If you have a mental checklist for spotting AI images, here is the uncomfortable news: it has a shelf life, and it is shorter than you think. Count the fingers. Read the text on the sign. Check the reflections. Each of these was, for a while, excellent advice. Each of them has been quietly retired by the next model release. The checklist isn't wrong so much as it is perishable, and nobody puts an expiry date on the carton.
This is the thesis behind everything we do here, so it's worth saying plainly. The tells are a moving target. Not because the people teaching them are careless, but because the thing they describe keeps changing underneath the description.
Watch how a tell dies. It always goes the same way.
A new image model ships. People notice it can't do hands: six fingers, fused knuckles, a thumb where no thumb should be. "Count the fingers" becomes the universal advice, repeated in every explainer and group chat. It works, for about two years.
Then the labs read the same complaints you do. Hands were a known failure, so hands got fixed. FLUX.1, released in 2024, was singled out by reviewers for being unusually good at hands, and the rest of the field followed. A clean hand stopped meaning anything. The advice didn't get worse; the world moved past it.
Text inside images was the next reliable giveaway. For years, diffusion models turned any sign, label, or menu into a smear of dream-language: letters that looked like letters until you tried to read them. Then that broke too. OpenAI shipped ChatGPT Images 2.0 on April 21, 2026, its next-generation image model with markedly better text rendering, and competing labs have been closing the same gap. OpenAI hasn't said exactly what changed under the hood, but the direction of travel is clear: in-image text, once a structural weakness, is now something the leading systems mostly get right.
So that's two flagship tells, patched in roughly the order people learned them. Reflections, skin texture, background clutter: these are on the same conveyor belt, at different stages. The pattern is the point: any artifact specific enough to teach is specific enough to fix.
The reason your checklist can't keep up is simple arithmetic. The models ship faster than the folk wisdom about them. In a single twelve-month stretch across 2025 into 2026, the major image families pushed out version after version: Midjourney moved from V7 to a V8 line, OpenAI shipped ChatGPT Images 2.0, and the rest of the field raced to keep pace. Every one of those releases is, among other things, a list of last year's tells, marked "addressed."
A checklist is a snapshot. The models are a video. You cannot win a moving fight with a still photo.
Here is the part people resist, so let's put a number on it. When you ask ordinary, motivated people to sort real images from AI ones, they land barely above a coin flip.
In one large experiment (roughly 287,000 image judgments from more than 12,500 participants), overall accuracy came in at 62%. That sounds like a passing grade until you remember the floor is 50%. Other studies land lower, near 54%, and a broader read across dozens of studies finds human detection clustering around chance. People did best on human portraits, where we are exquisitely tuned to faces, and worst on landscapes, where there's no familiar geometry to catch a lie.
Sixty-two percent is not a skill but a slightly weighted coin. And it's the score of people who knew they were being tested and were looking for the trick. In the wild, scrolling, half-distracted, primed to believe whatever confirms the day's mood, the real number is worse.
So the honest position is that the tells are eroding, the humans are near chance, and confidence is the most dangerous thing you can bring to the screen.
If checklists rot, what's left? Not nothing. The durable skill was never a list of artifacts. It's a posture.
Slow down. Almost every bad call is a fast call. The image arrives, it confirms something, you pass it on. The single highest-yield habit is refusing to react at the speed the feed wants. A trained eye isn't one that knows more tricks; it's one that pauses before it's sure.
Check the boring parts. Not the dramatic center of the image: the edges, the captions, the source. Who posted this first? Does the account have a history or did it appear last Tuesday? Does the "photo" exist anywhere a real photo would have left a trail?
Ask for provenance, not vibes. The most reliable signal in 2026 is not in the pixels at all but in the receipt attached to them. Cameras from Leica, Canon, and others now sign images at capture with Content Credentials, and platforms like TikTok have started surfacing them. The rollout is uneven: Nikon added it via firmware in 2025 and had to suspend the whole program within days, after a researcher found a way to make the camera sign a forged image. And a missing credential proves nothing on its own. But verifiable origin beats a hunch every time. We dig into how that machinery works, and where it fails, in content provenance and C2PA.
Refuse to be sure for free. Certainty should cost you something: a check, a search, a second look. If you didn't pay for it, you don't actually have it.
This is also why a list of current tells, while genuinely useful, is a starting point and not a destination. We keep an honest, dated rundown of what still works on today's images, but read it knowing the ground is moving. The deeper habit of thinking clearly about AI media is what survives the next model.
There's no trick coming that ends this. The labs will keep patching the giveaways, and the giveaways will keep getting subtler, and the gap between what a model can fake and what a person can catch will keep narrowing. A checklist meets that future already obsolete. A reflex (slow, skeptical, provenance-first) meets it on even footing.
Reflexes aren't built by reading. They're built by reps. Which is the whole quiet argument for doing a little of this every day: five items, a few seconds of genuine doubt, a number at the end that tells you exactly how good your gut actually is. That's the puzzle. The tells will keep dying. The habit is the part that lasts.
What's the single most reliable way to tell if an image is AI-generated in 2026?
There isn't one, and anyone selling you a single trick is wrong. The most durable approach is provenance (where did this come from, who first posted it, does it carry Content Credentials), combined with the discipline to not declare it real or fake until you've actually checked. Visual tells can point you at an answer; they don't hand you one.
Do the old tells like counting fingers still work?
Mostly not. Hand rendering has improved a lot since the era of six-fingered messes, and text inside images, once a dead giveaway, got dramatically better with releases like ChatGPT Images 2.0 in April 2026. Old tells aren't useless, but they only catch lazy or older outputs. A clean image proves nothing.