Audio evidence and voice cloning: what lawyers and clients should preserve

This is general information for lawyers and clients, not legal advice. Evidence rules, and the law on recording calls, differ by country and state, so check the rules that apply to your matter. We link every source.

Why audio is a weak spot

Rule 901(b)(5) of the US Federal Rules of Evidence lets a witness authenticate a voice by opinion, based on hearing it “under circumstances that connect it with the alleged speaker” (Cornell LII). Rule 901(b)(6) covers telephone conversations by showing a call was made to a number assigned to a particular person or business. Both were written for a world in which a convincing imitation of a voice was hard to make. They are tests of familiarity and circumstance, not of whether a waveform was synthesised.

A UC Berkeley School of Information article about researchers Hany Farid and Sarah Barrington says that a voice can be cloned convincingly from as little as 30 seconds of reference audio and that accusations of AI voice clones have now reached the courts. That does not mean every recording is suspect. It means a recording you intend to rely on, or one served on you, deserves the same care as a signed document.

Fraud is part of the same trend. The FBI’s 2025 Internet Crime Report says voice cloning can be used to request a wire payment, and describes “distress” scams in which a cloned voice mimics a loved one; victims reported over US$5 million in losses to those in 2025. IC3 counted more than 22,000 AI-related complaints in total, with adjusted losses above US$893 million. These figures count complaints that mention AI, not proof that AI caused each loss.

What courts have already seen

England and Wales, family court (heard 2019, reported 2020). Family lawyer Byron James told The Telegraph that an audio file used in a custody dispute made his client appear to threaten the other parent. His client insisted he had never said it, though the voice sounded like him. James obtained the original file, exported it and examined the metadata, and said it showed the recording had been manipulated. The Telegraph describes it as believed to be the first reported case of its kind in UK courts, and in a LexBlog write-up he says the file was shown to be a deepfake rather than a faithful recording of a conversation. This is the lawyer’s own account of a private family case, not a published judgment.

Maryland, school workplace (2024 to 2025). A clip purporting to capture Pikesville High School’s principal making racist remarks circulated in January 2024. AP reports that the school’s former athletic director, Dazhon Darien, entered an Alford plea (acknowledging the evidence against him without admitting guilt) to one misdemeanour of disturbing school operations and was sentenced to four months in jail in April 2025. According to The Baltimore Banner, a media-forensics expert found traces of AI-generated content and later editing that added background noise; police also tied a Gmail account to the sending of the clip, and investigators later found he had searched for AI voice cloning and used a website, with a paid account, to make it. CBS Baltimore reports three experts concluded the recording contained AI-generated content with human-edited background noise. The point for lawyers is how the case was investigated: through forensic analysis, account records and device evidence, not by listening.

California, civil litigation (2025). In Mendones v. Cushman & Wakefield, Inc. (Cal. Super. Ct., 9 September 2025), the Advisory Committee on Evidence Rules’ October 2026 agenda book says the judge imposed terminating sanctions after finding that certain audio exhibits and graphic material were deepfakes or materially altered. The court also questioned other submissions, including voicemail messages, but said it did not have the time, funding or technical expertise to conduct a forensic analysis of them.

For the current federal and England-and-Wales procedural position, including the draft Rule 901(c) and the CPR 32.19 notice to prove, see our guide to deepfake evidence and the rules. We do not repeat it here.

What to preserve when audio may matter

The one practical thread in these cases is that the original file and its context carried the weight. A checklist for a client or a lawyer at intake:

1. Keep the native file. Do not rely on a screen recording of a voicemail, a forwarded message or a file that has been converted or compressed. Export the voicemail or voice note in its original format where the app allows it.
2. Keep the whole recording. A clip can hide what came before and after. Keep the full call or message, and note any cuts you know about.
3. Keep the source device and account. Do not wipe, reset or trade in the phone. The Pikesville case turned partly on device and account evidence. Note which app recorded it, the device model and the account it came from.
4. Record who handled it and when. Write a short custody note: file name, who supplied it, from which device or account, the date and time, and why. For a fuller log, see our chain-of-custody log for digital exhibits. For how timestamping fits at intake, see our timestamping at matter intake guide.
5. Fingerprint the file at receipt. A hash lets you show later that a copy is identical to what you received. The free SHA-256 checker shows a file’s fingerprint in your browser, and how to tell if a file has been changed explains the comparison.
6. Collect corroboration. Phone-carrier call records, message logs, calendar entries and witnesses can support or undercut a recording. A recording that matches nothing else deserves more scrutiny.
7. Ask for the native file from the other side. If an opponent serves audio, ask early for the original file, its metadata and the device or account it came from, and diarise any deadline to challenge authenticity. In England and Wales, CPR 32.19 sets a short window; our rules guide explains it.
8. Do not rely on a “detector” score alone. Axios quotes a digital-investigations specialist saying the detection tools on the market are a black box about how they produce a percentage. Treat a score as a lead, not a finding, and consider a qualified examiner.

Recording laws vary. A Justia 50-state survey says most US states require only one party’s consent to record a call, while a smaller group requires every party’s; rules in the UK and elsewhere differ again, and covertly obtained recordings can raise admissibility and privacy issues. Take advice before relying on a recording you made yourself.

When to call a forensic examiner

A dated fingerprint does not decide whether audio is genuine. Where authenticity truly matters, a forensic audio examiner can look at the file’s structure, encoding history, noise and signal characteristics, and the device evidence around it. Those examinations have limits too: the Axios report on deepfakes in the judicial system reports that there are not enough forensic analysts to verify AI-faked evidence, and that even expert analysis may not be enough to persuade a jury beyond a reasonable doubt. See our comparison of forensic examination vs an independent timestamp for what each can show.

Budget and proportionality matter. The Mendones court said it lacked the resources to test every doubtful item. Early, targeted preservation is cheaper than a late examination of material that has been copied, converted and forwarded several times.

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. The fingerprint is made in your browser, and the file stays on your device. For audio, that means: you can show that the file you hold is the same one you fingerprinted on receipt, and that it existed by that date, before any dispute was on the horizon.

It does not show who is speaking, that a voice is genuine, that a recording was lawfully made, or that it will be admitted. A TimeBind made on a fake recording would simply date the fake. It is one layer next to the native file, the custody note, corroborating records and, where needed, an examiner. For a broader overview, see our guide for lawyers on AI forgeries, our page on legal timestamping, and what to do if the other side says your evidence is a deepfake. Plain-English definitions of voice cloning, deepfake and related terms are in the AI fakes and file-integrity glossary.

Sources checked

• Federal Rules of Evidence, Rule 901, Cornell Legal Information Institute: law.cornell.edu
• Justia, “Recording Phone Calls and Conversations Under the Law: 50-State Survey” (last reviewed September 2024): justia.com
• UC Berkeley School of Information, “Prof. Hany Farid and Ph.D. Student Sarah Barrington Speak Out About the Dangers of AI-Generated Voice Evidence Used In Court”, 2025: ischool.berkeley.edu
• FBI Internet Crime Complaint Center, 2025 Internet Crime Report (AI used in cybercrime): ic3.gov
• The Telegraph, “Doctored audio evidence used to damn father in custody battle”, 31 January 2020: telegraph.co.uk; LexBlog, “Family Lawyers: A Real-Life Case of a Deepfake Audio”, 4 February 2020: lexblog.com
• AP, “Former school athletic director gets 4 months in jail in racist AI deepfake case”: apnews.com; The Baltimore Banner: thebanner.com; CBS Baltimore: cbsnews.com
• Advisory Committee on Evidence Rules, agenda book for the October 2026 meeting (discussing Mendones v. Cushman & Wakefield, Inc.): uscourts.gov
• Axios, “Courts aren’t ready for AI-generated evidence”, 25 July 2025: axios.com

If you are sorting a suspicious call or a possible voice clone, a free checklist can help you decide what to do next: the AI Voice Scam Red-Flags Checker. It does not analyse an audio file or detect a clone, and it is not legal advice.

Start a dated record of the files you receive

Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.