Nine in ten people say they can't spot AI fakes. The tested number is worse.

The headline doing the rounds this week is a tidy one: nine out of ten people "can no longer distinguish real from AI-generated content." It comes from a new Malwarebytes report called Face Value, it is already mutating as it travels — I have seen "90% of people can't spot deepfakes" twice on my feeds this morning — and it invites my favorite question, the one that should be tattooed on every statistic before it is allowed outdoors: measured how? Because here is the thing about that number. Nobody in that survey was shown a deepfake and asked to spot it. They were asked how they feel about their chances. The 88 percent is a confession of doubt, not a test score.

Normally this is where I would deflate the scary number and we would all go home early. Not today. Today the twist runs the other direction: when a different study actually sat people down and tested whether they could tell real from fake — properly, with a balanced set and a scoring rule — the tested number came out worse than the worried one. The headline got the question wrong and the conclusion right. That is rare enough to be worth taking apart carefully.

What the survey actually asked

Start with what Face Value is, because it is a perfectly respectable piece of sentiment research being dressed up, by other people, as a capability test. Malwarebytes surveyed 1,500 adults across the US, UK, Austria, Germany, and Switzerland, weighted for balance. The findings, as stated in the report rather than the aggregators: 88 percent say it is becoming harder to tell what online content is genuinely human or real. 85 percent say it can be hard to tell scams apart from the real thing — up from 66 percent asking the same question last year. 84 percent say convincing video no longer feels like proof. Notice the verbs. Say. Feels. These are people reporting their own confidence, and their confidence is collapsing — 19 points in one year on the scam question, which is an enormous move for a tracked survey item and far outside any plausible margin of error on this sample.

Now the footnote-reading. Fifteen hundred respondents across five countries is roughly 300 per country, which puts the per-country error bars at five to six percentage points before you slice by age or anything else — so treat any country-level or demographic claim from this report as directional, not gospel. And the mutation matters: "it can be hard to tell scams apart from the real thing" (the actual item, 85 percent) is a much weaker claim than "can no longer distinguish real from AI" (the headline, nowhere in the questionnaire). A statistic that changes meaning between the methodology section and the headline is not lying, exactly. It is commuting without a ticket.

The study that actually ran the test

So what happens when you stop asking people how they feel and start checking? For that we have the Veriff Deepfakes Report, run with Kantar in February: 3,000 respondents across the US, UK, and Brazil, each shown 16 visuals — eight authentic, eight AI-generated or manipulated, a balanced design that deserves explicit credit because it is what separates a measurement from a vibe. Scoring ran from −1 (got everything wrong) through 0 (random guessing) to 1 (got everything right).

The American average was 0.07. Statistically, that is the sound of a coin landing. On the hardest video pair, only 30 percent answered correctly — meaningfully below chance, which means the fake was not merely passing as real, it was outperforming the real thing at seeming real. Meanwhile, about half of US respondents said they were confident they could reliably identify manipulated media. Half believe; 0.07 deliver. And the cues people say they rely on tell you why: unnatural skin texture (53 percent), odd facial movements (51 percent), "gut feeling" (36 percent). Every one of those is a description of last year's generators. Detection-by-vibes has a shelf life, and it expires with each model release.

The people who say they can no longer tell are telling the truth. It's the half who say they can that should worry you.

Two numbers, one direction

Put the two studies side by side and the composite picture is unusually coherent for this genre. One measures perceived ability and finds it collapsing: 66 to 85 in a year. The other measures actual ability and finds it was never there: 0.07. The perception is catching down to the reality, not the other way around. There is even a number for the gap between them — Veriff identifies a 'high-risk' 7 percent who combine poor detection, high confidence, and a habit of never verifying anything. That segment, not the worried 88, is where the losses concentrate, because worry at least makes people check.

And if you want the corroborating evidence that isn't a survey at all, it exists, with a dollar sign on it. According to FBI figures reported this week, Americans lost nearly $900 million to AI-assisted scams over the past year — outcome data, immune to question-wording. In the Malwarebytes sample, half of respondents say they encountered an AI-driven scam in the past year; 19 percent report some form of AI identity harm; 10 percent say someone generated sexually explicit imagery of them without consent. Self-reported, yes, with all the usual caveats about what people recognize and admit. But when the sentiment data, the tested data, and the loss data all point the same way, the burden of proof flips to whoever wants to argue it's fine.

The numbers worth keeping

Stripped of the headline inflation, here is the scorecard I would actually carry forward from this week:

• Tested deepfake detection, US average: 0.07 on a −1-to-1 scale — statistically indistinguishable from guessing (Veriff/Kantar, n=3,000, balanced 8/8 design)
• Hardest video pair: 30 percent correct — below chance; the fake beat the real at looking real
• Confidence-competence gap: roughly half believe they can reliably spot fakes; the scores say otherwise
• Self-reported difficulty telling scams from real: 85 percent, up from 66 percent in one year (Malwarebytes, n=1,500, five countries)
• Reported US losses to AI-assisted scams: nearly $900 million in a year, per FBI figures
• The 'high-risk' segment — bad at detection, sure of themselves, never verify: 7 percent

What would change my read? A tested study showing detection scores improving over time — they are not; the trend across the literature runs the other way as generators improve. Or evidence that the 19-point sentiment jump is an artifact of question wording or sample change — possible, the report doesn't publish enough methodology to rule it out, and I'd genuinely like to see the questionnaire. Malwarebytes, the inbox is open. Until then, the one-year collapse in confidence looks like signal, because it agrees with everything that isn't a survey.

The defenses that don't require eyesight

The practical conclusion is the one the deepfake-detection industry won't put on a slide: training your eyes is a losing strategy, because your eyes are competing with a model that retrains faster than you do. The defenses that hold up are procedural, and the survey respondents are quietly inventing them already — 13 percent of families in the Malwarebytes sample have set up a codeword to authenticate each other against voice-clone calls, and 19 percent have disabled their voicemail greeting so there is less clean audio to clone. I find that 13 percent weirdly moving. It is signal-detection theory arriving at the kitchen table: when the channel can be spoofed, you authenticate out-of-band. Banks call it two-factor. Grandmothers call it the codeword.

So, the verdict. The viral version of this story — "nine in ten can no longer tell real from fake" — fails the audit on wording: that is not what was asked, and a perception survey cannot support it. The defensible version is starker anyway: when ability was actually measured, it rounded to a coin flip, and the public's self-assessment has spent the past year converging on that fact from above. People are not becoming paranoid. They are becoming calibrated. The chart of public doubt only goes up, and my standing rule is that a chart that only goes up deserves a hard look. I looked. It is still going up, and this time the error bars agree with the alarm.