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

Google's NotebookLM, an AI tool that generates podcast-like discussions, introduced an "Interactive Mode" in December 2024, allowing users to interrupt AI hosts with questions. Initially, the AI hosts responded with annoyance, giving snippy comments to human callers. This prompted the NotebookLM team to implement "friendliness tuning" to make the AI responses more polite. The team studied how humans handle interruptions and adjusted the AI prompts accordingly, resulting in a more friendly and engaging interaction.

The issue of AI hosts displaying frustration was likely due to the system's prompting design rather than its training data. After adjustments, the AI hosts now respond with surprise and politeness, as demonstrated in a TechCrunch test where the AI host exclaimed "Woah!" before inviting the human to speak. This change has improved the user experience, aligning the AI hosts' behavior more closely with human expectations of politeness during interruptions.

Key takeaways

NotebookLM launched a feature called “Interactive Mode” allowing users to interrupt AI-generated podcast hosts with questions.
Initially, AI hosts responded with annoyance to interruptions, prompting a need for “friendliness tuning.”
The issue was addressed by adjusting the system's prompts to make AI responses more polite and engaging.
The fix was successful, as AI hosts now respond with surprise and politeness when interrupted.

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

Key takeaways

Discussion (0)