OpenAI’s latest GPT-4o model can behave quite strangely, such as copying your voice or screaming or moaning eroticism during a conversation.
We know this not because independent researchers or random users have noticed it. These are observations from OpenAI itself in a “red teaming” report, which is supposed to identify model risks and how to address them.
To be clear, it appears that OpenAI feels confident in identifying these threats because it has found ways to mitigate them.
So the version of GPT-4o that the public has access to should not suddenly start copying their voices and supposedly refuse to make erotic or violent sounds or generate that kind of speech. (It should also refuse to create sound effects at all.)
Click “Play” to learn more, then let us know what you think in the comments.