Reading the Files:
AI is now the interface

Three and a half million pages, 180,000 images, 2,000 videos, 246 gigabytes. No person can read that. The question of who gets to know what is in these files is therefore a question about software.

Into the gap stepped journalists, engineers and volunteers. The New York Times built its own tooling — scraping DOJ results into spreadsheets, semantic search, AI tagging, text extraction from images, audio and video, and a system to scan all three million pages. Its own account is careful: AI “cannot determine newsworthiness,” reporters treat outputs as tips, and humans make the final editorial calls.

Outside newsrooms, a dozen public tools now offer to answer questions about the files — EpsteinGPT, Jmail, askepstein, epsteinunboxed, Sifter Labs and others. Most are small volunteer projects. None is audited. All of them are, for most people, the only practical way in.

The most dangerous failure mode is specific, and the Times names it. AI is “prone to error and hallucination, particularly around sensitive issues like redactions.”

This site already documents what happens when redaction goes wrong: the January 2026 release exposed at least 31 people who had been victimised as children, and the harm was permanent within hours. A model that fills in a blacked-out name is not producing a formatting error. It is producing a re-identification.

And the failures are not hypothetical. One volunteer database returned “gibberish” transcripts from handwriting. A widely repeated “23,000 emails” figure turned out to be wrong — only about 3,000 were emails. A CJR study found AI search tools inventing citations to articles that do not exist, with premium tools performing worse because they are tuned to sound authoritative.

Which is the honest summary. These tools are doing genuinely valuable work that the government failed to do, and the good ones link every claim to a source document. But the public record of this case is now substantially mediated by unaudited software — and the only real safeguard is that every claim remains traceable to a page.

There is a second, backward-looking question people reasonably ask: did Epstein fund artificial intelligence, and did any of it serve his aims?

He funded some. Marvin Minsky, a founder of the field, and Joscha Bach, whose work is on computational models of cognition, both received support through the MIT orbit. The Media Lab took roughly $800,000 directly and $7.5 million arranged.

This archive audited all ten of his funded research programmes against his stated aims. AI is filed as “adjacent, not instrumental.” Mind-uploading is a transhumanist objective and cognitive architecture is its notional substrate — but the funded work was basic research decades from any such application.

Four of the ten programmes had no possible connection to anything he wanted. Cosmology cannot breed anybody. Quantum information theory has no route to heritable modification. The honest finding was that he was not buying a capability — he was buying a room.

Where AI does connect substantively is downstream, and it is not about him. The surveillance thread runs from cameras in a townhouse to Palantir, Carbyne, facial recognition sold into Nigeria, and Paragon’s Graphite now under a $2 million ICE contract. Those are machine-learning systems, and they are operating now.

And Minsky’s own position should be stated. He was named in a deposition, denied it, and died in 2016. No finding was ever made.