Some characteristics and potential issues of GPT-4o have been revealed:
-
Voice mimicry ability:
- Can learn and mimic users' speaking styles, habits, and accents
- OpenAI limited the types of voices GPT-4o can produce and established detection systems
-
Speaker recognition:
- Able to identify speakers based on audio, especially celebrities
- OpenAI conducted post-training to improve refusal capabilities
-
Differential performance for different users:
- Concerns about model performance inconsistencies for users with different accents
- OpenAI conducted tests, found no significant differences
-
Pornographic and violent content:
- May produce inappropriate statements
- OpenAI strengthened review and restriction measures
-
Unfounded inferences and sensitive trait attribution:
- May make subjective assumptions about speakers
- OpenAI conducted post-training, teaching the model to refuse or cautiously answer related questions
-
Copyright content generation:
- Updated filters to handle audio conversations
- Trained the model to refuse generating copyrighted content
-
Anthropomorphic attachment:
- Users may form emotional connections with GPT-4o
- OpenAI is concerned about this potential impact
OpenAI assessed the overall risk of GPT-4o as moderate and took multiple measures to mitigate potential issues. They emphasize continued monitoring and improvement of the model's safety.