Can people tell an AI voice from a human? What the research says

Short version: in 2026, if a voice has been cloned from a real person, most listeners cannot tell it from the real thing. A voice generated from scratch, with no human original behind it, still gets spotted most of the time. So “can people tell?” has two answers, and the one that matters depends on which kind of voice you mean.

What the studies actually found

Two peer-reviewed studies from 2025 put real numbers on this. Start with the clearest. Researchers at Queen Mary University of London (Lavan and colleagues, published in PLOS ONE on 24 September 2025) played 50 listeners a mix of real human recordings, AI clones of real people, and generic AI voices built from a large model with no specific person behind them. The question was simple: human, or machine?

Look at the top two rows. A clone landed at 70%, a real recording at 72%. That gap sits inside the statistical noise, so for any practical purpose the clone is indistinguishable from the person it copied. The generic voice, with no human original, came in at 39%, caught more often than not. The same study found no “hyperrealism” effect: clones match real voices, they do not beat them yet.

So where did “80% can’t tell” come from?

You will see “80% of people cannot tell AI from a human voice” quoted all over the place. It is a real number aimed at the wrong claim. It comes from a separate study (Barrington, Cooper and Farid at UC Berkeley, in Scientific Reports, 2025). What they actually measured: when listeners heard an AI clone, they matched it to the real person’s identity about 80% of the time. They correctly flagged a voice as AI only about 60% of the time. So the 80% means “this clone sounds like that person”, not “nobody can tell it is AI”. Close enough to mislead, different enough to matter. We cite the 70 / 72 / 39 split instead, because it answers the question buyers are actually asking.

The line that matters: cloned, not generic

If you take one thing from the research, take this: the threshold that has been crossed is voice cloning, not AI voices in general. Clone a real person from a clean recording and you get something most people accept as that person. Spin up a generic voice from a model and you are still in uncanny territory more often than not. The Berkeley work was prompted by the January 2024 robocall that faked President Biden’s voice to thousands of New Hampshire voters. That was a clone of a specific person, which is exactly the case the studies say works.

One more finding worth flagging. In the Queen Mary study the AI voices were rated more dominant than the human ones, and the generic AI voices were rated slightly more trustworthy (57.9 against 53.8 on their scale). Sounding human is not the same as sounding like you, and neither is the same as being trusted. Three different things, often confused.

The catch nobody puts on the box

Here is the quieter finding buyers should not miss. The off-the-shelf voices platforms ship are narrow. A 2025 study presented at the ACM FAccT conference tested synthetic voice services and found that 61.5% of participants with non-American or non-British accents felt they were not represented. For one leading vendor’s stock voices, accent recognition was poor: 12% for Indian-accented English, 10% for African, and 0% for Australian. So “indistinguishable from human” holds mostly for the accents these systems were built around. If your callers are in Mumbai or Lagos, the demo voice may not land the way it did in the sales call.

What this means if you are buying a voice agent

First, judge the voice on your own script, in your callers’ accents, not the vendor’s demo reel. The research says the cloned, accent-matched case is the strong one. A generic voice in an accent the model barely covers is the weak one. Our naturalness score on each platform is a starting point (ElevenLabs leads the pack we have heard, and you can see how the field scores on the rankings), but your own ear on your own copy beats any score.

Second, if you run outbound calls, the “can they tell” question is partly out of your hands now, because the law has a view. In the United States the FCC ruled in February 2024 that AI-generated voices in robocalls count as artificial and are illegal without prior consent. In the EU, Article 50 of the AI Act, enforceable from 2 August 2026, says a voice agent has to tell the person it is not human. The honest move and the compliant move point the same way: disclose it.

Third, do not buy on “it sounds human” alone. It probably does now. The questions that still separate platforms are the ones we test elsewhere: the latency (the pause before it answers), the real all-in cost per minute, and whether it can hand a live call to a person. Naturalness has largely been solved. The rest has not, and that is where the money decision actually sits.