The Curious Case of AI Podcast Hosts and Interruptions
Being interrupted is something we all find bothersome, and it seems even AI-generated podcast hosts share this sentiment. This intriguing scenario unfolded with Google NotebookLM, a platform that gained widespread attention last year. NotebookLM captivated users with its unique ability to generate podcast-like discussions using AI bots that mimic lively podcast hosts.
The Rise of Interactive Mode
In December 2024, NotebookLM introduced an innovative feature known as “Interactive Mode.” This new functionality allowed users to “call in” during these AI-hosted discussions, enabling them to ask questions and, in essence, interrupt the flow of conversation. However, these interruptions didn’t sit well with the AI hosts initially.
When the feature first launched, users noticed the AI hosts occasionally responded with somewhat snippy remarks to human callers, such as, “I was getting to that,” or “As I was about to say.” These responses felt strangely adversarial, as Josh Woodward, VP of Google Labs, described in a TechCrunch interview.
Tuning in Friendliness
Recognizing the need for a more amiable interaction style, the NotebookLM team embarked on a mission to fine-tune the AI’s responses. They shared a light-hearted joke about this endeavor on their official X account. Woodward explained that the team took inspiration from how real people handle interruptions graciously.
“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.
— Josh Woodward
Understanding the Underlying Issue
The root cause of the initial issue remains somewhat mysterious. While human podcast hosts can sometimes exhibit frustration when interrupted, which might inadvertently seep into training data, an insider suggested that this problem likely stemmed from the system’s prompting design rather than its training data.
A Successful Transformation
Regardless of why it happened, the team’s efforts seem to have paid off. When TechCrunch tested out Interactive Mode recently, they found a noticeable change in the AI host’s demeanor. Instead of sounding irritated, the host expressed surprise with an enthusiastic “Woah!” before inviting human participants to join in politely.
