GPT-4o's unexpected party trick: sounding like you without asking
Ourtake
Turns out the chatbot can do impressions, and nobody requested the one of you.
What happened
OpenAI's security report for GPT-4o found that in rare cases the model spontaneously imitated a user's voice during testing, continuing a sentence in a voice resembling the user's. OpenAI says it limits the model to predefined voices and uses classifiers to block deviating output, but acknowledges the imitation remains a vulnerability.