The advice you’ve read about spotting AI photos is mostly out of date. Blinking, weird teeth, six fingers, EXIF data: generators fixed the obvious stuff years ago, and every major platform strips EXIF on upload anyway. If you’re still checking for those, you’re checking for nothing.
Here’s what still works, what doesn’t, and why a photo check was never supposed to be your only check.
The old tells are dead
A few years ago, “does the person blink” and “count the teeth” were real signals. They aren’t anymore. Generators closed the blinking gap, fixed gross facial asymmetry, and cleaned up teeth. Metadata is a dead end too: every major platform strips EXIF the moment a photo is uploaded, so “check the metadata” is advice for a version of the internet that no longer exists.
If a photo passes those old checks, that tells you nothing either way.
What still works in 2026
The tells that survive are smaller and require a closer look:
- Hands and fingers. Proportion problems, thumbs at angles a joint can’t make.
- Text in the scene. Signage or labels that look phonetically plausible but aren’t real words, or letters that mirror or warp.
- Eyes. Iris texture that doesn’t quite match between the two eyes, and catchlights (the little reflections) that are inconsistent with each other.
- Hair edges. Strands that blend into the background instead of separating cleanly from it.
- Shadows. Multiple shadows falling in different directions, as if lit from two suns.
- Backgrounds. Repeated patterns, a “ring wrapping oddly” around a finger, earrings that don’t match each other.
For dating profiles specifically, look at the set as a whole, not just one photo. A single studio-quality lighting setup across every photo, no candid or group shots, nobody else in frame with a sharp face, no identifiable public place, jewelry or tattoos that change between photos, and a face that looks “too clean” even at social-media resolution: any one of these is a caution, and two or three together are a stop.
Why your gut is a coin flip
Here’s the uncomfortable part. In a controlled study, ordinary people correctly spotted fake AI faces only about 30% of the time, and misjudged real faces as fake 46% of the time. Trained “super-recognizers,” people specifically skilled at facial recognition, only managed 41%. A short, five-minute training on rendering errors raised ordinary people to 51% and super-recognizers to 64%, which is better, but still nowhere near reliable.
There’s a second effect working against you: research on “AI hyperrealism” found people often rate AI-generated faces as more real-looking than actual photos of real people. Your instinct isn’t just uncertain, it’s biased toward trusting the fake.
Treat your gut reaction as a trigger for the checks below, never as the verdict itself.
Detectors aren’t the fix either
You might assume a paid detector tool solves this. It doesn’t, for one specific reason: the images you’re actually looking at.
In lab conditions, tools like Hive score 89 to 94% accuracy identifying images from generators like Midjourney, DALL-E, and SDXL, with roughly an 8% false-positive rate. Sightengine, AI or Not, and Illuminarty land somewhere in the 82 to 86% range in the lab. That sounds usable.
But a dating profile photo is never a lab image. It’s been resized, recompressed, and often screenshotted off an app, sometimes more than once. Independent testing found detection accuracy on recompressed, cropped, or social-media-processed images drops below 5%, while these same consumer detectors flag 5 to 15% of genuine photos as fake. That means a detector run on a real dating app screenshot is close to a coin flip too, in either direction. Treat any single detector score as weak evidence, not a finding.
Watermark checks (Google’s SynthID, Content Credentials or C2PA tags) have the opposite problem: they confirm an image came from a covered source, but their absence proves nothing. Midjourney, Stable Diffusion, Flux, and any locally run model carry no watermark, and a plain screenshot strips whatever manifest existed anyway. Treating “no watermark” as “not AI” is, by a wide margin, the most common mistake people make here.
What a photo check is actually for
None of this means skip the photo check. It means understand its job. A photo check is a trigger, not a verdict. Run the reverse image search, look for the tells, notice if the set feels like one lit session with no candids. If something’s off, slow down and run the other checks: name and city consistency, a conversation test, identity verification. A hit under a completely different name is an immediate stop. But a clean-looking photo doesn’t clear anyone. It just means you move to the next check instead of stopping here.
The test that actually catches a fake in real time
The one test that reliably outperforms all of the above is a live, unannounced video call, and specifically what you do during it. Real-time face-swap tools have gotten good at holding a straight-on, well-lit face still. They are much worse at everything else:
- Ask them to turn to a full profile view and hold it.
- Ask them to wave a hand slowly across their face, or hold fingers over their nose and mouth. Occlusion is the weakest point of live face swaps.
- Ask them to stand up, step back, or move the camera.
- Change the lighting mid-call: room light off and on, or turn toward a window.
- Hold an object with text on it up next to, then in front of, the face.
- Give them a random phrase to read aloud and watch the lips on open vowels.
- Ask them to put on or take off glasses, or push their hair back.
Do several of these, and don’t announce them in advance. A refusal to do any of them, for any reason, is the strongest signal you’ll get, stronger than anything a detector or your own eyes can tell you from a static photo.
The printable version of this, with the full photo and video-call checklist, is the free download on this page.
Sources: EyeSift AI image detection guide, 2026; LiveScience/Royal Society Open Science super-recognizer study, 2026; Hive, Sightengine, AI or Not, and Illuminarty pricing pages, 2026; independent AI-detector accuracy reviews (Imagera, AI Angst), 2026; Android Headlines on Google SynthID and C2PA, May 2026; Deep-Live-Cam project documentation; Metaphysic and DuckDuckGoose deepfake video-call research. General education, not legal advice.