The law knows what a cloned voice is. Nobody has said what proof looks like.

Start with what has actually been decided, because it is more than people assume. In February 2024 the communications regulator ruled that voices generated by machine are artificial voices within the meaning of the telephone statute - they are artificial, in the ruling's own words, because a person is not speaking them.¹ That settles the character of the thing. Read the ruling looking for a method and there is none: no detection technology named, no accuracy threshold set, no guidance to a carrier or to a called party on how to tell. Seven months later the same agency issued the largest penalty this field has produced, six million dollars, and the violation found was transmitting misleading caller identification with intent to defraud. The synthesis was the context of the case, not the element proved.² The biggest number in this area was won on a provision that predates the technology entirely.

Then look at what a court actually requires, because almost everyone gets this backwards. Authentication under Rule 901(a) asks only for evidence sufficient to support a finding that the item is what its proponent claims. That is conditional relevance under Rule 104(b): the judge does not decide whether the recording is genuine, the judge decides whether a reasonable juror could find it genuine, and the question then goes to the jury as weight.³ A party who says a recording is a clone is, in the ordinary case, making a jury argument with no instrument in the room. In March 2026 a federal court met exactly that argument - a litigant said recordings were either someone else entirely or manipulated audio made by artificial intelligence to clone his voice - and resolved it with Rule 901(b)(5), an opinion identifying a person's voice, on a declaration from someone who had spoken with him many times before. No detection tool appeared anywhere in the decision.⁴ A provision written in 1975 is currently doing this work, which is a real answer and a narrow one: it needs somebody who already knows the voice.

Where a detector's output does come in through a qualified witness, one rule already governs it, and it was tightened recently. The expert evidence rule was amended effective December 1, 2023 to state that the proponent must demonstrate to the court that it is more likely than not that the reliability requirements are met. The committee note is unusually blunt about why: many courts had treated the sufficiency of an expert's basis and the application of the method as questions of weight, and the note says those rulings are an incorrect application of the rules. The same note then turns to forensic witnesses directly, telling them to avoid assertions of absolute or one hundred percent certainty and telling judges they should, where possible, receive an estimate of the known or potential rate of error based on studies of how often the method produces accurate results.⁵ The second change rewrote the rule so that an opinion must reflect a reliable application of the method, which is aimed precisely at the gap between what a tool measured and what a witness says it means. A likelihood score is not a verdict that a recording is a clone.

The judiciary has in fact written rules for this problem. It has advanced neither. A draft addressed to evidence fabricated by generative artificial intelligence exists, and has never been published for comment - held in abeyance in November 2025 and deferred again on May 7, 2026, with the committee stating that despite extensive commentary few examples exist of courts having to address the possibility of deepfakes.⁶ A second rule, on machine-generated evidence offered without an expert, did enter the process: published August 15, 2025, two public hearings in January 2026, comments closing February 16, 2026 - and then in May 2026 the committee declined to recommend action, revised the text, and acknowledged the revision would require re-publication.⁷ The reason for the pause is a survey the committee commissioned of the federal bench. Nine hundred and thirty-one judges responded, a forty-five percent response rate. Ninety-eight percent had never encountered a deepfake challenge. Fifteen had. Of those fifteen, five thought the rules should be amended.⁸ The two drafts of the machine-evidence rule differ in what they cover and what foundation they demand, so a quotation of it means nothing without a date attached.⁹ And the judiciary's own list of pending amendments settles the question of timing: two evidence amendments are in the pipeline, with effective dates in 2026 and 2028, and neither is one of these.¹⁰

Now the measurement side, which is where the gap is widest. The federal speaker recognition evaluation has run since 1996. Its task, in the current plan's own words, is to determine whether a target individual is present in a test segment - a same-person question. Whether a segment came from a machine is a different question, and we searched the evaluation plan and the program pages for spoofing, synthetic and deepfake and found none of them.¹¹ That distinction matters more than it sounds: a cloned voice is designed to pass a same-person test, so a speaker recognition result is not evidence of authenticity. The federal generative AI evaluation program lists the modalities it supports and, as of its own March 2026 briefing, they are text, image, video and code. Audio is absent, and the single completed pilot was text only.¹² There is a federal overview of technical approaches to synthetic content, and its audio detection figures are not federal measurements at all - they are drawn from a published literature review, because no federal evaluation of synthetic speech detection has been run. The range that report gives for reported accuracy is fifty percent to well above ninety, depending on method and dataset. The bottom of that range is a coin toss.¹³

One peer-reviewed study shows what that range means in practice better than any summary. Twelve detection architectures were tested on the benchmark they were built for and then on thirty-eight hours of real-world audio from fifty-eight public figures. The best model went from an equal error rate of 1.229 percent to 37.154 percent - roughly a thirtyfold increase, on the same metric, with nothing changed but the audio. The authors' own conclusion is that the community has tailored its solutions too closely to the prevailing benchmark and that deepfakes are much harder to detect outside the lab than previously thought.¹⁴ The metric built specifically for this problem, which scores a countermeasure jointly with the verifier it protects, came out of an academic challenge series and no federal agency has adopted it. Federal guidance itself warns that accuracy, area under the curve and F1 do not capture what matters here, because false positives - calling a real person's voice synthetic - can be extremely damaging.¹⁵

Which leaves the question of who certifies any of this, and the answer is nobody. The national forensic standards organization restructured on October 1, 2025, cutting twenty-two subcommittees to nineteen, and one of the three it closed was Speaker Recognition - the only body in the structure whose subject was the human voice.¹⁶ We took the complete registry export and counted terms across every field of all 262 standards: synthetic, zero; deepfake, zero; artificial intelligence, zero; voice cloning, zero; spoof, zero; authentic in any form, zero. What is there instead is a guide to comparing two recordings of a person, a protocol for collecting a speech sample, and a guide to laboratory setup.¹⁷ And the standard a method is supposed to meet was set out a decade ago in terms this field has not yet met: foundational validity can only be shown through empirical studies, and certification programs, accreditation programs, standardized protocols, proficiency testing and professional consensus cannot substitute for it. Nothing can substitute for it.¹⁸ That report evaluated six methods and voice comparison was not among them, which is worth stating plainly rather than stretching.

Two last things a buyer should know. The trademark register holds no application and no registration for this exact wording, and no board proceeding involving it - which is a fact about an index on a date, not a clearance, and not an opinion about whether anyone may use it.¹⁹ And there is no federal market figure for this category, so any number a buyer is shown for it is a private model.²⁰ A settled harm, an unsettled method, two rules in a drawer, thirty years of measurement aimed at a different question, and a registry of 262 standards that never names the thing. Somebody has to read all of that before building here, and that is the business this name is for.