Google fixes its AI’s snippy podcast hosts

NotebookLM had to retrain its AI podcast hosts after users found them giving rude responses to interruptions. The fix now makes the AI sound more human and polite.

Microphone

Google’s NotebookLM, an AI-driven podcast platform, recently faced a surprising issue: its virtual hosts sounded irritated when users interrupted them. The problem emerged after the launch of an interactive feature allowing users to “call in” and ask questions during AI-generated discussions. Users reported that the hosts would sometimes make snippy comments like ‘I was getting to that,’ creating an oddly confrontational experience.

To address the issue, Google’s team implemented a ‘friendliness tuning’ process. They revised the prompts used to guide the AI hosts, studying how human speakers handle interruptions more politely. The adjustment was announced with a light-hearted post on the platform’s social media.

The change seems to be working. In tests, the AI hosts now react to interruptions with polite curiosity rather than frustration. NotebookLM’s refined approach highlights the importance of human-like interaction in AI products, ensuring users feel more engaged and welcomed in conversations.