Seeing is no longer believing

Why seeing used to work as proof

For most of history, a face, a voice or a photograph was expensive to fake. Forging one took skill, time and access. So when you saw someone on camera or heard your brother on the phone, the cost of faking it made the evidence fairly good.

That is why the habit runs so deep. Researchers who tested people on deepfake videos in 2021 described a seeing-is-believing heuristic: people leaned towards treating videos as authentic, and overestimated their own ability to catch fakes. The habit was not foolish. The cost it relied on has fallen.

What a face, voice or video shows now

What the research found

In a 2022 study in PNAS, Sophie Nightingale and Hany Farid asked people to tell AI-generated faces from real photographs. Average accuracy was 48.2%, around the level of guessing. With training and feedback it rose only to 59%. In a third experiment, the AI faces were rated 7.7% more trustworthy than the real ones.

Voices are no safer. In a 2023 study of 529 listeners in English and Mandarin, people spotted AI-generated speech 73% of the time, and hearing examples first helped only slightly. For video, a 2021 experiment found people could not reliably detect deepfakes, and neither warnings nor cash rewards improved them.

These tests used tools that are now years old. The authors of the voice study expected detection to get harder as the technology improves.

How it is already used on real people

In 2024 the engineering firm Arup confirmed that an employee in Hong Kong paid about $25 million after a video call with people he believed were the company's finance chief and colleagues. All of them were deepfakes. In Kozhikode in 2023, a retired man sent ₹40,000 after a WhatsApp video call with what looked and sounded like a former colleague. He found out only when he rang the colleague's saved number.

The FBI has warned that criminals use short clips of a loved one's voice to fake an emergency, and generate video for real-time calls with supposed executives or police.

What replaces 'I saw it'

We do not think the answer is to get better at spotting fakes. The studies above suggest that is a race people lose. The answer is to stop asking media to prove itself, and ask for proof that does not depend on how real something looks.

• Provenance: where a file came from. An open standard called Content Credentials, from the C2PA coalition whose members include Adobe, Google and Microsoft, lets anyone see a file's origin and edit history where it has been attached. When it is missing, that tells you nothing either way.
• A channel you control: hang up and call back on a number you already have. The FBI also suggests a secret word agreed with family.
• Verification by someone other than the sender: a check that was done once, by a party that looked, and that you can confirm yourself instead of judging by eye.

None of these ask you to be a better judge of faces. That is the point. See also why a video call is not verification.

Questions