Consent Before the Clone: Why Voice Cloning Is a Governance Risk We Can Mitigate Now
Credit: Tesfu Assefa

Consent Before the Clone: Why Voice Cloning Is a Governance Risk We Can Mitigate Now

Michael T. Neely argues that in the age of deepfake AI, a few seconds of audio can turn a familiar voice into a tool for fraud. But his experience in the music industry suggests that consent checks, legal protections and public awareness could limit the damage.
Nov 17, 2026
6 min read
0.00 MPXR
This article is part of Mindplex’s Public Call Series on AI Risk & Mitigation. The series convenes rigorous perspectives on the challenges, governance, and societal implications of advanced AI.

A convincing copy of a person's voice no longer requires a studio, a skilled impersonator, or even the person's cooperation. It requires seconds of audio—a podcast clip, a live stream, a song—and a tool that anyone can open in a browser. That is why voice cloning belongs in a different category from most of the risks debated under the heading of artificial intelligence. It is not a capability forecast. It is a present-day governance problem, already running through the music business, already used for fraud and impersonation, and—this is the part that deserves more attention—already mitigable with tools we have.

I come to this as a publisher who has spent more than forty years in the music business, and more recently as a witness and filer in the rooms where this is being decided: testimony to the Georgia Senate's study committee on AI in the creative industries, a practitioner case study on cloned-voice consent included in the UN Internet Governance Forum Policy Network on AI's work, and comments to the Federal Trade Commission and the U.S. Copyright Office. From those rooms, the picture is consistent. The technology is not waiting for permission, but governance does not have to wait for catastrophe either. Three mitigations are available now: procedural consent standards, state-level statute, and public literacy.

The risk, stated plainly

A voice clone does its damage in circulation, not in creation. A fabricated recording of an artist "releasing" a song they never made, a cloned executive authorising a payment, a fake endorsement in a singer's own voice—each one travels faster than any correction, and each one is cheap to make and expensive to answer. In my Copyright Office comment, "Verify Before You Pay," I addressed the adjacent problem in streaming: fraudulent synthetic content is not only published, it is monetised, paid out by systems that never asked whether the voice was authorised. The pattern is the same across these cases. At the point of upload and the point of payment, nobody is required to establish that the person whose voice is being used ever agreed to it.

It is tempting to treat this as a problem of bad actors, and bad actors certainly exist. But a governance lens shows something more useful: impersonation at scale is a predictable output of systems built without a consent step. Where verification is absent, cloning without consent is not an abuse of the system. It is the system, working as designed. That is good news, in a way—because systems can be redesigned.

Credit: Tesfu Assefa
Credit: Tesfu Assefa

Mitigation one: consent as a procedure, not a slogan

Everyone in this debate says they favour consent. The disagreement is over what consent means in practice. A term buried in a platform's conditions is not consent in any sense a performer would recognise. A workable standard is procedural, and it has three parts.

First, authorisation before cloning: a service should not clone an identifiable person's voice without that person's verified agreement—verify before you amplify. Second, provenance that survives sharing: synthetic content should carry a record—that it is synthetic, who made it, and on whose authority—that persists when the file is downloaded, clipped and reposted, because a label that falls off at the first share is transparency in form only. Third, contestability: the person cloned must have a working channel to challenge a replica and get a decision from a named, accountable human, with deadlines and real remedies—removal, correction, demonetisation. A reporting form that produces no answer is not a remedy; it is a suggestion box.

None of this requires new science. It requires the consent step to be placed where the leverage is: at the cloning service, at the point of upload, and at the point of payment.

Mitigation two: statute, starting where law already moves

Federal action on digital replicas has been proposed—the NO FAKES Act is the leading example—but the law that exists today is being written in the states. Tennessee's ELVIS Act, updated in 2024, was the first statute to protect voice explicitly as an attribute of the person, and it did so because the music industry in Nashville understood the threat concretely and early. My case study for the Policy Network on AI examined the ELVIS Act as a working model: what it covers, where its limits are, and what other jurisdictions can borrow from it.

In Georgia, I testified before the Senate study committee examining AI's impact on the creative industries, and later asked to speak at the state's 21st Century Community Forums for AI Progress. The argument I made there is the one I would make anywhere: a state statute cannot govern the whole internet, but it can do three things now—name the right (your voice is yours), name the wrong (commercial cloning without consent), and name the remedy (a route to takedown and damages that an individual artist can actually use). States are also where the evidentiary record gets built. Every hearing forces the question that abstract debate avoids: not whether voice cloning is powerful, but who answers when it is used against someone.

Mitigation three: public literacy

The third mitigation is the least technical and the most neglected. Most people still hear a familiar voice and believe it. That instinct—reasonable for all of human history—is now an attack surface. Public literacy about synthetic media is not a call for universal suspicion; it is a call for a new default: a voice recording is a claim, not a proof, and verification is a normal step rather than an insult. Newsrooms are learning this. Banks are learning it, expensively. The music business, where a fake release can chart before the real artist wakes up, is learning it fastest of all.

Literacy also has a governance function. Rules are easier to pass and harder to evade when the public understands the harm in concrete terms. The constituents who moved Tennessee were not technologists; they were artists, families and fans who understood immediately what it means when a voice—including the voice of someone who has died—can be made to say anything.

A risk we can mitigate now

Voice cloning will be used to make remarkable art, with permission, by artists who choose it. Nothing in this argument is against the tool. The argument is about sequence: consent before the clone, provenance with the content, and a person who can answer back when the system

gets it wrong. Those are not speculative safeguards for a hypothetical future. They are procedures, statutes and habits that exist today in at least prototype form, in at least some jurisdictions, ready to be generalised.

Governance debates about AI often stall between two unhelpful poles: the promise that the technology will govern itself, and the fear that nothing can govern it at all. Voice cloning refutes both. The risk is specific, the victims are identifiable, the mitigations are practical, and the first drafts of the law are already on the books. What is needed now is not more alarm. It is more adoption.

About the Writer

More from Mindplex

Keep reading

Three more ideas worth your time.

Browse Magazine

Discussion

Join the discussion

Sign in to share a response with the community.

Type @ to mention someone Type / or use + to add a block Highlight text, then choose Link
Loading editor

Comments cannot be edited after posting because they become part of the reputation record. Give yours a quick review first.

ab

aaron bergman

1 hour ago

Your call for provenance that survives sharing is doable. Though the technology might not exactly frame it, I see a concrete, deployable tech stack which can turn your governance principle into reality.

The best approach combines audio watermarking and a cryptographic manifest. This method uses an invisible watermark, like Meta’s AudioSeal, which is embedded into the audio file itself. This means the watermark sticks around even after the audio is compressed or altered, so it can still be tracked when shared or reposted. The accompanying C2PA manifest acts like a digital fingerprint. It’s a secured record that includes details like the model used, the time of creation, and proof that the user allowed the sharing.

This two-part system tackles the problems mentioned in the article, since a single watermark can be removed and a manifest might not always go with the file. Together, they create a reliable way to verify the audio and maintain its history, even in the chaotic world of Deepfake AI infested social media.