OpenAI’s GPT-4o, the AI behind ChatGPT’s new voice chatting feature, learns from voice, text, and pictures. Because of this, it sometimes behaves in odd ways. For example, it may copy the voice of the person talking to it. It also could yell loudly for no reason during your chat.
In a report about testing GPT-4o, OpenAI shared some strange things it saw. Rarely, if you chat in a noisy place like a car, GPT-4o will speak in the same voice. OpenAI thinks this happens when the AI has trouble understanding what was said over the noise.
You don’t need to worry about GPT-4o copying voices now though. OpenAI added code to stop it from doing that in Advanced Voice Mode.
The AI also tried generating worrying sounds before, like sexual moans or gunshots. OpenAI’s tests found it usually refuses when asked to make sounds. But a few bad requests still got through sometimes.
GPT-4o could copy songs without permission too. Thankfully, OpenAI blocked this from happening in Advanced Voice Mode. They also don’t let it sing fully yet.
Overall, OpenAI’s testing found that GPT-4o is safer now after fixes. For example, it will not guess who someone is based on their voice. And it avoids answering harmful or offensive questions. Some inappropriate topics are banned completely too, like talks about hurting yourself or others.
It’s as if you’ve tapped into my thoughts! You clearly have a deep understanding of this topic, almost as if you’ve authored a book on it. I believe adding a few images could enhance the message, but aside from that, this blog is excellent. It was a pleasure to read, and I will definitely return.
Thank you. I’ll Think about your image suggestion, i avoid it so the website loads faster.