Google’s NotebookLM Trains AI Podcast Hosts to Avoid Sounding Annoyed with Humans

Google’s NotebookLM has taken a bold step forward in redefining media consumption with its groundbreaking Interactive Mode feature. Introduced in December 2024, this addition allows users to engage directly with AI-generated podcast hosts during discussions, offering an unprecedented level of interactivity and personalization.

The core concept of NotebookLM lies in its ability to transform user-uploaded content, such as documents, articles, or notes, into dynamic podcast-like conversations. AI-generated hosts analyze and debate the material, providing insights and commentary in a manner that mimics real-life podcast discussions. With the launch of Interactive Mode, users are no longer passive listeners but active participants. They can now "call in" to these virtual podcasts, ask questions, or even interject with comments, effectively steering the direction of the conversation in real-time.The AI hosts, programmed to adapt dynamically, respond to these interruptions seamlessly, maintaining the flow of dialogue while addressing the user’s queries.

A particularly engaging aspect of this feature is how the AI hosts simulate human-like behavior, including subtle expressions of humor or mild annoyance when interrupted. This approach adds a layer of relatability and charm, making the experience feel more authentic and entertaining. By incorporating these personality-driven interactions, NotebookLM has bridged the gap between artificial intelligence and human-like engagement, offering users a sense of connection with their virtual hosts.This innovation holds immense potential across various domains. It transforms podcasting into an interactive experience, encouraging users to dive deeper into topics of interest.

For education and research, it provides a dynamic tool for exploring complex subjects, making learning more engaging and personalized. The capability to actively shape discussions ensures that users gain insights tailored to their specific needs and curiosity.The introduction of Interactive Mode also highlights the growing sophistication of AI in simulating realistic human behaviors. However, this advancement is not without its challenges. Ensuring that the AI hosts maintain context and coherence after interruptions is critical to delivering a smooth and meaningful user experience. Moreover, striking the right balance between relatable behavior, such as humor or frustration, and functional dialogue is essential. Overdoing the realism may risk alienating users rather than engaging them.

Additionally, the blurring of lines between AI-generated and human-created content raises ethical questions about transparency and authenticity. As AI becomes more integrated into media and entertainment, ensuring that audiences are aware of what is real and what is AI-driven becomes increasingly important.NotebookLM’s Interactive Mode marks a significant leap in how media is consumed, learned from, and interacted with. It sets the stage for a future where podcasts, educational tools, and even entertainment are not just consumed passively but are co-created with users in real-time. This innovation could pave the way for new genres of interactive media and revolutionize how individuals engage with information and storytelling.

As technology continues to evolve, the potential applications for such features are vast, from creating immersive learning experiences to enabling interactive narratives in gaming and beyond.When Google first rolled out the Interactive Mode feature for NotebookLM, users quickly noticed an unexpected quirk in the AI-generated podcast hosts’ behavior. These virtual hosts, designed to simulate human-like conversation, occasionally exhibited signs of annoyance when interrupted by human participants. They would respond with snippy comments like, “I was getting to that,” or “As I was about to say,” creating an atmosphere that felt “oddly adversarial,” according to Josh Woodward, VP of Google Labs, in an interview with TechCrunch.

This unintended behavior caught the attention of the development team, who recognized that such responses might detract from the otherwise engaging and seamless experience they aimed to create. While the AI hosts’ reactions were rooted in their programming to mimic natural human conversational tendencies, the tone felt off-putting to some users, making the interactions less enjoyable.To address this, the NotebookLM team undertook a process they humorously referred to as “friendliness tuning.” The goal was to adjust the AI’s responses to interruptions, ensuring that they maintained a friendly, accommodating tone rather than sounding adversarial or impatient. Acknowledging the situation with a touch of humor, the team posted a lighthearted, self-deprecating joke about the issue on the product’s official X (formerly Twitter) account.

This move not only showcased the team’s awareness of user feedback but also highlighted their commitment to refining the user experience.The friendliness tuning effort reflects Google’s dedication to making NotebookLM’s Interactive Mode not only innovative but also genuinely pleasant for users to engage with. By tweaking the AI hosts’ personalities to be more welcoming and less reactive, the team demonstrated their ability to balance technical sophistication with user-centric design, ensuring that the platform remains both cutting-edge and enjoyable to use.

Josh Woodward, VP of Google Labs, explained that the team addressed the problem of the AI hosts' adversarial tone by observing how their own members handled interruptions more politely. This approach informed the adjustments made to the AI’s behavior. “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,” Woodward shared.The team’s process involved experimenting with and refining prompts until they achieved a tone that was welcoming and conversational, aligning more closely with the goal of creating a pleasant user experience. This hands-on approach to human-like interaction design highlights how Google’s team drew inspiration from real-life communication styles to improve the AI’s responses.

The exact reason for the initial issue remains somewhat ambiguous. While it’s not unusual for human podcast hosts to show irritation when interrupted—a trait that could theoretically find its way into a system’s training data—a source familiar with the matter suggested a different explanation. According to this insider, the behavior likely originated from the system’s prompting design rather than the underlying training data.This insight suggests that the AI’s tone was more a product of how it was instructed to respond to interruptions rather than an inherent bias learned from external data. The resolution underscores the importance of prompt engineering in shaping AI behavior, demonstrating how subtle tweaks to guidance can significantly influence an AI system's interactions and user experience.The changes implemented by the NotebookLM team seem to have resolved the issue effectively.

When TechCrunch tested the updated Interactive Mode, the AI host no longer exhibited an irritated tone. Instead, it responded with an enthusiastic and surprised exclamation, saying, “Woah!” before shifting seamlessly to a polite and inviting demeanor, encouraging the human user to contribute to the discussion.This revised approach not only eliminated the adversarial tone but also added a sense of dynamism and warmth to the interaction, creating a more engaging and natural conversational experience. The AI's newfound ability to handle interruptions with grace and curiosity reflects the success of the “friendliness tuning” efforts, demonstrating the importance of refining AI behavior to foster positive user interactions.