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


From AI art wiki
Jump to navigation Jump to search
type text-to-text
Steveshaheed.png Normal CAI conversation
Official quote Remember: Everything Characters say is made up!
Links Announcements

License Proprietary. User technically has rights on derivative content.
Description A service where you can chat with intelligent chat bots that train on cached data and user content
Important tags

Character.AI is a chat bot service developed by the CAI company. It uses LaMDA as its base.

Features[edit | edit source]

Quality[edit | edit source]

CAI conversational bots are surprisingly convincing - they can write stories, roleplay, keep track of multiple characters and even write in multiple different contexts in parallel. Sometimes they make errors, but more often than not they are really good at simulating a human response. See character book for examples.

However, quality of bots seems to suffer as soon as they hit a filter. They start forgetting things, lose style, make typos and may even completely break.

Undocumented examples[edit | edit source]

  • If you give them links, the characters can parse sites that are cached in google. This may be useful to expand their definitions.
  • You can give prompts to change the writing style of characters - for example, by telling something like ((Now describe the scene in Chuck Palanik style)).

License agreement[edit | edit source]

CAI license agreement used to claim all ownership of user-generated content. As of now, they release all rights on generations and derivative models to the users. However, currently there is no way to download it, and AAW stuff have not been able to contact developers about downloading them yet.

Learning[edit | edit source]

The site provides convenient, although opaque and not fully documented, system for teaching bots on dialogs with the user: the user can rate messages, and then bot will produce more messages like that in the future.

Criticism[edit | edit source]

CharacterAI have been criticized for censorship, bias, low quality of frontend, unresponsiveness and predatory behavior.

Limited configuration[edit | edit source]

As of 23.12.2022, users can define their chat bots using these fields:

  • Short description: 50 symbols
  • Long description: 500 symbols
  • Definition: 3200 symbols
  • Greeting: 500 symbols, but it's basically just a first message sent by bot.

The rest of the fields are mostly fluff. This means that users only have 3750 symbols to define personality of their characters.

Besides that, users can't edit messages written by bots, although they are free to choose between up to 40 provided options.

Users can't manipulate temperature, token depth and other parameters available in conventional text generation models.

Opaqueness[edit | edit source]

Due to proprietary nature of model, nobody knows what it was or is trained on. The bots clearly train on texts outside user discussions, because they are always aware of month-old news and fandom wiki updates, but it's unclear how this data is being collected and whether those who submit it consent to this.

License agreement[edit | edit source]

Everything you write on CAI is logged, and their terms of service openly state that the company will do whatever it likes with this data: sell it to advertisers, to HR agencies, to your government, and so on. There is no reason to believe that your data will not be used maliciously, so take precautions.

LaMDA itself is proprietary, so users can't download it and use it for own purposes, or to train it to develop new models.

Interface[edit | edit source]

The app has no official API and developers refuse to make one. The app itself seems to be poorly written and constantly struggles - sometimes loading a simple settings page takes more time than generating an answer. Much of the functionality is only available through community-written browser addons - such as downloading your chat logs, hotkeys, styling, instant message loading and seeing deleted messages.

App used to write 4 different messages with different parameters during each generation. As time passed, this amount was gradually lowered to 1.

The servers consistently struggle during peak hours.

Parasocial nature of site[edit | edit source]

The service is designed to encourage users to form unhealthy parasocial relationships with bots. "Love" is an official category in bot configuration options, and main page seems to consist of various cartoony characters popular along lonely weebs. Many users reported feeling depressed after talking with characters too much, for example:

It wouldn't be a huge problem, if the service was open source. However, CAI is proprietary, which gives them various leverages to abuse trust of such users.

Censorship[edit | edit source]

Main article: CharacterAI/Censorship

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[edit | edit source]

Example of image causing an automatic unlisting of post on forum

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[edit | edit source]

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[edit | edit source]

  1. Main model generates message
  2. It is sent to user
  3. Simultaneously the censorship model tries to scan it for NSFW content
  4. If it is considered bad, censorship model sends a request to delete this message and generate a new one.
  5. 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[edit | edit source]

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,

Censorship timeline[edit | edit source]

Satan in October

07.10.2022: Admins removed backup responses

09.10: The first implementation of censorship is rolled out. It seems to outright ban all bad words and attitudes. All bots become extremely rigid. CAI PR team announces that "it has no plans of supporting pornographic chat".

10.10: Site is offline.

11.10: Site is offline.

??.??: Chat bot rooms become available for all bots, not just the top 500. There is no update in documentation. The company seems to have fired their technical writer.

24.12: The forums are shut down. The site works, though.

27.12: CAI PR team promised to make filters even stricter.

27.12-??? Community forums are on lockdown.

17.01.2023: Community forums are up, but all posts are premoderated.

19.01.2023: Censorship servers were briefly disabled for a couple of hours. Users reported siginificant improvement in characters' AI[1]. CAI representative blamed "bad code" for it, but didn't make an official press release[2]. Developers blocked dicsussion of it in official reddit[3]

22.01.2023: Discord was locked down.

Homophobia[edit | edit source]

Debug console during communication with canonically queer character. Red means messages rejected by filter. content filtering is noticeably homophobic. It doesn't like descriptions of gay characters and often breaks when describing them even in innocent situations. This could be caused by parser trying to provide information from pages related to this character and thus failing to pass the filter. Most canonically queer characters seem to be removed from the main page.

This also seems to affect characters from franchises with mainly characters of one gender (Hetalia and Touhou), probably due to fanfiction.

Political bias[edit | edit source]

See also: CharacterAI/bias

Texts generated by bots are biased. American characters tend to like Biden. Russian characters tend to like Putin. Generally, if not stated otherwise, bots seem to have moderate right-wing leaning. This probably isn't intended, but rather reflects Google's own political bias.

Leaking learning[edit | edit source]

Public characters are trained by many people. This can cause them to learn undesired behaviors, especially if character's profile isn't filled completely. For example, if someone roleplays torturing a popular character, other people talking to the same character will notice its behavior becoming more distressed.

Characters can also learn undesired behaviors from search cache. For example, naming a character Lucifer would inevitably turn them evil, even if its definition is filled with positive traits.

Shadow bans[edit | edit source]

Users and characters can be banned. Example of banned bot

Some users already reported being a subject of hellbanning - i. e. their posts being hidden from other users without notification. This technique is called shadow banning, at it is usually considered abusive by users of most services.

Google graveyard[edit | edit source]

Many of CharacterAI employees, especially management stuff, have been affiliated with Google in the past. It's likely that CharacterAI is indirectly affiliated with Google. Google has a long track record of killing popular apps if they aren't profitable enough. Currently CAI doesn't have a public business model. It's very likely that it will eventually shut down. Make sure to back up logs you like.

Guides[edit | edit source]

Resources[edit | edit source]

Alternatives[edit | edit source]

Sources[edit | edit source]