The New York Times Built an AI to Read 3 Million Epstein Files

In January 2026, the U.S. Department of Justice released a huge mixed-media collection about Jeffrey Epstein — roughly three million pages of documents, far more than any human team could read. The New York Times responded by building what its researchers call the Epstein Files Engine: an AI tool that acts less like a writer and more like a very fast research assistant.

Here's how it worked. A reporter typed a question in everyday language. The AI translated that question into the kind of precise database query a programmer would normally write, then searched three separate collections at once: the Epstein releases, the Times's own archive, and outside news headlines about Epstein. It came back with answers that included citations — links to the exact documents — so reporters could verify everything themselves rather than trusting the machine.

The results were substantial. More than 100 journalists used the tool, and it contributed to at least 20 published stories. The team also built a feature called Diff that spots duplicated text and repeated images, which helped the AI flag what was genuinely new instead of resurfacing the same familiar material. That matters because in a giant document dump, the real value is finding the handful of details that nobody has noticed yet.

The paper's central argument is the most interesting part for anyone worried about AI replacing journalists. The researchers say newsroom AI works best not as an autonomous writer churning out articles, but as an interface to source material and institutional knowledge — a way for humans to reach information faster and then apply their own judgment. The catch: this approach depends on well-organized, searchable data and careful human verification. Point it at messy or incomplete records, and the citations are only as trustworthy as the documents behind them.