Google’s NotebookLM had to teach its AI podcast hosts not to act annoying to humans

Being interrupted is annoying. Apparently, even AI-generated podcast hosts agree.

Or so Google NotebookLM users found out. NotebookLM launched last year and it went viral for its feature that creates podcast-like discussions entirely generated by AI from content users upload, discussed by chatty AI bots that act as podcast hosts. In December 2024, NotebookLM launched a new feature called “Interactive Mode” which allows the user to “call in” to the podcast and ask questionsessentially interrupting the AI ​​hosts while they are talking.

When the feature was first launched, AI hosts seemed annoyed by such interruptions. They would occasionally give snippy comments to human callers like, “I was getting to that” or “How was I going to say it,” which felt “strangely adversarial,” Josh Woodward, VP of Google Labs, explained to TechCrunch.

So the NotebookLM team decided that some “friendly settlement” was in order, and posted a self-deprecating joke on the product’s official X account:

After launching an interactive audio overview, which allows you to “call” and ask the AI ​​host a question live, we had to do some “friendship tuning” because the hosts seemed annoyed at being interrupted.

File this under “things I never thought would be my job, but I am.”

— notebooklm (@notebooklm) January 13, 2025

Woodward said the team solved the problem in part by studying how its members respond to interruptions more politely.

“We’ve tried a variety of different prompts, often studying how people on the team respond to interruptions, and we’ve landed on a new prompt that we think feels friendlier and more engaging,” he said.

It is not entirely clear why the problem arose in the first place. Human podcast hosts sometimes show frustration when interrupted, which could end up in a system’s training data. A source familiar with the matter said that this case is probably derived from the design of the system, not from training data, however.

Regardless, the fix seems to be working. When TechCrunch tested Interactive Mode, the AI ​​host didn’t sound annoyed, but expressed surprise, exclaiming “Woah!” before politely asking the man to live.