Everybody thinks they can spot a fake. Scroll Hinge for ten minutes and you'll catch yourself doing it: "that's AI, that's AI, that's definitely AI." It feels obvious. It feels like a superpower.
It isn't. In a study published in the British Journal of Psychology, researchers at UNSW and ANU tested 125 people, including 36 "super-recognisers" with exceptional face-recognition ability. Average participants performed only slightly better than chance, and even the super-recognisers beat them by a slim margin. The part that should bother you: confidence stayed high regardless of accuracy. People were sure they'd nailed it while getting roughly half wrong.
So when a photo pings your radar as fake, you're often not detecting the AI. You're detecting something, and it's usually not what you think.
This matters in both directions. If you're swiping, you want to know what genuinely signals a fake profile versus what just signals a good camera. If you're using AI photos yourself, you want to know which of your images are quietly getting you left-swiped for reasons nobody will ever tell you.
The 60-Second Fake Check
- Look at the light on the face and the light in the background. Do they come from the same direction, at the same softness?
- Find the edges: hair against sky, shoulder against wall, glasses frame against skin. Look for a faint halo or a too-clean cut.
- Read any text in frame. Signs, labels, tattoos, book spines. AI still turns writing into confident gibberish.
- Check skin at 200%. Pores and stubble shadow, or a matte airbrushed surface with no texture anywhere?
- Look for asymmetry. Real faces are lopsided. Two eyebrows at identical heights is a warning, not a compliment.
- Scan the whole profile as a set. Same light quality, same depth of field, same distance from camera in every shot?
- Count the mundane. Is there a single boring photo, bad angle, or unflattering moment anywhere in the lineup?
That last one catches more fakes than anything on the anatomy checklist. Hold that thought.
The Tell Moved, and Most Advice Didn't
Every listicle still tells you to count fingers and check the ears. That advice was correct in 2023. It's mostly dead now.
There's real research behind why people believed it. A 2025 study of 664 participants in Royal Society Open Science found that a five-minute training session on specific rendering errors lifted accuracy from 31% to 51%, and from 41% to 64% among super-recognisers. What did the training teach? Misaligned teeth, odd hairlines, and misshapen ears.
Those were the errors of the models people were tested on. Current image models have largely closed the hand and ear gap. If your entire detection method is finger-counting, you're screening for a generation of tools nobody uses anymore, and you'll wave through fakes while flagging real photos of people with slightly weird ears.
The tells that survived are not anatomical. They're about physics and context.
The Real Reason AI Photos Look Fake
Here's the mechanism, and understanding it is worth more than any checklist.
AI faces don't fail by being ugly or broken. They fail by being too normal. Dr Amy Dawel, who co-led the ANU work, put it plainly: rather than obvious glitches, AI faces "tend to be unusually average, highly symmetrical, well-proportioned and statistically typical."
An image model learns a face from an enormous pile of faces. What comes out leans toward the statistical middle. Symmetry gets cleaned up. The crooked incisor smooths over. One nostril stops being bigger than the other. Skin loses the small blotchiness that every actual human has.
Individually, none of that reads as an error. Collectively, it reads as a face with nothing to explain. Your brain has spent your entire life learning real faces, and real faces are lopsided, blotchy, and specific. A face with no specifics doesn't trip the "fake" alarm so much as the "boring" one.
There's an ugly twist here. A Psychological Science study found that White AI-generated faces were judged human more often than actual human faces, 69.5% versus 52.2%. They called it AI hyperrealism. Synthetic faces can out-human humans, because averageness reads as familiar and familiar reads as trustworthy.
So the practical version, for your profile: the problem with AI photos usually isn't that she thinks "that's fake." It's that she feels nothing and swipes. Generic doesn't get reported. It gets ignored, which is worse, because you never find out.
How to Spot AI Generated Photos on Dating Apps
Forget the fingers. This is the list that still works in 2026.
What Actually Gives AI Photos Away Now
| Signal | What You're Looking For | How Reliable |
|---|---|---|
| Light mismatch | Face lit softly from the left, background shadows falling right | High. Physics is hard to fake and easy to check |
| Text in frame | Signs, menus, logos, tattoos rendered as plausible nonsense | High, when text is present at all |
| Edge halos | Faint outline where hair meets background, or a suspiciously clean cut | Medium-high. Zoom in on the hair |
| Skin with no texture | Matte, poreless, uniform tone across forehead and cheeks | Medium. Beauty filters produce the same thing |
| Over-symmetry | Matched eyebrow heights, perfectly even smile, flawless dentition | Medium. Some people really are symmetrical |
| Set-level sameness | Every photo shares one look: same light, same lens, same crop distance | Very high. This is the strongest one |
| No boring photos | Six images, zero of them mediocre, zero candid, zero awkward | Very high, and almost nobody checks it |
| Hands and ears | Extra fingers, melted earlobes, asymmetric ears | Low now. Fixed in current models |
Notice the two "very high" rows are both about the profile, not the photo. That's not an accident.
The Set Is the Tell
A single AI image can be flawless. Six of them from the same tool on the same afternoon cannot hide what they are.
Real camera rolls are a mess. Your photos come from four years, nine locations, three phones, and a friend who doesn't know how to hold a camera. The light is wrong in some. Your eyes are half-closed in one. You look tired in another because you were.
An AI batch shares a fingerprint: consistent depth of field, consistent skin rendering, consistent flattering angle, consistent competence. Every shot is the best shot. Nobody's life produces that.
This is also the fastest way to audit your own profile. Line up your six photos and ask whether they look like a camera roll or a catalogue. If it's a catalogue, you have a problem, regardless of how good each individual image is. Mixing AI shots with genuinely real ones is not a moral compromise, it's the thing that makes the lineup read as human. The dating profile photo order guide covers which slots take which type, and the dating app photo mistakes breakdown covers the ones that predate AI entirely and still sink more profiles than anything on this page.
When You're Screening Someone Else
If you're on the receiving end, the stakes are different. You're not judging aesthetics, you're checking whether a person exists.
The money side of this is not small. The FTC reported $298 million lost to romance scams that started on social media in 2025, with nearly 60% of romance scam victims saying it began on a social platform. AI faces made the cheapest part of that operation, the photos, effectively free.
Screens That Actually Work
- Ask for a video call early. Not as an accusation, just as a normal next step
- Ask for a specific photo: 'send me a pic of your coffee right now.' Real people can. Scripted profiles stall
- Reverse image search the photos. Catfish accounts often reuse stolen or previously posted images
- Notice whether the conversation ever gets specific about their actual week
Screens That Waste Your Time
- Counting fingers. Current models get hands right
- Assuming a professional-looking photo means fake. Some people just hired a photographer
- Treating smooth skin as proof. Every phone ships with beauty processing turned on
- Running one AI detector and trusting the verdict
That last point deserves numbers, because people lean on these tools far too hard.
Don't Trust the Detectors
NewsGuard tested five leading AI image detection tools in May 2026 against 45 images. The tools declared authentic images to be AI-generated 13.33% of the time collectively. One tool, ScamAI, flagged 6 of 15 genuine press photos as fake. ZeroGPT flagged 3 of 15. Two others got all 15 right.
Same image, three detectors, three answers. That's the state of it.
There is a more reliable layer, and it's provenance rather than detection. Since May 2026, OpenAI attaches C2PA Content Credentials and Google's SynthID watermark to images from its models. The metadata can be stripped from a file. SynthID lives in the pixels and survives screenshots and compression. So a positive signal means something. The absence of one means nothing at all, since plenty of generators mark nothing.
The apps are solving this from a different angle. Tinder's Face Check asks for a live video selfie, confirms a real person is physically present, and compares facial geometry to your profile photos. That's an identity check, not an AI detector, and Match Group keeps extending the pattern. It doesn't care whether a tool touched your image. It cares whether the face is yours. Policy specifics per app are in does Tinder ban AI photos and does Hinge allow AI photos.
How to Not Get Clocked
Everything above inverts into a fix list.
Making Your Photos Read as Real
- Upload unfiltered inputs. If you feed the model smoothed skin, it will hand back poreless skin and you'll blame the tool.
- Vary your input lighting and angles hard. Same-bathroom selfies produce one flat rendering of your head in every output.
- Reject any output where the light on you disagrees with the light in the room.
- Kill photos with text in frame unless the text is legible and correct.
- Keep at least half your lineup as genuine camera roll photos, including one that isn't flattering.
- Pick scenes your actual life can back up. A logistics guy with six yacht photos fails on story, not pixels.
- Generate a lot, keep three or four. Reject on any single hard tell, however good the pose is.
- Show the set to a friend and watch their face for the first half second. That reaction is your real detector.
Point five is the one guys resist and it's the highest leverage on the list. A slightly worse photo that's obviously real does more for the credibility of the whole set than another perfect one. Perfection is the signal now.
The deeper mechanics of getting good output, what to upload, how to curate, are covered in the full guide on how to make AI dating photos look real. If your raw material is the problem rather than the tool, start with how to take good dating photos instead, because no generator fixes a camera roll of twelve identical bathroom mirror shots. And if you're still choosing a tool, the best AI dating photo generators comparison sorts the dating-first ones from the LinkedIn headshot tools in costume, which matters, because studio-grey corporate output reads as its own species of fake on Hinge.
The Line That Matters
There's a difference between better lighting and a different face. Reshaping your jaw, slimming your body, or adding hair you don't have isn't optimization, it's a loan against your first date at a terrible interest rate. She'll see you in a bar on Thursday. That's the only detector with real consequences, and it has never had a false positive.
Whether any of this is fair game at all is a reasonable thing to chew on, and worth reading are AI dating photos ethical before you decide where your own line sits. The short version of my position: changing the room is fine, changing the person is not.
The Short Version
Key Takeaways
- You are worse at spotting AI faces than you think, and confidence doesn't track accuracy.
- Hands and ears are dead tells. Current models get them right.
- The live tells are lighting mismatch, garbled text, edge halos, and texture-free skin.
- AI faces fail by being too average, not by being broken. Symmetry and flawlessness are the giveaway.
- The strongest signal is set-level: six photos sharing one look, with nothing mediocre anywhere.
- AI detectors are unreliable in both directions. NewsGuard measured a 13.33% false positive rate across five tools.
- To avoid getting clocked: unfiltered inputs, varied lighting, real photos mixed in, and at least one that isn't flattering.
The whole thing collapses into one question, and it's not "can they tell?" It's "does this look like me?" Get that right and detection stops being your problem, because there's nothing to detect. Get it wrong and you'll pass every algorithm and fail the only test that counts. If you want the full lineup strategy from scratch, start with the best dating profile photos for men.