Listening closely can give us information. It cannot carry every decision on its own. To see why, we need to ask what a detection score actually measures.
One test asks whether an audio clip is synthetic. Another asks whether two voices belong to the same person. A third might ask whether the person making a request has authority to make it. These tasks overlap, but they are not interchangeable. A system that answers the first has not automatically answered the other two.
Barrington and colleagues made part of this distinction explicit in a 2025 study. Separate groups judged identity resemblance or whether clips sounded human. The much-cited 83.3% was the median share of real-and-clone pairs judged to be the same speaker. It was not a fraud success rate. In the separate classification task, average correct judgments were 67.4% for genuine clips and 60.8% for synthetic ones. The listeners were not testing the voices of their own friends or relatives. Barrington and colleagues, 2025
A voice may therefore resemble a person while still sounding artificial, or sound fully human without proving whose voice it is. This is one reason to resist a single headline score for trust.
A calculation we can inspect
For this brief, I recalculated the aggregate counts in Mai and colleagues’ 2023 listening experiment. This was a small, reproducible analysis of published tables. It did not involve new listeners, recordings or a rerun of a voice model.
The study involved English- and Mandarin-speaking participants. In one task, they judged single clips as real or fake. In another, they chose the fake from a pair. These were explicit detection tasks using selected speech from two female source identities. The four counts in the single-clip table were:
| What the clip was | Called real | Called fake | Total |
|---|---|---|---|
| Genuine | 1,761 | 837 | 2,598 |
| Synthetic | 681 | 1,841 | 2,522 |
| Total judgments | 2,442 | 2,678 | 5,120 |
Source: published Table 3; totals and percentages independently recalculated for this brief. These are judgments, not different people. Mai and colleagues, 2023
The correct judgments are the 1,761 judgments of genuine clips called real plus the 1,841 judgments of synthetic clips called fake. That gives 3,602 out of 5,120, or 70.35% overall accuracy. The often-repeated 73.00% answers a narrower question: what share of the synthetic-clip judgments were correct? Its denominator is 2,522 synthetic judgments, not all 5,120 judgments.
The other side deserves equal attention. Listeners called genuine clips fake in 837 of 2,598 judgments, a 32.22% false-alarm rate. They missed synthetic clips in 681 of 2,522 judgments, a 27.00% miss rate. A single accuracy figure hides the difference between these errors.
The arithmetic also exposed a limit in what can be reconstructed. The listed single-clip participant groups imply 5,200 judgments if each supplied twenty. The published result tables contain 5,120, a gap of eighty. I could not establish the cause. The percentages above use the actual published cells, without filling the gap or inventing an exclusion rule. The separate paired-task counts reconcile to 5,380 judgments, with 85.59% correct.
Because each listener made repeated judgments, these observations are not independent people. I have not attached confidence intervals that pretend otherwise. Participant-level files were not retrieved during this review. The result is an audit of published aggregates, with an unresolved reporting gap, rather than a new estimate from raw participant data. The methods note, aggregate counts, offline calculation and results make these choices inspectable.
What should change after seeing the table?
The table gives no estimate of how many calls in your life are fake. Nearly half its single-clip judgments concerned synthetic speech because of the study design. That mixture cannot be treated as a rate of fraud in ordinary life.
It also shows why we should care about the cost of a false alarm. If we respond to every suspected fake by cutting off contact, genuine people will bear some of that cost. A worried relative may be refused help. A person using a communication aid may be treated as dishonest. The better response to an uncertain signal is usually a way to check, not a final judgment about the person’s motives.
The higher score on pairs is worth understanding too. Having both a real and a fake version to compare changes the task. An incoming caller rarely provides a matched, genuine recording of the same words. A gain on the paired test cannot be assumed to appear on an ordinary call.
Nor does one brief familiarization exercise prove that teaching cannot help. Mai’s familiarization used five labelled examples. That narrow intervention is different from practicing how to find a known contact or check a payment request. The evidence gives us no reason to abandon those separate questions.
These are historical experimental results. Newer voices, different languages and new listening conditions may produce different errors. The point of reproducing the counts is to understand the evidence accurately, not to turn a 2023 result into a permanent limit on human ability.
Automated detection deserves a fair test
Computer detectors have also faced problems when their test setting changes. Müller and colleagues showed this in 2022 by comparing performance on a standard benchmark with a set of real and synthetic public-figure recordings gathered from the world outside that benchmark. Some detectors’ errors rose sharply. The study demonstrated a transfer problem under that design; it did not prove that all later detectors would fail. Müller and colleagues, 2022
There has since been substantial progress. Huang, Mao and Qian reported a mean equal error rate of 1.65% across eight public test sets for their larger detector in 2026. Its hardest set had a 6.47% rate. An equal error rate summarizes a threshold where the false-alarm rate and miss rate are equal. It is not the chance that a particular flagged caller is fake. Their separate commercial-generator test used synthetic clips only, so that test could not tell us how often genuine speech would be wrongly flagged. Huang, Mao and Qian, ACL 2026
These findings rule out an easy story in which detectors are simply useless. Good tools may add valuable evidence. The question is whether a tool has been tested for the voices, languages and channels where it will be used, with both kinds of error reported. A provider should also explain what happens after a flag. Does a human review the request? Is there a route for a genuine caller to resolve the problem?
Even a strong detector cannot make the whole decision. A genuine human voice can deliver a false request. A synthetic voice can carry an honest one. Audio screening and permission to act should therefore remain separate checks.
There is a privacy choice as well. Uploading a family recording to an unfamiliar service sends another person’s voice and perhaps private conversation to that service. Before using such a tool, check whose permission is needed, what is retained and whether the same question can be resolved through direct contact. A detection label is not worth collecting more private material than the decision requires.
We should use better detection where it earns its place. We should also build a way forward when it is wrong or unavailable. The aim is a decision process that can survive an imperfect signal.