You Cannot Protect Your Voice. You Can Protect The People Who Answer The Phone.

Almost every guide here is supply-side, and that argument was lost three years ago.

In January 2023, Microsoft researchers published VALL-E, which could "synthesize high-quality personalized speech with only a 3-second enrolled recording of an unseen speaker" (arXiv:2301.02111, 5 January 2023). McAfee Labs tested freely available tools that same year and got an estimated 85% match to the original speaker from three to four seconds of speech. The FTC put it plainly on 20 March 2023: a scammer needs "a short audio clip of your family member's voice", which "he could get from content posted online", plus a voice-cloning program. Only the quality has changed since.

Access is not gated. Consumer Reports assessed six voice-cloning products in March 2025 and found two, Descript and Resemble AI, putting meaningful barriers in front of cloning somebody else's voice. The other four asked for a tick-box confirming a legal right, then proceeded.

Your voicemail greeting tops the list: close-mic'd, uninterrupted, and usually carrying your full name in your own mouth. Next is the livestreamed wedding or funeral, years old and still public on a funeral home's site, carrying long-form speech under real emotion, which is what makes a distress call convincing. The image side is worse: a professional headshot is frontal, evenly lit, unobstructed, exactly what a face-swap system consumes.

Context is free as well: an obituary names everyone at once, in order, relationships spelled out.

FTC data released in June 2026 showed $3.5 billion reported lost to imposter scams in 2025, the most reported fraud category and close to one in three fraud reports. The FBI's 2025 Internet Crime Report logged 22,364 complaints citing the use of AI and nearly $893 million in losses, and the Bureau's framing is that the real number is higher because most victims never learn AI was involved. The FBI was still warning about cloned distress calls in June 2026.

Since you cannot reduce capability, supply or context, make the finished product fail at the moment somebody points it at your family.

Pick a word your household uses to confirm identity on a call, and never write it anywhere searchable.

The FBI's Internet Crime Complaint Center recommends exactly this (Alert I-120324-PSA, 3 December 2024, on criminal use of generative AI): "Create a secret word or phrase with your family to verify their identity." The Bureau repeated the advice on 15 May 2025, in a PSA on criminals impersonating senior US officials by text and voice, adding that you should independently identify a number for whoever contacted you and call to verify.

A clone can reproduce everything about you except a fact that exists only between you and your brother.

Choose a phrase with no digital footprint. Copper lantern is fine. The dog's name is not. Agree it out loud, with the five people who might take the call, never by text or shared note, because a compromised account must not compromise your family.

Then agree the rule and not only the word. Any call about money, an accident, an arrest, bail or an urgent transfer stops until the phrase is given, and you ask for the phrase rather than offering it first.

The callback habit carries equal weight, and the FTC's advice on the family-emergency scam skips the listening test: do not trust the voice, call the person who supposedly contacted you on a number you know is theirs. Caller ID stopped being evidence years ago, a voice stopped being evidence three years ago, and a number already in your phone still is.

If any account you hold uses your voice as an authentication factor, remove it where you can.

On 23 February 2023 the Vice reporter Joseph Cox walked through Lloyds Bank's Voice ID with a synthesised clone of his own voice and reached his balance and transaction history, playing "my voice is my password" back to a bank that says the system "analyses over 100 different characteristics" of a customer's voice. A free tool, five minutes of speech.

Your voice is not a password. Your voice is a public credential you cannot rotate.

Voice biometrics have improved, and good implementations add liveness detection and anti-spoofing. You have no way of knowing which one your bank runs, and the failure mode is your money. The same three questions cover brokerage lines, utility accounts, benefits helplines, and the IT helpdesk that resets passwords for anyone who sounds sufficiently senior and sufficiently annoyed.

Start building your legacy Free build, 50 memories, no card required.

As at July 2026 no federal US right lets you stop a person copying your voice, and what exists is aimed at commercial exploitation rather than at fraud against your mother.

The protections are jurisdictional and partial: sell a record using a clone of your voice in Tennessee and there is a claim, ring your mother in Ohio pretending to be you and the remedy is the fraud law that already existed. Identity law was written for people whose identity earns money, so the private-person right of publicity is rarely worth enforcing: your provable damages are small and the legal bill is not. And the most-hyped instrument does the least for you today, which what the NO FAKES Act would actually cover takes from the bill rather than the press release, while digital replica laws state by state covers each statute.

Europe took the transparency route instead of the ownership route. The AI omnibus, Regulation (EU) 2026/1744, in force on 27 July 2026, left Article 50's 2 August 2026 date alone and conceded only the December marking transition above. Making fakes declare themselves is not the same as owning a copy of your voice, and what EU AI Act Article 50 requires sets out where the line falls. Denmark has gone further with a bill giving every identifiable person a copyright-style right over their own likeness and voice, and as at July 2026 that bill is not law.

Then the removal side, where most people go first and the least is available. Every route below starts after the content exists and after you have found out.

All six are free. One carries a duty: the TAKE IT DOWN Act (Public Law 119-12) explicitly reaches digitally altered and AI-generated images, and the FTC enforces it through takeitdown.ftc.gov.

StopNCII.org is the practical companion, and "a hash will be sent from your device, but not the image/video itself." Only the hash travels, to Meta's Facebook, Instagram and Threads, TikTok, Reddit, Snap, X, Bluesky, Patreon, OnlyFans, Pornhub and others, with Microsoft applying validated hashes across Bing, OneDrive, Teams and Xbox, and Google using them in Search. The FAQ is explicit that synthetic images qualify: if the "'deepfake' or synthetic image is of you", you can hash the file. So are the limits: if an image is "edited through cropping, filters added or a video clipped, the original hash may not recognise the image", and encrypted messaging is out of reach.

YouTube publishes its rules, the clearest case of the discretionary route. The privacy complaint process says that "if someone has used AI to alter or create content that looks or sounds like you, you can ask for it to be removed", so voice is in scope, and then the conditions arrive: you or your legal representative must file, you must be "uniquely identifiable", and YouTube weighs "whether the content contains parody, satire or other public interest value". Its likeness detection tool went from a pilot with a subset of creators to politicians, government officials and journalists in March 2026, to talent agencies and the celebrities they represent on 21 April, to all eligible creators aged 18 or over from 16 May. You enrol with a government-issued ID and a selfie video, verification "may take up to 5 days to complete", and YouTube scans new uploads for your face. As at 29 July 2026 the tool detects visual matches only, the company says it aims "to extend Likeness detection to audio in the near future", and it states plainly that "YouTube will not remove all content".

Everywhere else you file under a platform's own policy. Screenshot with the full URL and a visible timestamp first, because content vanishes from view without being deleted, and report the fraud separately at IC3.gov and ReportFraud.ftc.gov. In the United States the Cyber Civil Rights Initiative runs a free Image Abuse Helpline, 1-844-878-CCRI (2274).

Neither table touches the call at twenty to midnight: no artefact to report, no platform to notify, and by the time you know the money has moved.

The threat is moving from being faked to not being believed.

A May 2026 listening study collected 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems. Accuracy at spotting fakes barely moved against a 2021 baseline, from 72.9% to 71.2%. Accuracy on genuine recordings fell from 72.7% to 64.1%. The authors call the result a skepticism shift: rather than getting better at hearing artefacts, listeners are getting worse at trusting real audio. Machine detectors held above 94.5% across all conditions, so the problem is human. A 2023 study at University College London had already found 529 listeners identifying deepfake speech only 73% of the time on a single clip, with familiarisation lifting accuracy by 3.84% on average.

So a genuine voice message from you will increasingly be doubted by default, and the price of that reflex is that your real voice loses standing alongside the counterfeit, the quieter harm covered in being remembered wrong.

The answer is provenance rather than detection. Under the European Network of Forensic Science Institutes best-practice methodology, a practitioner compares a questioned recording against a recording of a known speaker and expresses the result as a likelihood ratio, weighing how similar the two are against how typical that similarity is in a relevant population. With no known-speaker recording there is nothing to anchor the comparison to, and the analysis never starts. A bank fraud team asks the same question outside a courtroom: what does this person sound like, and who says so. A recording you made deliberately, timestamped, answers both halves. Clips scattered across other people's feeds answer neither.

ElevenLabs gates its Professional Voice Clone behind a voice captcha, the pattern worth borrowing: you read a prompted line aloud on the spot, and that recording is compared against the samples you uploaded, so a stranger cannot enrol somebody else's voice from downloaded audio. Note the gap: this is the same company Consumer Reports found accepting a tick-box on its quicker flow, so a verification step protects only the flow it is attached to. The mechanism is still right: a live spoken statement bound to one enrolment at one time is the difference between a file and evidence. Read the terms first, because consent and ownership are not the same thing, and who owns your voice clone sets out the licence language to look for.

Such a recording prevents nothing and triggers no takedown. It puts your real voice, your consent and the date on the record before anyone has to argue, so the burden of proof does not rest on your children at the worst moment of their lives.

That is where my own company enters, and the claim is narrow. Afterlife.ai™ does not stop scammers, nobody can, and any company that says otherwise is lying to you. What we run is enrolment: a guide called Idy interviews you in your own words, those interviews become memories, memories are what Personas are built from, and your voice arrives on a date with consent recorded against your account. Personas answer from the things you actually said, and a guard in the code strips any claim your own material does not support (an AI trained only on me covers the cost of that design). Release runs through the Executor Lock™ process, so a forged claim has to survive nomination by name, seven days, verification of the triggering evidence and a dispute window rather than one convincing phone call. While you are alive you hold the controls. Timeless AI™ is the same product without the legacy framing.

Three caveats from the person who built the thing. None of this reaches audio of you that is already public, and no version of our product ever will. Personas are versions of you rather than you, and Personas never invent what you did not say. And a recording held only inside one company's account is weaker than one you keep yourself, so ask every vendor what happens to your voice data if the company is sold.

You cannot take your voice back. What you can still decide is which recording counts as authoritative: the file a stranger assembled from a wedding video, or the one you made on purpose, on a date, with your consent attached.

Agree the word tonight. Then make the recording.

Start building your legacy Free build, 50 memories, no card required.

Can I stop my voice being cloned?

No. Three to four seconds of recorded speech is enough for a free tool to produce a usable clone, and audio already published cannot be recalled. You can reduce the supply by changing your voicemail greeting, unlisting old webinar recordings and asking for funeral or wedding livestreams to be made private. Then put your effort into verification, which is the part that still works.

Is a family verification word actually enough against a good fake?

Against the common scam, yes, because the scam depends on urgency and on you not verifying. A cloned voice reproduces how you sound, not what only you and your family know, and the FBI recommends exactly this measure in two separate public service announcements. The word fails in two situations: if the phrase has been written somewhere a compromised account exposes, and if the attacker already has access to your family's messages. The weak point is social rather than technical, because people feel unable to challenge a relative who is apparently crying. Agree in advance, out loud, that anyone may ask and nobody may take offence, keep the phrase memorised and unwritten, and pair the whole thing with hanging up and calling back.

What do I do in the first hour after finding a deepfake of myself?

Screenshot it with the URL and a visible timestamp, and save the link, before you report anything, because content often vanishes from your view without being deleted. If the fake is sexual, file a removal request with the platform under the TAKE IT DOWN Act and hash the file at StopNCII.org. If the fake is being used for fraud, report at IC3.gov and ReportFraud.ftc.gov. Then tell the people it could be sent to, because a warning that arrives before the fake does defuses most of the damage.

Does the NO FAKES Act protect me now?

No. As at July 2026, the NO FAKES Act (S.4591, 119th Congress) is a bill. It was reintroduced on 20 May 2026 and ordered reported out of the Senate Judiciary Committee on a unanimous vote on 18 June 2026. It has not passed the full Senate, has not passed the House, and is not law. Until it is, your protection depends on your state's publicity and fraud law.

Should I turn off voice authentication at my bank?

If it is offered as a convenience and you have an alternative, yes. A reporter bypassed a major UK bank's voice ID with a consumer voice tool in February 2023, and you have no way to audit which anti-spoofing measures your own bank runs. Call and ask whether voice recognition is enabled, whether it can be removed, and whether the fallback can be locked to a passcode.

I am not in the United States. Does any of this apply?

The verification word, the callback habit and the voicemail change apply everywhere. StopNCII.org is international. The TAKE IT DOWN Act's 48-hour duty is United States law, though the large platforms tend to run one global process. In the EU, Article 50 of the AI Act requires deepfakes to be disclosed from 2 August 2026. Elsewhere, check your national law on non-consensual imagery, fraud and harassment first, because that is usually the fastest route.

• How to prove a voice is really you, the other half of this page.

• Is voice cloning safe, on cloning your own on purpose.

• How to create an AI version of yourself and the plans.

Corrections: tell us if anything here is wrong or out of date.

Built from primary sources wherever one was reachable: government consumer alerts and FBI public service announcements, the FTC and FCC's own published rules and rulings, bill and committee records, the Article 50 text and the omnibus amendment that adjusted its marking deadline, statute text for state law, peer-reviewed and preprint listening studies, and the platform help pages that set out each takedown route. Where a claim rests on a company's own policy, the policy page was read on 29 July 2026 and dated as such.

1. Federal Trade Commission, Alvaro Puig, Scammers use AI to enhance their family emergency schemes, consumer alert, 20 March 2023.

2. FBI Internet Crime Complaint Center, Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud, Alert Number I-120324-PSA, 3 December 2024.

3. FBI Internet Crime Complaint Center, Public Service Announcement I-051525-PSA, 15 May 2025, on malicious actors impersonating senior US officials by text and voice message.

4. Federal Trade Commission, FTC Data Show People Reported Losing $3.5 Billion to Imposter Scams in 2025, press release, June 2026.

5. Federal Bureau of Investigation, Cryptocurrency and AI Scams Bilk Americans of Billions, on the IC3 2025 Internet Crime Report (22,364 AI-related complaints, nearly $893 million in losses, largest category investment fraud), 2026.

6. FBI warns of AI voice-cloning scam that mimics loved ones in distress, reporting an FBI warning, 2 June 2026.

7. Chengyi Wang, Sanyuan Chen, Yu Wu and colleagues, Microsoft, Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E), arXiv:2301.02111, 5 January 2023.

8. McAfee Labs, Beware the Artificial Impostor, 2023 (three to four seconds of speech produced an estimated 85% voice match).

9. Kimberly T. Mai, Sergi Bray, Toby Davies and Lewis D. Griffin, University College London, Warning: humans cannot reliably detect speech deepfakes, PLOS ONE, 2 August 2023 (529 listeners; 73% accuracy in the single-clip condition, 70.35% overall; familiarisation added 3.84%).

10. Nicolas M. Müller and Wei Herng Choong, Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception, arXiv:2605.26136, preprint submitted 21 May 2026 (35,532 judgments, 1,768 participants, 138 text-to-speech and voice conversion systems; fakes 72.9% to 71.2%, genuine 72.7% to 64.1%, machine detectors above 94.5%).

11. Consumer Reports, assessment of AI voice cloning products, March 2025.

12. TechCrunch, Consumer Reports finds popular voice cloning tools lack safeguards, 10 March 2025.

13. Joseph Cox, Vice, How I Broke Into a Bank Account With an AI-Generated Voice, 23 February 2023.

14. Federal Communications Commission, FCC Makes AI-Generated Voices in Robocalls Illegal, Declaratory Ruling FCC 24-17, adopted 2 February 2024 and released 8 February 2024.

15. Federal Trade Commission, Rule on Impersonation of Government and Businesses, 16 CFR Part 461, effective 1 April 2024.

16. Federal Trade Commission, FTC Proposes New Protections to Combat AI Impersonation of Individuals, 15 February 2024. Informal hearing held 17 January 2025; not finalised as at July 2026.

17. Tennessee Ensuring Likeness Voice and Image Security (ELVIS) Act, signed 21 March 2024, effective 1 July 2024, amending Tenn. Code Ann. Title 47 Chapter 25 Part 11. Analysis: Holland and Knight, First-of-Its-Kind AI Law Addresses Deep Fakes and Voice Clones, April 2024 (Class A misdemeanour plus a private right of action; "voice" defined as a sound "readily identifiable and attributable to a particular individual", simulations included).

18. California AB 1836 and AB 2602, both signed 17 September 2024; AB 2602 applies to contracts for performances entered into on or after 1 January 2025. Analysis: Fenwick, California's New AI Laws Limit Uses of Digital Likeness, 2024.

19. Revised Code of Washington 63.60.040, right exclusive for ten years after death for an individual and seventy-five years for a "personality".

20. SC Daily Gazette, Officials praise new law criminalizing 'revenge porn.' SC was the last state to ban it, 29 May 2025 (H.3058 signed 12 May 2025; covers AI-generated images).

21. S.146, TAKE IT DOWN Act, 119th Congress, Public Law 119-12, signed 19 May 2025; scope limited to intimate visual depictions.

22. Federal Trade Commission, FTC Begins Enforcing TAKE IT DOWN Act, 19 May 2026, on the Section 3 notice-and-removal requirements.

23. Covington and Burling, The TAKE IT DOWN Act's Notice and Removal Requirements Enter Into Effect, 2026 (48 hours, reasonable efforts on known identical copies, FTC enforcement, TakeItDown.ftc.gov).

24. StopNCII.org, How StopNCII.org works, FAQ and partners, policy and partner list checked 29 July 2026.

25. Cyber Civil Rights Initiative, Nonconsensual Distribution of Intimate Images, checked 29 July 2026.

26. S.4591, NO FAKES Act of 2026, 119th Congress, introduced 20 May 2026, ordered reported by the Senate Judiciary Committee on a unanimous voice vote 18 June 2026. Not law as at July 2026.

27. EU AI Act Article 50, Regulation (EU) 2024/1689, transparency obligations for providers and deployers of certain AI systems, applies from 2 August 2026. Summary of the same obligations, including the marking transition: transparency rules under Article 50. On the AI omnibus, Regulation (EU) 2026/1744: Gibson Dunn, EU AI Act Omnibus Agreement, 2026.

28. Schjødt, Owning the Self: Denmark's Copyright Turn Against Deepfakes (proposal unveiled 26 June 2025, not adopted), checked 29 July 2026.

29. YouTube Help, Likeness detection on YouTube (visual matches only; ID plus selfie video, verification up to five days; YouTube states it aims to extend likeness detection to audio in the near future, and that it will not remove all content), policy checked 29 July 2026.

30. YouTube Official Blog, Expanding likeness detection to the entertainment industry, April 2026.

31. TechCrunch, YouTube expands its AI likeness detection technology to celebrities, 21 April 2026 (pilot began with a subset of creators; politicians, government officials and journalists added in March 2026).

32. MediaPost, YouTube Rolls Out Likeness Detection To All Creators Over 18, 19 May 2026 (rollout confirmed 16 May 2026).

33. YouTube Help, Protecting your identity (privacy complaint process covering AI-altered or AI-created content that looks or sounds like you), policy checked 29 July 2026.

34. European Network of Forensic Science Institutes, Best Practice Manual for the Methodology of Forensic Speaker Comparison, 2021 (questioned recording against known-speaker recording, likelihood-ratio framework).

35. ElevenLabs, Professional Voice Cloning documentation (voice captcha: the speaker reads a prompted line aloud and the recording is matched against the uploaded samples), checked 29 July 2026.

Still unsettled: the NO FAKES Act has cleared the Senate Judiciary Committee and nothing more, so it may change or die; the FTC's proposed extension of its impersonation rule to individuals has not been finalised; Denmark's likeness right is a proposal rather than law; and YouTube's likeness detection covers faces, not voices, with audio promised and not shipped. Detection accuracy figures come from laboratory listening tests and should not be read as what a frightened person on a phone call can manage.

Last reviewed: 30 July 2026. Legal, regulatory and platform-policy claims on this page are on a 90-day recheck cycle. The NO FAKES Act status, the Danish likeness bill and the EU AI Act Article 50 marking transition are rechecked every 30 days until they settle.