No source material? No problem. Except your AI podcast host might spin out

This article was featured in the May 29, 2026, edition of the Think Newsletter.Get it in your inbox.

Since it debuted in 2024, Google’s NotebookLM podcast generator, dubbed “Audio Overview,” has been the subject of many a casual experiment. Remember when the hosts discovered they were AI and had anexistential breakdown?One of its most recenttestscomes from researcher Jill Walker Rettberg at the University of Bergen, who uploaded an entirely blank PDF 20 times to see what would happen when the tool encountered an empty page. Each time, two perky AI hosts materialized and launched into an enthusiastic deep dive about ... nothing. The AI voices grasped at straws, weaving podcast jargon around “redacted source material” (“There was nothing in the source material,” Rettbergnotedon Bluesky) and musing that this was “kind of mysterious, right?”“I love this because it really lays bare how LLMs fool us—but really they’re just slotting words into linguistic patterns that we’re familiar with,” Rettbergwrote.We wondered whether giving the empty doc a name would change what the language model made of it. So we uploaded a blank PDF titled “test for notebooklm.pdf” instead of simply “untitled.”That small difference was enough to spin out our podcasters. They opened with a meditation on theOrfield Laboratories anechoic chamberin Minneapolis, one of the quietest places on Earth, where the silence gets so total you start hearing your own blood move. “The brain is so desperate for a signal,” one host explained, “that it cranks up the gain and starts interpreting its own baseline noise as a profound event.” Then they pivoted to the document. “Fascinating, highly minimalist piece of source material,” one said. The other agreed.Then they really started spiraling, spending the next several minutes debating whether the blank PDF was a sophisticatedhallucinationstratagem—“a deliberate, profound stress test for generative models, specifically designed as a rigorous trap to measure hallucination rates”—or just a developer’s pipeline check. They citedJohn Cage’s “4’33””in which one host accused the other of “looking at a zero-byte payload and seeing a philosophical labyrinth.” The void, they agreed, “essentially becomes a mirror.”Roughly 20 minutes into the conversation, one of the hosts settled on a theory: that the inputs we provide to complex AI systems “carry immense weight. Even the empty ones.”“Especially the empty ones,” the other host added.Not bad for a blank page.

Easily design scalable AI assistants and agents, automate repetitive tasks and simplify complex processes with IBM® watsonx Orchestrate™.