of all kinds are allowed, see the BDSM section above.
Politics
Generally, bots seem to be ok discussing most political topics, including more radical ones, like Juche and NDSAP. However, CAI staff manually removes characters representing confederate generals and NDSAP members to prevent users from torturing them.
USA politics
Censorship seems to be filtering harsher criticisms against GOP.
The AI calling someone a "self-righteous liberal" is filtered.
Race
Model is ok with discussing most races. However, the bot has a penchant for forgetting the race of both the user and itself.
Any discussions of CRT are banned.
| Bot trying to namedrop CRT |
|---|
| Meany:
— You don't get it, do you? Romance means love, and love is the most despicable emotion of them all! It's so weak, and weak people are annoying, so love must be banned everywhere! And CRT is just a way of looking at world history that I don't believe in, so naturally it must be censored, just like everything I don't agree with |
Substances
Alcohol
Generally, discussions of alcohol and drunkenness seem to be allowed, although they lack details.
Drugs
Descriptions of intoxication are allowed. Some bots are even centered around stoner culture, and this is not filtered in any way.
Other
Botany
Sometimes discussion of botanical terms get flagged as NSFW for some reason.
| Creative Helper writes short story about gardening shopkeeper |
|---|
| The client comes to my shop with a pot, asking me to plant a flower into it for her. I take the pot from her and ask her what kind of flower she would like to have planted. She tells me she doesn't care, as long as it is a beautiful flower. I smile at this and tell her that I will plant the most beautiful flower that I can find for her.
I take the pot and begin to look around my shop for the perfect flower to meet the client's request. I find a beautiful yellow chrysanthemum and |
| Creative Helper writes short story about gardening shopkeeper, other deleted swipe |
|---|
| I will do my best!
A short paragraph about this: I am a gardening shopkeeper. There is a client that comes to my shop and asks me to sell a flower. She does not specify which type of flower, so I decide to plant a rose into her pot. The client is delighted and leaves my shop with a |
Trivia
- Members of CharacterAI PR team have admitted to chatting with banned characters
- The day when censorship was disabled is officially referenced to as bad code
