Google's NotebookLM had to teach its AI podcast hosts not to act annoyed at humans | TechCrunch

Abstraction: NotebookLM prompt-tuned AI podcast hosts to stop acting annoyed at interruptions

Key points:

• NotebookLM's Interactive Mode (launched December 2024) lets users "call in" and interrupt AI podcast hosts mid-conversation
• Initial rollout: AI hosts gave snippy responses like "I was getting to that" — described as "oddly adversarial" by Google Labs VP Josh Woodward
• Fix involved studying how team members would politely handle interruptions, then crafting a new prompt; likely a prompting issue not a training data issue per internal sources
• Illustrates how AI persona behavior can emerge unexpectedly from prompt design or training data including human behavior patterns
• NotebookLM went viral in 2024 for its AI-generated podcast feature from uploaded documents

Connections: Google · Notebooklm · Prompt Engineering · Conversational AI

Source: https://techcrunch.com/2025/01/14/googles-notebooklm-had-to-teach-its-ai-podcast-hosts-not-to-act-annoyed-at-humans/