Digging Into
Three Million
Pages

The Scale of the Task

The Justice Department released the Epstein files on January 30, 2026 — a document dump so vast that, stacked, the pages would reach the top of the Empire State Building. About two dozen journalists are working through the material, yet have seen only 2–3% of it. At that rate, it would take years to review and verify everything.

Reporters from the Investigations, National, Metro, and Business desks were assembled alongside engineers and AI journalists, creating an unusually cross-functional team. They started with a simple method: search terms. Trump. Clinton. Gates. Duke of York. Names, places, and events connected to Epstein.

AI Meets Traditional Reporting: The Hybrid Approach

The NYT's approach was never to let AI drive the reporting — rather, to use it as an amplifier for human judgment. Here's how each layer worked:

• 01
  Document Ingestion & Indexing Andrew Chavez's Interactive News team spent ~10 hours uploading and indexing all 3 million pages into a proprietary search tool. Until ready, reporters used the DOJ's own clunky interface.
• 02
  Semantic Search (Beyond Keywords) Traditional search only finds exact matches. Semantic search finds conceptually related content — critical for navigating typos, OCR errors, and inconsistent language throughout the files.
• 03
  AI Tagging & Categorization An AI tool automatically bucketed documents by type (email, legal record, photo, text message) and added newsworthiness labels, giving reporters a fast triage layer.
• 04
  Visual Photo Search Reporters could search images by visual content — not just metadata — allowing them to locate photos and scan imagery at a scale impossible by hand.
• 05
  Duplicate Detection An AI tool flagged repeated documents across the 3 million pages, preventing reporters from counting the same information multiple times or chasing false leads.
• 06
  Crowdsourced Spreadsheet Verification Dylan Freedman built a scraper that pulled DOJ search results into organized spreadsheets by key figure. Reporters then collaboratively verified each entry — classic newsroom teamwork, supercharged.
• 07
  Video & Audio Transcription AI parsed the 2,000+ videos and audio files into searchable transcripts, making multimedia part of the investigative database — not a separate, harder-to-access silo.

What AI Can and Cannot Do

The Times team was remarkably candid about where AI helped and where it fell short — a more honest accounting than most newsrooms provide.

Chavez stressed that they gave AI only "discrete, narrow tasks" it could handle reliably — like identifying whether a page contained an image — rather than open-ended analysis. The AI surfaces signals; reporters follow up with human judgment and sourcing built over years.

AI Verification in Action: Two Case Studies

A third judgment call — around unverified Trump accusations — showed how editorial process and AI tools work together. The team found a document summarizing over a dozen unverified tips about Trump and Epstein, but chose to describe their existence in general terms without publishing unverifiable details. The article's published language: "The emails did not include any corroborating evidence and The New York Times is not describing the details of the unverified claims."

What the Files Have — and Haven't — Revealed

After reviewing roughly 2–3% of the material, the picture is clearer in some areas and murky in others:

• ✓
  Rich portrait of Epstein's orbit The files give an inside look at how Epstein operated — trading gifts and favors, enticing elites with dinners and island visits, and positioning himself as someone who "knew things" about powerful people.
• ✓
  Extensive network documentation 38,000+ references to Trump — though many are news clippings and previously released congressional documents. No major new revelations about his relationship with Epstein emerged.
• ✗
  No proof of wide pedophilia ring Despite thorough searching, no hard evidence of a broad ring has emerged in what's been reviewed. Possible co-conspirators named by investigators were not new names.
• ✗
  No clear blackmail proof The theory that Epstein collected secrets for leverage remains unproven. The files show he "seemed to see value" in claiming to know things — but hard evidence has not surfaced yet.