How to detect deepfake evidence: a screening guide for lawyers

When should lawyers screen for AI generated evidence?

Screen evidence as soon as a case may turn on a photo, recording, video or document whose origin you cannot confirm independently. Litigators, in-house counsel, family and employment lawyers, insurance defense teams, paralegals and investigators all receive material like this.

• A recording produced by the other side: a voicemail or call in which your client appears to threaten, admit or agree to something.
• Photos of injuries, property damage or a scene, supplied without the original device or file.
• Video in custody, harassment or employment disputes, often forwarded through messaging apps.
• Letters, contracts or screenshots of messages that a client or witness says they never wrote.
• Your own client’s evidence, which you need to stand behind before you file it.
• Genuine evidence that the other side now dismisses as “a deepfake”.

That last case matters as much as the others. As generators improve, claiming that real evidence is fake becomes an easy line of attack, and you need a record showing how you checked.

Is AnyFact a forensic tool for court?

No. AnyFact is a screening tool that estimates whether a file shows signs of AI generation. It is not a forensic expert service, it does not provide expert testimony, and its results are not court-certified.

• It does not maintain a chain of custody for the files you upload.
• It does not certify that a file is authentic or unaltered.
• It is not designed to find manual edits such as spliced audio, removed frames or a changed figure in a PDF.
• Its PDF report documents a screening result; it is not an expert report and should not be presented as one.

How are courts handling deepfake evidence?

Courts are applying existing authentication rules and starting to sanction fabricated exhibits. In Mendones v. Cushman & Wakefield, Judge Victoria Kolakowski of the Alameda County Superior Court in California dismissed the case in September 2025 after finding that a video exhibit presented as a real witness was a deepfake. The court noted that it lacked the time, funding and technical expertise to forensically test the other suspect exhibits.

In US federal courts, Rule 901(a) of the Federal Rules of Evidence requires the proponent to produce evidence sufficient to support a finding that an item is what the proponent claims it is. The federal Advisory Committee on Evidence Rules also published a proposed Rule 707 on machine-generated evidence for public comment in 2025. Rules differ by jurisdiction and are changing, so check the current position where you practice. This page is not legal advice.

What is at stake with a deepfake in court, either way?

Relying on fabricated evidence can sink a case and expose the lawyer who offered it to sanctions and lasting damage to their credibility with the court. Failing to spot a fake from the other side can mean a client loses on evidence that never should have been admitted.

Wrongly calling genuine evidence fake does harm too. It can discredit a truthful witness, retraumatize a victim whose recording is real and cost you the court’s trust when an expert later confirms authenticity. A screening result of “probably AI” is a reason to investigate, never a basis for an allegation in a filing.

How to detect deepfake evidence: a screening workflow

Treat screening as part of evidence intake, not as an afterthought before trial. The steps below keep the original intact and tell you early whether an expert is needed.

1. 01Preserve the original. Secure the native file and, where possible, the device or account it came from. Record a hash value, the date received and everyone who handled it.
2. 02Establish provenance through people: who recorded or created it, on what device, when, where, and who else was present.
3. 03Read the metadata and any content credentials on a working copy. A content credentials checker shows C2PA and EXIF/XMP data that may name a camera, an editor or an AI tool.
4. 04Test internal consistency: timestamps against known events, weather and lighting against the claimed time, background sound against the claimed place.
5. 05Screen the working copy with the matching detector and read the signals and limits it reports. Never upload the only copy of anything.
6. 06Decide whether to retain a forensic examiner. If the item is contested, central to the case, or the screen is uncertain or probably AI, the answer is usually yes.
7. 07Document each step so you can explain later how and when you checked.

Which AnyFact detectors fit legal screening, and what do they return?

Four detectors match the evidence lawyers see most. Each returns a verdict (probably human, uncertain or probably AI), the signals behind it, a plain-language explanation, the limits of the method and a downloadable PDF report.

Longer recordings need trimming to the passage in dispute. For an overview of how AnyFact combines image, video and audio analysis, see the deepfake detector page.

Limits and confidentiality for legal teams

Screening is weakest on exactly the files lawyers often get: compressed messaging-app copies, short clips, phone-quality audio and screenshots. A real recording that was cut or re-ordered by hand may come back probably human, because the tool looks for generation, not editing.

Before uploading client material, check your professional duties of confidentiality and any protective order. AnyFact deletes uploaded files after 30 days and processes them with Modal, Supabase and Anthropic’s Claude. It has no public API yet and works through its web app.

Frequently asked questions