Deepfake evidence and the rules: US federal practice, England and Wales, and the EU AI Act

TimeBinder adds a layer of protection against AI and human fraud and disputes. For lawyers handling digital exhibits, the starting point is knowing what the rules say today, and what is still only a draft.

Deepfakes and altered files are now a routine worry for anyone who relies on a screenshot, a recording or a PDF. The harder question is what the rules actually require. This guide reads three primary sources for lawyers working in the United States, England and Wales, and the EU: the US Federal Rules of Evidence and the rule-makers’ own papers, Part 32 of the Civil Procedure Rules, and Article 50 of the EU AI Act.

It is general information, not legal advice. Evidence rules differ by court, by state and by member state, so check the rules that apply to your matter. Where something is a draft, we say so, and we link every source so you can read it yourself.

United States (federal): a low authentication bar, and a draft rule on deepfakes

The current rules. Under Federal Rule of Evidence 901(a), the proponent must produce evidence “sufficient to support a finding that the item is what the proponent claims it is.” Rule 901(b) lists examples, including distinctive characteristics taken together with all the circumstances (901(b)(4)) and evidence describing a process or system that produces an accurate result (901(b)(9)).

Rule 902(13) and 902(14) let a party authenticate certain electronic records, and data copied from a device or file, by a certification from a qualified person, with advance notice to the other side. The 2017 Committee Note to 902(14) explains that such data is ordinarily authenticated by hash value, and that the certification can establish authenticity only: the opponent can still object on other grounds, such as hearsay or relevance.

Deepfakes: a draft, not a rule. The Advisory Committee on Evidence Rules has been studying deepfakes. In its report of 17 May 2026 it said that, at least for now, an amendment to Rule 901 is not warranted, because few examples exist of courts having to address deepfakes; a Federal Judicial Center survey found only fifteen judges who reported dealing with deepfake issues. It is nonetheless holding a working draft of a new Rule 901(c) “in abeyance,” ready if it concludes an amendment is needed. As drafted, the opponent would first have to present evidence sufficient to support a finding of fabrication by generative AI; if they did, the item would be admitted only if the proponent shows it is more likely than not authentic.

The draft has not been approved or published for comment. It was on the agenda of the Committee’s 15 October 2026 meeting, and we have not seen the outcome, so check the Committee’s page for later developments.

Update, 15 October 2026 meeting. The Advisory Committee on Evidence Rules met in Boston on 15 October 2026. According to a Reuters report carried by Investing.com, OpenAI and Meta representatives told the committee that detecting deepfakes reliably will become harder; these are statements to the committee as reported, not court findings. A trade newsletter reports that the committee sent both the Rule 901 deepfake proposal and the Rule 707 proposal to subcommittees for further study, with its next meeting on 14 April; we have not found official minutes, so check the committee’s page. A dated, checkable record shows that an exact file existed no later than a recorded time and still matches; it does not show that the file is genuine.

A separate proposal: Rule 707. Proposed Rule 707 concerns machine-generated evidence offered without an expert. It is about reliability of AI output, not about whether a file is fake. It was published for comment in 2025; the Committee later voted to withdraw it as published and study a revised version (same agenda book).

A real example. The Committee’s agenda book describes Mendones v. Cushman & Wakefield, Inc. (Cal. Super. Ct., 9 September 2025), in which the judge imposed terminating sanctions after finding that audio exhibits and graphic material submitted by the plaintiffs were deepfakes or materially altered. State courts apply their own rules, which often follow the federal ones; check yours.

England and Wales: general rules, and a notice to prove

We did not find anything in CPR Part 32 that deals with deepfakes by name. The mechanisms it does contain are general.

• Rule 32.19. A party is deemed to admit the authenticity of a document disclosed under Part 31 unless it serves notice that it wishes the document to be proved at trial. The notice must be served by the latest date for serving witness statements, or within 7 days of disclosure, whichever is later. A party who doubts a disclosed document therefore needs to act early.
• Rule 32.1. The court may control the evidence: the issues it needs evidence on, the nature of that evidence and how it is put before the court.
• Rule 32.14. Proceedings for contempt of court may be brought against a person who makes, or causes to be made, a false statement in a document verified by a statement of truth without an honest belief in its truth.

The judiciary’s own AI guidance for judicial office holders (October 2025) says: “AI tools are now being used to produce fake material, including text, images and video,” and that “Courts and tribunals have always had to handle forgeries, and allegations of forgery, involving varying levels of sophistication.” Judges, it says, should be aware of the potential challenges posed by deepfake technology.

EU: the AI Act labels deepfakes, but it is not an evidence rule

Article 50 of the AI Act applies from 2 August 2026. It defines a deepfake (Article 3(60)) as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Deployers must disclose such content clearly, at the latest on first exposure, and providers of generative systems must mark outputs in a machine-readable way. Providers of systems already on the market before 2 August 2026 have until 2 December 2026 for the marking duty. The Commission says fines can reach 15 million euros or 3% of worldwide turnover.

Our reading is that this is a transparency duty on the businesses that build and use generative AI. It does not set out how a court should test an exhibit, which stays with national procedural law. It also has limits that matter for evidence: the Commission FAQ says personal, non-professional use is outside the AI Act, so an absent label shows nothing about whether a file is genuine.

What the three have in common: provenance, originals and a record made before the dispute

None of the three sources tells a court to trust a file because it looks right. Each points, in different words, to the same practical questions.

For the technical side, see our guides on forensic examination vs an independent timestamp and screenshots as evidence.

A short checklist for the lawyer holding the exhibit

1. Get the native file. Ask for the original, not a forward, a screenshot or a re-saved copy.
2. Work from a copy. Opening and re-saving can change a file. Keep the original untouched.
3. Write the custody note. Matter number, exact file name, who supplied it, from which device or account, the date and the reason. Our chain-of-custody log template lists the fields.
4. Record the hash. Note the file’s fingerprint at receipt, so any later copy can be compared. Our guide to checking whether a file has changed shows how.
5. Date the file. A dated record made now shows the file existed no later than that time, before any dispute begins.
6. Diarise the procedural dates. In England and Wales, the CPR 32.19 deadline. In US federal cases, the notice period for a Rule 902(13) or 902(14) certification. Check your own court’s rules.
7. Do not rely on a detector score. It is an indicator, not a finding. See what to do if the other side says your evidence is a deepfake.

If the other side calls your exhibit a deepfake, or you doubt theirs

Under the current US rules the authentication bar is low, and the draft 901(c) would ask a challenger for an initial showing before the proponent has to prove more. In England and Wales, the practical lever is the notice to prove under CPR 32.19, served on time. In both systems, the side that kept the original, the custody note and an earlier dated record is better placed to answer. Our guide to AI forgeries and backdating for lawyers covers the red flags in more depth.

How a TimeBind fits, and what it does not show

A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches its fingerprint. It fits in the “when did this exact file exist” and “does the copy match” rows above. A SHA-256 fingerprint is the kind of hash value the 2017 Committee Note describes, but a certification under Rule 902(14) has to come from a qualified person, and TimeBinder does not provide one.

A TimeBind shows that this exact file existed no later than the recorded time and still matches its fingerprint. TimeBinder makes the fingerprint in your browser and the file stays on your device. It does not show who made the file, who owns it, or that anyone agreed to it. It does not show that the content of the file is true, and it does not decide how any court will treat the file. TimeBinder takes no position on what a particular court will accept. It complements your engagement terms, your professional rules and legal advice; it does not replace them. This is general information, not legal advice.

Sources checked

Checked on 11 November 2026. We quote only what these documents say and mark where a point depends on your own rules or local law. Anything we could not verify from a primary source is left out.

• Federal Rule of Evidence 901: text and notes
• Federal Rule of Evidence 902: 902(13), 902(14) and the 2017 Committee Notes
• Advisory Committee on Evidence Rules, report of 17 May 2026: the working draft Rule 901(c), the Rule 707 status and the Federal Judicial Center survey
• Advisory Committee on Evidence Rules, agenda book, 15 October 2026: agenda items on deepfakes and Rule 707, and the Mendones summary
• Civil Procedure Rules, Part 32: rules 32.1, 32.14 and 32.19
• AI guidance for judicial office holders, October 2025: the passage on fake material and deepfakes
• European Commission, Article 50 AI Act FAQ: scope, dates, deepfake definition and fines
• TimeBinder pricing: Pooled and ASAP prices

Date the exhibit before anyone doubts it

When a file arrives, record its fingerprint and date it straight away, then file the certificate with your custody note. Your first Pooled TimeBind is free, with no credit card required. After that, Pooled TimeBinds start from US$5, and an ASAP TimeBind is US$15. Prices checked 11 November 2026. See plans and pricing. Open TimeBinder