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

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

Or so Google NotebookLM’s users discovered. NotebookLM launched last year and went viral for its feature that creates entirely AI-generated podcast-like discussions from content users upload, discussed by chatty AI bots acting like 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 questions, essentially interrupting the AI hosts as they talk.

When the feature was first rolled out, the AI hosts seemed annoyed at such interruptions. They were occasionally giving snippy comments to human callers like, “I was getting to that” or “As I was about to say,” which felt “oddly adversarial,” Josh Woodward, VP of Google Labs, explained to TechCrunch.

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

Woodward said the team fixed the problem partly by studying how its own members would answer interruptions more politely.

“We tested a variety of different prompts, often studying how people on the team would answer interruptions, and we landed on a new prompt that we think feels more friendly and engaging,” he said. 

It’s not totally clear why the issue cropped up in the first place. Human podcast hosts sometimes display frustration when interrupted, which could end up in a system’s training data. A source familiar with the matter said this case most likely stemmed from the system’s prompting design, not training data, however.

Regardless, the fix appears to be working. When TechCrunch tried out Interactive Mode, the AI host did not sound annoyed but did express surprise, exclaiming “Woah!” before politely asking the human to chime in.