How to Spot AI Images: The Tells That Still Work (and the Ones That Don’t)
Spotdle·Sep 26, 2026·6 min read
The finger-counting era is over. Here’s what replaced it.
ImagesIM
If you learned to spot AI images by counting fingers, your method has expired. It was good advice for about two years. It is now mostly wrong.
The models read the same internet you do, including every article complaining about mangled hands. Hand rendering has improved enough, on ordinary poses, that a clean hand tells you almost nothing anymore, though complex grips and unusual angles can still trip a model up. The same fate has caught most of the tells people still repeat. So let's be honest about where we actually stand, then look at what holds up.
Humans are not good at this. In a 2025 experiment spanning roughly 287,000 image judgments from more than 12,500 people, the overall success rate was 62%: better than a coin flip, but not by much. People did best on human faces and worst on landscapes, the natural and urban scenes where there's no familiar anatomy to anchor your suspicion.
The software isn't a magic shortcut either. A 2026 zero-shot benchmark tested 23 detector variants across 2.6 million images spanning 291 different generators and found no universal winner: the best detector averaged about 75% accuracy, the worst 37.5%, and against modern commercial generators like Flux and Midjourney v7, even the best tools dropped to 18–30%. Compression, noise, and unusual lighting throw them off badly. A detector score is a hint, not a ruling. We get into why these tools keep losing ground in why the tells are getting harder.
So no single look, and no single tool, settles it. What follows narrows the field. It does not prove a thing on its own.
The generators have gotten very good at the subject in the center of the frame. They are still worse at everything that subject depends on: physics, repetition, and small print. That's where to look.
Diffusion models compose a plausible-looking scene; they do not run a lighting simulation. So check whether the light agrees with itself.
Shadows that disagree. Two people lit from the same window, but their shadows fall in different directions, or one casts a hard shadow and the other none at all. Shadows are a frequent point of failure because the model is matching textures, not tracking a light source.
Reflections that lie. Mirrors, sunglasses, shop windows, still water. A reflection that doesn't match what's in front of it, or skips an object entirely, is a strong tell. Reflections require the model to render the same scene twice and keep it consistent, which it often can't.
Melted joins. Look where two things meet: where a hand grips a railing, where glasses sit on a nose, where a strap crosses a shoulder, where teeth meet lips. Edges that fuse, blur, or quietly merge into each other are still common.
Surfaces that are too clean, and color that bands#
Real photographs carry sensor noise, dust, micro-scratches, skin texture, uneven wear. AI output often looks airbrushed at the pixel level: pores smoothed away, fabric without lint, metal without fingerprints. On its own "looks too polished" is a weak signal now, because plenty of real commercial photography is heavily retouched too. But pair it with color banding (smooth gradients in skies or backgrounds that step in visible bands instead of blending) and you have something more useful.
This is the most durable tell in 2026. Models spend their effort on the focal point and improvise the rest.
Garbled distant text. Street signs, book spines, posters, license plates that dissolve into letter-shaped nonsense the moment they're out of focus.
Smeared crowds. Faces in the background that turn to putty, bodies that share a leg, people who blur into the architecture.
Broken repeating patterns. Brick courses that drift, railings with uneven gaps, tiled floors that bend, window grids that don't line up. Repetition is hard for a model that's painting plausibility rather than measuring a wall.
Despite real progress, embedded text is still where many models stumble: extra strokes on a letter, a word that's almost right, signage that reads as gibberish at a glance. If there's writing in the frame, read it slowly. It often won't survive the read.
Be ruthless about retiring dead tells. Clinging to them is how you get fooled with confidence.
Counting fingers. Covered above. The obvious hand failure is now the exception.
"It just looks too good." Modern output is supposed to look good. Polish is the default, not a flaw.
Hunting for a watermark or logo. Visible watermarks are trivially cropped or never applied, and absence proves nothing. The real provenance signals are now invisible. More on that below.
Any single tell, used alone. One clean hand, one sharp background, one readable sign. None of it clears an image. The skeptics who get caught are usually the ones who found one "proof" and stopped looking.
When it matters, stop squinting and check the source.
Reverse-image search first. Drop the image into a reverse-image tool and look for an earlier, higher-resolution, credited version from a real outlet. Finding the photo on a news wire from last year is worth more than any pixel-level hunch. Finding nothing (an "image" with no history) is itself informative.
Then check for content credentials. A real provenance standard now exists and is finally shipping. On May 19, 2026, Google expanded verification for both C2PA "Content Credentials" (cryptographic metadata describing how an image was made) and its own SynthID pixel watermark into Search, through Lens, AI Mode, and Circle to Search, with Chrome verification coming in the following months. OpenAI is among the companies now bringing SynthID to more of its own generated content. Pixel 10 was the first phone to embed Content Credentials in every photo at capture, and Meta will start labeling Pixel-captured photos and videos with Content Credentials when they're shared to Instagram.
The catch, and it's a big one: credentials are fragile. A screenshot or a re-upload strips the C2PA metadata, so most images you'll actually encounter still won't carry it. The pixel-level SynthID watermark is meant to survive when the metadata doesn't, which is exactly why both layers exist. So a credential that says "AI-generated" is reliable; the absence of one tells you almost nothing yet. We walk through reading and trusting these signals in content provenance and C2PA.
Knowing how these images are assembled is half the battle (see how AI images are made), and the same physics-and-edges instinct carries over to spotting AI video, where motion adds new ways to slip.
The honest summary: there is no single tell left, and there may never be one again. What still works is a habit: check the physics, read the background, distrust your first impression, and verify the source before you share. Your best detector is your own willingness to slow down.
That's the whole game in today's puzzle, too: five images, a little doubt, and the patience to look twice.
Does counting fingers still catch AI images?
Rarely. Hand rendering has improved a lot since the days of extra fingers and fused knuckles, and a normal-looking hand no longer tells you anything on its own, though complex grips and unusual angles can still trip a model up. Treat it as a weak signal at best.
Can an AI-image detector just tell me the answer?
Not reliably. A 2026 zero-shot benchmark of 23 detector models found the best one averaging about 75% accuracy, with no universal winner, and against the newest commercial generators, even the best tools dropped to 18–30%. They are a hint, not a verdict. Provenance and reverse-image search beat any single detector score.