Fact-Checking & AI Detection Tools
Accurate research requires more than finding a name, running a search, or asking an artificial intelligence system for an answer. Every claim published by EpsteinWiki must be traced to reviewable evidence, checked against its original context, and classified according to what the evidence actually establishes.
This guide identifies tools that can help contributors locate records, verify documents, examine media, detect possible synthetic content, recover deleted webpages, and compare competing claims. These tools support investigation. They do not replace human review, original documents, chain of custody, or editorial judgment.
Snapshot
| Principle | EpsteinWiki Standard |
|---|---|
| AI answer | Research lead, never a final citation |
| Search result | Discovery aid, not proof |
| OCR text | Machine generated transcription requiring comparison with the page image |
| AI detector score | Indicator requiring corroboration |
| Metadata | Evidence about a file, not proof of the depicted event |
| Court database copy | Verify docket, filing number, date, and document identity |
| EFTA result | Open the document and confirm the cited page |
| Name match | Confirm identity before connecting the person to the record |
| Missing result | Does not prove that evidence does not exist |
| Serious allegation | Requires primary evidence and second editor review |
The Most Important Rule
No fact checking or AI detection tool can determine guilt, complicity, credibility, criminal intent, or the truth of an allegation by itself.
A tool may help establish that:
- A document contains a particular name.
- A photograph appeared online before a claimed date.
- A file contains certain metadata.
- A video may contain evidence of editing.
- An image carries a provenance credential.
- A court filing exists on a particular docket.
- Several accounts repeated the same claim.
The tool cannot automatically establish why the name appears, whether the statement is true, who performed an edit, whether accounts coordinated, or whether a person committed wrongdoing.
Epstein Files Research Tools
Epstein Data
Epstein Data provides full text searching across the Department of Justice Epstein document production, including indexed documents, photographs, spreadsheets, audio, video, and transcribed material.
Use it to:
- Search EFTA identifiers.
- Locate exact names, phrases, email addresses, companies, telephone numbers, and locations.
- Find related documents across datasets.
- Review extracted text and searchable transcriptions.
- Open the underlying document connected to a result.
Every EpsteinWiki citation should link to the relevant record whenever an Epstein Data copy is available.
Limitations: Search results depend on OCR, transcription, indexing, and entity extraction. Handwriting, damaged scans, unusual fonts, redactions, poor audio, and formatting may produce errors. Always open the underlying evidence and compare the extracted text with the original page, image, audio, or video.
Ask Epstein AI
Ask Epstein AI is an evidence research system designed to search records, trace relationships, organize claims, and identify supporting or contradictory material.
Use it to:
- Generate leads for further document review.
- Locate records associated with a person, institution, event, or date.
- Identify possible relationships requiring verification.
- Find contradictions or gaps in a developing theory.
- Produce initial research questions.
Limitations: AI generated responses can omit context, merge people with similar names, misread OCR, misunderstand legal language, or create unsupported connections. Every cited claim must be checked against the original record. Ask Epstein AI is not itself evidence.
Department of Justice Epstein Library
The official Department of Justice Epstein Library houses materials released under the Epstein Files Transparency Act. Its full library search can help locate records across the federal production.
Use it to:
- Confirm that a document appeared in an official release.
- Identify the original dataset.
- Review DOJ supplied file names and organization.
- Compare mirrored copies with the government hosted version.
- Monitor later additions or corrections.
Limitations: The Department of Justice warns that handwritten and unusually formatted materials may not be reliably searchable. The library may also contain sensitive information that escaped redaction. Contributors must not republish exposed survivor or private identifying information merely because it appeared on a government website.
FBI Vault
The FBI Vault Jeffrey Epstein collection contains FBI records released through the Freedom of Information Act.
Use it to:
- Confirm FBI Vault provenance.
- Download the full released file part.
- Review surrounding pages.
- compare FBI pagination with other copies.
- Identify redactions and FOIA markings.
Limitations: A document’s presence in the FBI Vault does not establish that every allegation inside it was substantiated. FBI files can include tips, leads, summaries, administrative records, third party statements, and unverified claims.
Court Record Tools
CourtListener and RECAP
CourtListener provides access to opinions, dockets, oral arguments, and documents collected through the RECAP archive. Its RECAP search contains millions of documents originally obtained through PACER.
Use it to:
- Search federal dockets.
- Locate complaints, motions, exhibits, orders, and transcripts.
- Track filing history.
- Create docket alerts.
- Obtain free copies of records already contributed to RECAP.
Limitations: RECAP does not contain every PACER document. Docket entries may exist without an attached file. Sealed, restricted, withdrawn, or newly filed material may be unavailable.
PACER
PACER is the federal judiciary’s official electronic records system. The PACER Case Locator searches a nationwide index of federal cases.
Use it to:
- Confirm the official federal docket.
- Verify filing numbers and dates.
- Retrieve documents missing from RECAP.
- Check whether a filing was amended, sealed, restricted, terminated, or superseded.
- Review the complete procedural history.
Limitations: PACER requires registration and may charge access fees. A filed allegation is not a judicial finding. Always distinguish complaints, testimony, exhibits, motions, orders, verdicts, and judgments.
Web and Claim Verification Tools
Internet Archive Wayback Machine
The Wayback Machine can locate historical versions of public webpages. Its Save Page Now function can preserve many currently available pages.
Use it to:
- Confirm what a website displayed on a particular date.
- Recover deleted biographies, press releases, schedules, and statements.
- Document changes to webpages.
- preserve sources vulnerable to deletion.
- compare earlier and later versions.
Limitations: Archived pages may omit embedded media, scripts, linked files, private material, or images. An archive timestamp shows when the crawler captured the page, not necessarily when the content was originally created.
Google Fact Check Explorer
Google Fact Check Explorer searches fact checks published with structured ClaimReview data. It can also help identify earlier reviews of an image.
Use it to:
- Determine whether a viral claim has already been investigated.
- locate competing fact checks.
- Search by person, phrase, subject, or image.
- Identify primary sources cited by fact checkers.
Limitations: A result reflects the work and classification of the publishing organization. Read the full analysis and inspect its sources. Absence from the database does not mean a claim is true.
Reverse Image Search
TinEye searches for earlier and modified copies of an image. Google Lens, Bing Visual Search, and other reverse image systems may return different results because each searches a different index.
Use reverse image search to:
- Find earlier appearances.
- Locate higher resolution copies.
- identify stock photography.
- Discover alternative captions.
- Detect cropping, mirroring, recoloring, or reuse.
Limitations: The oldest indexed result may not be the original source. A match can establish reuse without proving who created the image or whether its caption is accurate.
InVID Verification Plugin
The InVID Verification Plugin provides keyframe extraction, metadata viewing, magnification, reverse image search, and contextual tools for photographs and videos.
Use it to:
- Break a video into searchable frames.
- Search individual frames for earlier appearances.
- Inspect visual details.
- Compare thumbnails.
- gather available platform context.
Limitations: Filters and visual anomalies are investigative indicators. They do not automatically prove that an image or video was altered.
Local File Inspection Tools
Local tools are preferable when a file contains sensitive, private, unpublished, or legally restricted information. Uploading a file to a public detection service may disclose it to another company or create an uncontrolled copy.
ExifTool
ExifTool reads metadata from many image, video, audio, document, and archive formats.
Use it to inspect:
- File creation and modification fields.
- Camera and device information.
- GPS data.
- Editing software fields.
- embedded thumbnails.
- XMP, EXIF, IPTC, PDF, and container metadata.
Limitations: Metadata can be changed, removed, copied, generated during export, or replaced by a social platform. Missing metadata does not prove fabrication.
MediaInfo
MediaInfo displays technical and tag information from audio and video files.
Use it to examine:
- Container format.
- Codec.
- Duration.
- Frame rate.
- Resolution.
- Bit rate.
- Audio tracks.
- Writing application.
Limitations: Technical characteristics can identify inconsistencies or processing history but rarely establish who edited a file or why.
ffprobe
ffprobe examines multimedia streams and returns detailed information about containers, video, audio, subtitles, timestamps, and stored metadata.
Use it when a precise, exportable technical report is needed.
Limitations: ffprobe reports what exists inside the examined copy. It does not reconstruct missing provenance or authenticate the depicted event.
Tesseract OCR
Tesseract is an open source optical character recognition engine that extracts searchable text from images.
Use it for:
- Scanned court records.
- Government files.
- Photographed documents.
- Tables and typed correspondence.
- Search preparation.
Limitations: OCR can confuse names, dates, punctuation, page numbers, account numbers, and handwritten text. Never quote OCR without comparing it with the page image.
OCRmyPDF
OCRmyPDF adds a searchable OCR layer to scanned PDF files.
Use it to create searchable working copies while preserving the original file separately.
Limitations: The searchable layer contains machine generated guesses. The OCR copy must be labeled as a working derivative and must not replace the preserved original.
Cryptographic Hash Tools
SHA 256 hashing tools create a digital fingerprint of a file. Hashes help determine whether a preserved copy changed after acquisition.
Use a hash to:
- Identify the exact file examined.
- Compare copies.
- document transfers.
- detect later changes.
- connect analysis results to a specific version.
Limitations: A matching hash establishes that two copies contain the same bytes. It does not prove that the original file was authentic or accurately described.
AI Generated Media and Provenance Tools
Content Credentials
The official Content Credentials verification tool checks for C2PA provenance information embedded in supported media.
A Content Credential may identify:
- The signing application or device.
- Creation or editing actions.
- Ingredients used to make the final file.
- Whether artificial intelligence was involved.
- Whether the signed manifest still validates.
Limitations: A valid credential documents signed provenance assertions. It does not establish that a caption is true or that the depicted event happened. A missing credential does not prove that content is synthetic.
Google SynthID
Google SynthID places imperceptible watermarks in content created or edited through supported Google artificial intelligence systems. Gemini can check supported images, audio, and video for a SynthID watermark.
Use it to ask whether media carries a Google SynthID signal.
Limitations: A negative result does not establish that the item is human created. It may have been produced by another system, processed in a way that affected detection, or created without SynthID.
DeepFake O Meter
The University at Buffalo DeepFake O Meter runs media through multiple research detection models rather than returning only one model’s assessment.
Use it to:
- Compare detector outputs.
- identify disagreement among models.
- generate leads for technical review.
- document model specific scores.
Limitations: Models may perform differently on unfamiliar generators, compressed media, short clips, unusual languages, or edited files. Conflicting results should be expected. Upload only public, lawful, nonsensitive material after reviewing the service’s current policies.
Why AI Detectors Cannot Deliver a Verdict
The National Institute of Standards and Technology report on synthetic content explains that watermarking, provenance tracking, metadata, and automated detection each have strengths and limitations. No single technique provides a comprehensive solution.
Detector accuracy may change according to:
- The system that generated the content.
- Whether the detector was trained on that generator.
- Compression and resizing.
- Screenshots and screen recordings.
- Cropping or transcoding.
- Added noise, subtitles, filters, or music.
- File length and resolution.
- Language, accent, and voice quality.
- Whether human editing followed generation.
- Changes made after the detector was released.
A score of 80 percent does not mean there is an 80 percent legal probability that the file is fake. The meaning of a score depends on the model, testing conditions, calibration, and documentation supplied by the developer.
AI Text Detection
EpsteinWiki must not use an AI text detector to accuse a survivor, witness, contributor, journalist, or public figure of fabricating a statement.
Text detectors can be affected by writing style, editing, language proficiency, sentence length, formatting, and the use of grammar assistance. Human writing may be falsely classified as artificial, while edited AI text may evade detection.
If authorship matters, examine:
- File history.
- publication chronology.
- account ownership.
- version records.
- drafts.
- metadata.
- communication records.
- testimony from people involved.
- consistent or inconsistent factual knowledge.
- independent corroboration.
An AI text detector result may be recorded as a lead. It must not be presented as proof of authorship.
Using Artificial Intelligence for EpsteinWiki Research
Artificial intelligence may assist with:
- Initial document sorting.
- OCR correction suggestions.
- Search term development.
- Timeline extraction.
- Entity identification.
- Transcript preparation.
- Translation drafts.
- Comparison of document versions.
- Identifying contradictions.
- Generating questions for human research.
Artificial intelligence must not be used to:
- Invent missing text.
- reconstruct redactions as fact.
- create quotations.
- infer guilt from associations.
- merge people solely because they share a name.
- describe an allegation as a finding.
- replace review of the original evidence.
- generate evidence citations that were not opened and checked.
- publish private or survivor information.
- make final editorial decisions.
Required AI Verification Process
When AI helps locate or summarize evidence, contributors must:
Step 1: Open the Underlying Record
Do not cite the AI response, search snippet, or extracted entity card.
Step 2: Confirm the Identifier
Verify the EFTA number, Bates range, docket number, exhibit, dataset, file name, date, and page.
Step 3: Read the Surrounding Context
Review the complete message, page sequence, transcript passage, filing section, or document chain.
Step 4: Classify the Statement
Determine whether it is a documented fact, allegation, testimony, denial, inference, judicial finding, conviction, unresolved claim, or editorial analysis.
Step 5: Check Identity
Confirm that the named person is the correct individual. Use occupation, location, organization, dates, contact details, and surrounding correspondence.
Step 6: Seek Corroboration
Look for genuinely independent records. Multiple databases containing the same source are not independent confirmation.
Step 7: Record Limitations
Disclose OCR errors, incomplete records, redactions, missing attachments, uncertain identity, conflicting testimony, or unavailable originals.
Step 8: Obtain Review
Serious allegations, survivor material, uncertain identities, and difficult evidence classifications require second editor review.
Privacy and Upload Safety
Before uploading any file to an online service, determine:
- Whether the service stores uploaded content.
- Whether files may be used for system training.
- Whether a human reviewer may access the file.
- Whether deletion is available.
- Whether the file contains personal information.
- Whether the file identifies a survivor or minor.
- Whether the material is sealed, privileged, intimate, or restricted.
- Whether uploading would violate consent, law, policy, or chain of custody.
Sensitive files should be examined locally whenever possible.
If material may contain child sexual abuse material, do not download, copy, upload, hash, forward, or inspect it for research. Report suspected online child exploitation through the NCMEC CyberTipline. Call 911 or the appropriate emergency service when someone is in immediate danger.
Recording Tool Results
Every material tool result should include:
| Field | Required Record |
|---|---|
| Tool | Name of the tool |
| Version | Version or model when available |
| Date | Date and time used |
| Input | Exact file, URL, query, or identifier |
| Hash | SHA 256 value for examined files when appropriate |
| Output | Result produced by the tool |
| Interpretation | What the result may indicate |
| Limitation | What the result cannot establish |
| Corroboration | Other evidence supporting or contradicting the result |
| Reviewer | Contributor who performed the review |
| Decision | Publish, hold, escalate, correct, or reject |
Screenshots of detector results should not replace the written record.
Evidence Labels
Use the following labels when documenting tool assisted research:
| Label | Meaning |
|---|---|
| Verified | Supported by primary evidence and sufficient corroboration |
| Partially verified | Some elements are supported while others remain unresolved |
| Authentic file, disputed context | The file is genuine but its interpretation is contested |
| Altered | Material modification has been established |
| Synthetic | Reliable evidence establishes artificial generation or reconstruction |
| Misattributed | Authentic material is assigned to the wrong person, date, place, or event |
| Unverified | Available evidence is insufficient |
| Inconclusive | Meaningful review was completed without a reliable determination |
| False | Credible evidence directly contradicts the claim |
Always include the basis and confidence level. Never publish a label without showing the supporting evidence.
Common Fact Checking Failures
Searching Only One Spelling
Names may appear under nicknames, initials, married names, former names, OCR errors, alternate transliterations, email addresses, or company names.
Treating Search Counts as Evidence
A person may appear thousands of times because a document was duplicated, an email chain was repeated, a news clipping mentioned the person, or an index generated multiple entries.
Citing the Search Snippet
Snippets may omit qualifying text, combine nearby passages, or display inaccurate OCR.
Confusing Filing With Finding
A complaint contains allegations. A motion contains argument. Testimony records what someone said. An order records what a court decided.
Assuming a Government File Is Verified
Government collections can include raw tips, allegations, leads, news clippings, witness statements, and records received from outside sources.
Accepting AI Entity Matching
People with similar names may be merged. One person’s occupation, travel, emails, or allegations can be incorrectly assigned to another.
Uploading Sensitive Evidence
Convenient online analysis can compromise privacy, consent, provenance, and survivor safety.
Using One Detector
A single detector can produce false positives or false negatives. High stakes conclusions require multiple methods and qualified review.
Questions Contributors Must Answer
Before relying on any tool result, ask:
- What exact claim am I testing?
- Is this the original record or a derivative copy?
- Can I open the full document?
- Did I verify the EFTA, Bates, docket, exhibit, page, and date?
- Could OCR or transcription have changed the meaning?
- Have I confirmed the person’s identity?
- Does the record contain a fact, allegation, denial, testimony, or finding?
- Is the apparent corroboration genuinely independent?
- What evidence contradicts the proposed conclusion?
- Did an AI system infer anything not stated in the source?
- Could the tool be unfamiliar with the file format or generator?
- Was sensitive information uploaded to another service?
- Can another contributor reproduce the result?
- What does the evidence prove?
- What does it not prove?
Related EpsteinWiki Guides
- Editorial Standards
- Content Classification Guide
- Evidence Handling 101
- Chain of Custody Protocol
- Disinfo & Manipulated Media Guide
- Handling Contradictory Evidence
- How to Read Epstein Documents Without Making Things Up
- Corrections Policy
- Research Methodology
Sources
- Department of Justice Epstein Library
- FBI Vault Jeffrey Epstein Collection
- PACER Federal Court Records
- CourtListener RECAP Archive
- National Institute of Standards and Technology: Reducing Risks Posed by Synthetic Content
- NIST Generative AI Evaluation Program
- Coalition for Content Provenance and Authenticity
- Content Credentials Verification Tool
- Google DeepMind SynthID
- University at Buffalo DeepFake O Meter
- InVID Verification Plugin
- Internet Archive Save Page Now
- Google Fact Check Explorer
- TinEye Reverse Image Search
- ExifTool
- MediaInfo
- ffprobe Documentation
- Tesseract OCR
- OCRmyPDF
- NCMEC CyberTipline
Editorial Note
Tool availability, search coverage, pricing, privacy practices, and detection models can change. Contributors should review current documentation and terms before using an online service.
This guide does not endorse a tool as infallible or suitable for every investigation. EpsteinWiki conclusions must rest on inspectable evidence, transparent methods, careful classification, and human editorial review.
Last reviewed: September 6, 2026