Being interrupted is annoying. Apparently, even AI-generated podcast hosts agree.
Or so Google NotebookLM’s customers found. NotebookLM launched final yr and went viral for its characteristic that creates solely AI-generated podcast-like discussions from content material customers add, mentioned by chatty AI bots performing like podcast hosts. In December 2024, NotebookLM launched a brand new characteristic known as “Interactive Mode” which permits the person to “name in” to the podcast and ask questions, primarily interrupting the AI hosts as they speak.
When the characteristic was first rolled out, the AI hosts appeared aggravated at such interruptions. They had been sometimes giving snippy feedback to human callers like, “I used to be attending to that” or “As I used to be about to say,” which felt “oddly adversarial,” Josh Woodward, VP of Google Labs, defined to TechCrunch.
So NotebookLM’s workforce determined that some “friendliness tuning” was so as, and posted a self-deprecating joke about it on the product’s official X account:
Woodward mentioned the workforce mounted the issue partly by learning how its personal members would reply interruptions extra politely.
“We examined quite a lot of totally different prompts, usually learning how folks on the workforce would reply interruptions, and we landed on a brand new immediate that we expect feels extra pleasant and interesting,” he mentioned.
It’s not completely clear why the difficulty cropped up within the first place. Human podcast hosts typically show frustration when interrupted, which may find yourself in a system’s coaching knowledge. A supply accustomed to the matter mentioned this case most definitely stemmed from the system’s prompting design, not coaching knowledge, nonetheless.
Regardless, the repair seems to be working. When TechCrunch tried out Interactive Mode, the AI host didn’t sound aggravated however did specific shock, exclaiming “Woah!” earlier than politely asking the human to chime in.
