Published on 2026-07-25
Is this profile picture too perfect to be a real person?
A 2022 study found people can't tell AI faces from real ones better than a coin flip, and rate the fake ones as more trustworthy. Origin is the tiebreaker.
The face in the profile picture is lovely. Warm smile, good lighting, nothing obviously wrong. Which is exactly the problem, because 'nothing obviously wrong' turns out to be a terrible way of deciding whether a face is real. Researchers Sophie Nightingale and Hany Farid tested this directly: in a study published in the Proceedings of the National Academy of Sciences in 2022, participants tried to sort real photographed faces from ones generated by AI. Their average accuracy came out at 48.2 percent, on a task where random guessing gets you 50.
Worse than a guess, in the direction that matters
It gets stranger. The same research found people didn't just fail to spot the fake faces, they actively preferred them. Participants rated the AI-generated faces as more trustworthy than the real ones, 4.82 against 4.48 on their scale, a gap the researchers linked to a simple quirk: synthetic faces tend to look like an average of many faces, and average faces read as more trustworthy to us regardless of whether a real person sits behind them. A fake profile picture doesn't just slip past your instincts. It can actively charm them.
A separate 2025 study in Frontiers in Artificial Intelligence tried to see whether training helps. Researchers taught participants to hunt for specific problem areas, ears, eyes, skin texture, background, then tested them again. Accuracy crept up, from 67.4 to 71.1 percent, a real gain but a modest one. And it came at a cost: the same trained group got noticeably worse at correctly recognising genuine photos as genuine, mistaking more real people for fakes than they had before the training. Sharper eyes for glitches made people more suspicious across the board, real faces included.
The same study found something else worth knowing: not all fake faces are equally catchable. Faces from one older family of generator were spotted correctly only 25.3 percent of the time, worse than random chance, because their flaws had become too subtle for an untrained eye. Faces from newer diffusion-based tools were caught far more often, at 88.7 percent, at least in this particular test set. That gap should make you suspicious of any confident rule of thumb: 'AI faces always look a certain way' was only ever true of the generator that happened to be popular when you learned the rule.
What to actually look at
None of this means the visual clues are worthless, only that they're not decisive on their own. Ears are a genuinely useful place to look: real faces are slightly asymmetrical, and generators sometimes produce two ears, or two earrings, that don't quite match each other in shape or position. Backgrounds behind a profile photo can blur or melt in ways a real camera lens doesn't produce. Skin can look uniformly smooth in a way no real complexion manages, and hair at the edges sometimes merges strangely into whatever's behind it.
- Ears and earrings that don't match between the two sides
- A background that blurs, warps or repeats oddly close to the face
- Skin with no texture at all, an evenness real skin doesn't have
- Eyes with a slightly glassy or mismatched look between the two
- Hair strands that fade into the background rather than ending cleanly
A face can pass every glance you give it and still belong to nobody at all.
The test that actually settles it
Here's what visual inspection can't tell you: whether this person exists anywhere else. That's what a reverse image search is for, and it takes under a minute, according to media literacy specialists cited by AARP. Drop the photo into Google Images or a similar tool and see what comes back. A real person, especially one active enough online to be messaging you, tends to leave a trail: the same photo on another platform, tagged with their actual name, appearing in an old post, a graduation photo, a tagged group shot. A generated face has no trail to find, because it was never anywhere before someone made it for this one profile.
No results doesn't automatically mean fake, plenty of real people keep a small footprint. But paired with a profile that appeared recently, has almost no history, and is already asking you personal questions or moving a relationship along fast, an image with no past anywhere else stops being a curiosity and starts being the whole answer.
Sources
- AI-synthesized faces are indistinguishable from real faces and more trustworthy (Nightingale and Farid, PNAS 2022) (hfarid.org)
- Training humans for synthetic face image detection (Frontiers in Artificial Intelligence, 2025) (frontiersin.org)
- How to Use Reverse Image Search to Spot Fake Photos (AARP) (aarp.org)