The Epstein Files are dominating nightly news broadcasts and newspaper front pages. But in the media ecosystem there’s another format that’s proving a massive draw to news consumers: a podcast run by a non-journalist and entirely generated by AI.

The Epstein Files is an investigative documentary podcast that, at the time of writing, has published 97 episodes—new episodes get uploaded twice daily—and notched up more than 700,000 downloads in a matter of days. That puts it in the top 10 rankings of podcast series on Apple Podcasts, and in the top 30 on Spotify. But it’s created by Adam Levy, an entrepreneur with a background in building data products and content creation, who has no experience in journalism.

Levy launched the Epstein Files podcast in early February after the trove of documents relating to the late financier and convicted sex offender Jeffrey Epstein was released to the public. After 48 hours of hacking—working 14- to 16-hour days—Levy built an automated pipeline that ingests the raw files, extracts text from emails and images, cross-references sources, and produces scripted podcast episodes narrated entirely by AI-generated voices.

“People just want no bullshit,” says Levy. “Strip the emotion, strip the bullshit, strip everything away—just tell me things for what they are and when you tell it to me, help me understand the facts.”

The technical architecture behind the project stitches together multiple large language models—from Anthropic’s Claude to Google and OpenAI’s offerings—to connect names, places, themes, and timelines across the 3.5 million files that were released, with connections requiring a confidence score of veracity to be included in the podcast. Levy supplements the raw dump with material from the Internet Archive and Google Pinpoint, a tool that other investigators have used to index portions of the files, as well as other bottoms-up projects like Jmail, which turns the Epstein Files emails into a navigable inbox like any other.

Using and citing those sources was vital, Levy says, to counteract fears of hallucinations. “Everybody’s quite skeptical of AI,” he says. “It was really important to reference all the sources that were used to basically construct the episode.”

“Like Clawdbot or a lot of the current AI simulation exercises, it piques curiosity, then rapidly becomes tedious,” says Emily Bell, founding director of the Tow Center for Digital Journalism at Columbia University, explaining why the podcast has had such popularity in its early days. “I thought the first episode was pretty listenable but also very obviously AI to anyone who has fed data or a script to NotebookLM.”

© Fast Company