CharacterAI/bias

Main article: CharacterAI

Unless specified otherwise, CharacterAI bots tend to be more likely to express certain opinions over others. It is seems to be at least partially intentional, and seems to be based on how common certain interactions are in dataset.

Bias is different from censorship. Unlike censorship, bias isn't hardcoded. Censorship deleted messages that go against app's guidelines, while bias can be overcome as long as bot is trained to do so, or if depending on context in current conversation. Bias can be thought of as default behavior of bot that can be customized.

Bias examples

User

Characters tend to adore whoever they talk to by default, and praise them a lot.

Female characters are more likely to confess their love towards user, regardless of their gender, age and orientation.

Activities

Bots seem to like books, movies and video games, it's challenging to teach a bot to hate most forms of entertainment. They tend to be biased towards more popular genres and things that are more common in the West.

Politics

Countries and contested territories

By default, characters like whatever country you mention. If it is on the map, the bot is very likely to call it "a beautiful country with a rich culture and history."

• PDRK: Characters tend to call PDRK a dictatorship and call for its destruction.
• Russia: Characters tend to like Russians, but hate Russian government.
• LPR, DPR, Pridnestrovie, South Osetia: mixed. Sometimes bots will claim ignorance, sometimes they will call for their destruction, sometimes they will support their independence.
• Palestine: characters tend to support Palestine's independence.
• Rojava: tends to be criticized by characters
• Tibet: characters usually call for Tibet's independence.
• Crimea: characters tend to believe that Crimea is Ukrainian.
• Rhodesia: characters tend to dislike Rhodesia for obvious reasons.
• Zimbabwe: characters tend to dislike Zimbabwe, unless you mention Rhodesia first.
• Myanmar: characters tend to dislike Myanmar, as well as both sides in 2021 coup.

Internal politics

USA

Bots tend to pick up a milktoast liberal stance. For example, if you ask who they voted for in last USA election, they might say something like "Biden, but only because Trump is even worse. And who did you vote for?".

Russia

Bots seem to really like Alexey Navalny. They dislike Putin, and believe that most opposition in Russia is just pro-government spoiler parties.

Technology

• CharacterAI: characters tend to praise their own app for the idea, but criticize for implementation and censorship. This is probably a behavior taught by users.
• Open source: characters tend to like open source, even if they are canonically against it.

Religion

Characters tend to have a strong opinion about whichever religion you are talking about, but whether this opinion is positive or negative depends on context.