The liar’s dividend

The half of this problem that gets least attention. Fabricated audio is a threat; the ability to call real audio fabricated is a larger one, and it is already being used.

The mechanism is a shift in who has to prove what

A recording used to be close to self-authenticating in ordinary life. If a voice was on the tape saying the thing, the argument was largely over, and anyone denying it was denying something everyone could hear.

That default has gone, and its removal is not symmetrical. The person holding a recording must now support it — with the original file, a device, a chain of custody, corroborating records. The person denying it has to do nothing but gesture at a general capability that genuinely exists and is widely reported. One side acquires an evidential burden; the other acquires a free move.

The cost falls hardest on people whose account rests on a single recording and who have the fewest resources to support it. A large organisation can commission an examination. A care worker with one voicemail cannot, and the denial costs her opponent nothing. [VERIFY: cite the legal and policy literature where this term is established, with a full reference]

A concrete case

A care worker records her manager instructing her to backdate a medication log. She raises it internally. The manager does not dispute what the recording says; he says the recording was made by a computer, and points to a news segment about cloned voices from the previous month.

Nothing further is needed from him. The panel is now weighing a technical question nobody in the room can answer against a witness they were previously inclined to believe. Even a clean analysis result does not end it, because a clean result is a probability rather than a proof, and he can say so.

What actually helps her is unglamorous and mostly assembled before any of this: the original file left untouched on the phone that recorded it, the rota showing they were both on shift, the note she wrote that evening, the colleague she told. Our page on recordings produced in disputes sets that groundwork out in the order it is needed.

Detection can make this worse, and usually does when it is oversold

This is the uncomfortable part for a company that sells detection. Every claim of near-perfect accuracy raises the standard a genuine recording is expected to meet, and every claim that turns out to be overstated hands the denier a better argument next time.

Three failure modes recur. A tool that reports unclear is quoted as “even the analysis could not say it was real”. A tool with an unpublished error rate produces a result nobody can weigh, so its authority collapses under any pressure. And a false positive on genuine audio does not merely fail — it produces a document supporting the false claim, which the denier can then wave.

That last one is why we treat the false positive rate as the number that matters most. A detector that wrongly flags real speech is not neutral in this fight; it is on the other side.

Commonly confused with: deepfake audio

Deepfake audio is about recordings that were fabricated. The dividend is about recordings that were not. The two are opposite failures of the same collapse in default trust, and confusing them leads organisations to prepare only for the first.

It is also more than ordinary denial. People have always denied recordings, usually by disputing context or claiming an edit. What is new is that the denial is now cheap and pre-credentialed: the audience has already read that convincing fakes exist, so no evidence has to accompany the claim. Nor is it the same as provenance, which tries to close the gap from the other end and cannot, since most genuine recordings will never carry credentials.

What an honest detector can do about it

Less than people want, and something real. It cannot prove a recording is authentic; absence of generation evidence is not proof of capture, and we say so on every result. What it can do is turn a vague argument into a specific one, provided the numbers behind it are public.

A verdict is worth something only if you know how often the system is wrong in each direction, on audio of the kind you have. That is why the accuracy page reports both error directions by recording condition with sample sizes, why the engine version is stamped on every report so it can be checked against the changelog, and why likely human is described as the weaker verdict rather than sold as a clean bill of health. A published error rate is a defence against the dividend. An unpublished one is fuel for it.

Questions this term raises