Source-linked AI summary
Epstein Files Engine: Agentic Search for Investigative Journalism
Duy K. Nguyen, Teresa Mondría Terol, Dylan Freedman, Zach Seward
TL;DR
The Epstein release was too large and varied for exhaustive reading or keyword search, while reporters also needed to identify what was genuinely new and verify candidate leads. The Engine translated questions into queries across institutional corpora and used Diff to match duplicate pages; more than 100 journalists used it, and it contributed to at least 20 published stories.
Problem
The mixed-media Epstein release was too large for exhaustive reading, and reporters needed help turning initial hunches into effective queries and checking leads against prior reporting and source documents.
Source: §1 · p. 1Method
The Engine translated reporter questions into database queries over institutional corpora and paired its cited reports with Diff, a text-and-visual duplicate-matching method.
Source: §1 · p. 1Results
More than 100 reporters used the Engine, which contributed to at least 20 published stories across politics, finance, sports, education and entertainment.
Source: §1 · p. 1Takeaways & Limitations
The authors argue newsroom agents are most useful as interfaces to institutional memory and source material that help journalists ask questions and verify evidence, rather than as autonomous writers.
Source: §5.3 · p. 5The use analysis describes a floor rather than a distribution, and Diff was not compared with text-only or visual-only variants or comprehensively benchmarked against alternatives.
Source: §5.1 · pp. 4–5
Abstract
from arXiv · showhide
On Jan. 30, 2026, the U.S. Department of Justice released a mixed-media collection concerning Jeffrey Epstein, including about three million pages of PDFs. We describe the Epstein Files Engine, an A.I. agent The New York Times deployed to investigate the files. The Engine translated reporter questions into Google BigQuery SQL queries across three corpora: Epstein-related releases, the Times's archive and external, Epstein-related news headlines. It used an LLM to plan queries and returned citation-rich answers a reporter could verify and trust. More than 100 journalists used the Engine, and it contributed to at least 20 published stories. We report how reporters queried it and describe Diff, our text-and-visual duplicate matching method that amplified novelty signals and allowed the Engine to surface genuinely new information. We argue that newsroom agents serve newsrooms best not as autonomous writers, but as interfaces to source material and institutional knowledge.
1 Introduction
The Epstein Files Engine addressed a mixed-media release too large for exhaustive reading and a search process that demanded reporters already know how to frame effective queries. It translated questions into database queries and used citations and duplicate matching to support exploration and verification.
1 Introduction
- The DOJ release combined roughly three million PDF pages with emails, images, audio and video, making exhaustive reading and keyword search inadequate to organize it.The Times inventory counted 1.1 million emails, 200,000 images and more than 1,000 audio and video clips.Source: §1 · p. 1
- The Engine translated plain-English reporter questions into BigQuery SQL across institutional corpora and returned retrieved records with citations.The answers were intended to expedite open-ended exploration and verification.Source: §1 · p. 1
- More than 100 reporters used the Engine, which contributed to at least 20 published stories across several beats.The reported beats included politics, finance, sports, education and entertainment.Source: §1 · p. 1
- Diff combined text and visual signals to amplify novelty, with high recall but low precision at its deployed threshold; some false positives still exposed reporting-relevant structural similarities.Examples included differently redacted versions of the same page, repeated forms and record genres.Source: §1 · p. 1
2 Related work
Prior computational-journalism tools support document organization, search, analysis and lead discovery, while audience-facing systems use newsroom archives to answer reader questions. The Engine brings query generation and novelty signals into a newsroom investigative workflow.
2 Related work
- DocumentCloud, Datashare, Google Pinpoint, Semantra and Hagar support newsroom document organization, indexing, entity extraction, semantic search or retrieval-augmented investigation [14].Their described functions range from collaborative analysis of large collections to search without generated answers.Source: §2 · p. 1
- Journalism tools have also supported lead discovery through framing, investigative tip sheets and data visualizations [33].The Engine draws on this work by pairing an LLM with Diff-based novelty signals to catalyze lead discovery.Source: §2 · p. 1
- Audience-facing archive chatbots answer reader questions, while text-to-SQL benchmarks study enterprise workflows and Diff concatenates inspectable text and visual signals [5] [21] [22] [11] [27] [34].The Engine applies SQL generation to newsroom workflows; the authors contrast Diff’s concatenated signals with trained multimodal representations.Source: §2 · p. 2
3 System design
The Engine queried three institutional corpora and used an LLM as a query planner, keeping cited records as the source of truth. Diff combined page-level text and visual representations for duplicate matching, helping the system distinguish potentially new material from re-releases.
3.1 The Engine’s underlying data
- The Engine queried the DOJ release and earlier Epstein disclosures, the Times reporting archive, and an internal feed of external Epstein-related headlines.Extracted images were tagged and captioned so visual assets embedded in PDFs could be queried.Source: §3.1 · p. 2
- For a name query, the Engine sought both what the files said and what had not already been reported, weaving the three corpora to suggest leads.Figure 1 depicts generated queries across the corpora and the assembly of returned records into one cited report.Source: §3.1 · p. 2, Figure 1 · p. 2
3.2 The model as query planner
- An LLM planned SQL queries over the corpora, and the Engine synthesized returned records into a citation-rich report rather than generating an authoritative answer from fixed queries.The queries ran separately and their results were assembled into one report.Source: §3.2 · p. 2
3.3 Diff’s deduplication mechanism
- Diff represented each page with a text embedding, a visual difference hash and a bias term, enabling semantic matching across changed layouts and visual matching across noisy scans.Text embeddings and visual hashes were independently normalized; missing modalities were represented as zero. Cosine nearest-neighbor search combined the signals, and the deployed duplicate threshold was 0.92.Source: §3.3 · pp. 2–3
4 Deployment and use
The Engine moved from a manually relayed prototype to self-serve newsroom use, supporting reporters’ searches and verification across the investigation. Diff was deployed to favor finding duplicates, with validation showing a precision–recall trade-off and uncertainty in some human labels.
4 Deployment and use
- The prototype’s manual report relay became self-serve when the Engine went live to reporters on Feb. 9, after users were cautioned to verify its fallible reports against cited documents.Between launch and mid-June, more than 100 journalists posed 4,500 questions to the agent.Source: §4 · p. 3
4.1 What reporters asked
- A common named-person query returned document findings, associated people, and what was new relative to prior coverage; more than 60 such reports were produced in the investigation’s first weeks.Source: §4.1 · p. 3
- Reporters used the Engine to connect document findings about a person with prior Times and external coverage, often seeking what was new and sometimes newsworthy.Other requests explored financial trails, relationships, contact timelines, coded language, and authorship; the Engine said it was not instructed to recover redacted text.Source: §4.1 · p. 3
4.2 From query to published article
- Reporters checked cited material against original documents and developed leads through original reporting; at least 20 targeted reports contributed to published articles across five coverage areas.The reports contributed to articles in politics, finance, sports, education, and entertainment.Source: §4.2 · pp. 3–4
4.3 Validating Diff
- At τ=0.92, Diff’s population-weighted precision was 0.28 and recall was 0.86, with 95% bootstrap confidence intervals of [0.21, 0.35] and [0.69, 1.00].The deployed threshold was manually tuned in response to reporters’ qualitative feedback and favored recall; higher thresholds trade recall for precision.Source: §4.3 · p. 4, Table 2 · p. 4
- Pairs marked unsure made up 18% of the sample; resolving them as duplicates or distinct shifted precision to [0.21, 0.43] and recall to [0.59, 0.83].The validation sampled 400 pairs across similarity bands, oversampling around the deployed cutoff, and used one annotator’s labels under a newsroom-informed rubric.Source: §4.3 · p. 4
5 Findings and lessons
Reporters used the Engine for investigative questions beyond archive comparisons, while citation-grounded answers gave them material to inspect and trust. The paper also describes limits to its deployment evidence, shows how different resolutions of near-duplicates support discovery, and argues for reusable newsroom agents as interfaces to source material and institutional memory.
5 Findings and lessons
- Reporters used the Engine for unscripted investigative work, including searching for financial transactions, reconstructing networks and timelines, finding coded language, and analyzing writing styles.These uses extended beyond archive comparisons.Source: §5 · p. 4
- Verifiable citations gave skeptical reporters something concrete to inspect, while the chat interface invited exploratory questions they might not ask a colleague or know how to express as SQL.The authors describe this as a newsroom-side intimacy dividend.Source: §5 · p. 4
5.1 Limitations
- The deployment account reconstructs and quantifies journalist activity as a floor, not a distribution, and does not test corpus contributions in a controlled study.The role of the Times archive and external headlines in turning results into leads is supported by qualitative descriptions; the authors invite empirical efficacy tests.Source: §5.1 · pp. 4–5
- Diff’s validation followed the news moment, used a manually tuned threshold, and did not compare text-only or visual-only variants or test whether every duplicate was found.The authors say the validation was intended to justify including Diff in the Engine, not to comprehensively benchmark it against alternatives.Source: §5.1 · p. 5
5.2 What near-duplicates reveal
- At fine resolution, redaction variants can reveal obscured details; coarser matches connect document series for timeline work and group genres to surface similar records.The same structural similarities support different discovery uses depending on how coarsely they are measured.Source: §5.2 · p. 5, Figure 2 · p. 5
5.3 What generalizes
- The Engine’s modular infrastructure and playbook are designed for reuse with other primary-source corpora, and documentation could let journalists build their own Engines.The authors frame this as a way to make “chat with your documents” an institutional capability.Source: §5.3 · p. 5
- The authors argue newsroom agents are most useful as interfaces to institutional memory and source material, helping journalists pose questions and verify evidence rather than writing autonomously.They place judgment in journalists’ hands.Source: §5.3 · p. 5