One sentence is enough

The title is my takeaway from a new study by security researchers at Brno University of Technology, published in August. They tested how well people catch AI voices by ear. Spoiler, it doesn’t go well!

The test

They cloned nine famous voices, Morgan Freeman, Adele, Tom Holland, and six others, each from 3 minutes of YouTube interviews. They used three tools from three different years: Real-Time Voice Cloning from 2019, YourTTS from 2022, and ElevenLabs from 2024.

Then they played the clips to 82 IT professionals. People who work in tech, in a quiet room, with headphones on. They were told up front that fakes were in the set, and they could replay the real voice any time they wanted. It’s hard to imagine a better setup for a listener. In real life, nobody warns you.

Each clip had four sentences. In some, all four were made by AI. In others, they took a real interview and swapped just one sentence for an AI one.

Old fakes are easy, ElevenLabs is a coin flip

With the older tools, people did great. They caught the fakes about 9 times out of 10. Robotic tone, odd intonation, dropouts, etc.

With ElevenLabs, it flipped. People spotted only 43% of the fake sentences. A coin flip, from tech people who knew the fakes were coming. And keep in mind, these are their voices from 2024.

One sentence inside a real interview

This is the part that you should read two times.

When they placed one ElevenLabs sentence inside a real interview, listeners accepted it as real 77% of the time. The real speech around it seems to have helped it blend in.

People were not confused in general. They still judged the real sentences right 89% of the time. They just missed the fake one because the real voice around it made it believable.

Now think about what one sentence can do. A quote in an interview. An election campaign. A piece of evidence. A voice message that sounds exactly like your boss. Nobody needs to fake the whole thing when they can change the one sentence that matters.

The detectors struggled too

You’d think deepfake detectors would catch what people miss. The team ran the same clips through six research detectors. On the fully fake ElevenLabs clips, the detectors scored 0.09, against 0.48 for people (F1 score, 1 is perfect). The models did worse than the humans. On the swapped sentence clips, not a single detector got even one clip fully right.

The detectors were trained on academic datasets and fell apart on a commercial voice they had never heard. The authors say it plainly, detection alone is brittle.

Real voices get called fake too

There’s a second problem, and I think it’s worse. The same listeners marked 17.5% of the real clips as fake.

Part of that comes from the setup. People knew fakes were around, so they got suspicious. But that is exactly where we are all heading. Once everyone knows voices can be faked, everyone starts to doubt reality. And you can already see this on X, where people are '@grok is this real.'

The paper calls it the liar’s dividend. Real audio is getting distrusted. That is a fact.

Ears are not enough

The authors don’t sugarcoat it. Human ears alone are not reliable against synthetic voices, and teaching people to listen for artifacts doesn’t help when the artifacts are gone, which is basically happening right now. Their best listeners, the ones who caught the old fakes almost every time, lost nearly all their edge on ElevenLabs.

Their answer is layers. Verification procedures. Provenance, like C2PA. Watermarks built into the voice when it’s generated, so a machine can check it, and nobody has to rely on their ears. And detection that finds the fake part of a recording, not just a yes or no for the whole file.

On a phone call, the receiver will most likely have only the sound, without the original file or its metadata. That makes a watermark carried inside the audio especially useful.

What we do about it

This is why we built DeepMark. We put an ID inside the voice when it’s generated, and we read it back second by second, in real time, even on a phone call.

So if that swapped sentence came from a marked voice, we don’t guess. We show exactly which seconds are AI.

A mark doesn’t care how real a voice sounds. The better voices get, the more that matters.

It only works if voices are marked when they are made. That’s why we work with voice companies to mark what they generate, and why we help both agents and humans authenticate.

If your business makes voices or answers calls, let’s talk.