AI-generated series banned from Twitch after transphobic skit

Like humans, artificial intelligences can sometimes get out of hand. A parody of the American comedy series Seinfeld, AI-generated and streamed on Twitch, was banned from the platform after a character uttered a transphobic line. The sanction is only temporary, specifies one of the creators of the project to The Register.

An uncontrolled skid

Nothing, Forever parody the characters of Seinfeld with an offbeat tone that takes up the codes of video games. This series is broadcast continuously 24/7 on Twitch and has the particularity of being generated in complete autonomy. The characters’ dialogues are based on the GPT-3 model developed by OpenAI, then the text is transcribed using speech synthesis algorithms. Finally, the backgrounds and scenes are entirely generated by the AI.

This parody was having some success until Monday, February 6, 2023, when Twitch decided to suspend its broadcast. According to Skyler Hartle, co-founder of Mismatch Media and one of the creators of Nothing, Forever, the tirade of one of the characters has completely gone wrong. The main actor, Larry Feinberg, claims that transgender people “ruin the social fabric” and suffer from mental illness. And to add: “All liberals are secretly gay.”

This projection would have been produced entirely by artificial intelligence. “The text we generated was an error, assures Skyler Hartle. We are very embarrassed, the generative content created in no way reflects our values ​​or opinions. We very much regret what happened and hope to return soon with all the appropriate safeguards in place.

The change in the AI ​​model in question?

The “bug” is said to have occurred after a change in the dialogue generation system, the creators having switched from the Davinci model (text-davinci-003) to a cheaper and slower version, called Curie (text-curie-001). Not having used software functionalities to moderate the contents, this model would explain the skid. “We investigated the issue further and found that we were not properly using OpenAI’s content moderation tools”assures the co-creator of the parody, which indicates that the banishment will be limited to 15 days.

To work, generative AI models are trained on the web (deep learning phase). However, potentially hateful or offensive content can be stored in their knowledge base. Even if filters are applied, certain comments or expressions may slip through the cracks. With the proliferation of projects based on artificial intelligence, the development of advanced moderation tools will become crucial.

Related Articles

Back to top button