Start with the hands. Not because they are the only giveaway, but because they are the one place a fake so often falls apart while the rest of the picture holds together beautifully. Six fingers. A thumb bending the wrong way. Two hands that belong to different people. The face can be flawless and the hands still betray the whole thing.
That gap — gorgeous surface, broken logic — is the single most useful idea to carry into any image you are unsure about. Understand why it happens and you stop squinting at random pixels and start looking in the right places.
Why the fakes leave traces at all
Modern image generators are, at heart, extraordinarily sophisticated pattern reproducers. They learn statistical regularities from enormous collections of pictures and then produce new images that resemble what they have seen. They are very good at the overall look and texture of a scene. What they do not have is any model of the physical world. They do not know that a hand has five fingers because bones say so, or that a single lamp casts shadows in one direction. They know only what tends to appear near what.
So the tells cluster wherever reality imposes a consistent rule the model has to honour across the whole frame. Anatomy. Physics. Text. Continuity. Those are the pressure points. Everything below is really just a tour of them.
Look closely at the hard details
The most reliable places to inspect are the parts of an image that demand the model track many small, consistent elements at once.
- Hands, teeth, and ears. Hands are notorious — count the fingers, watch how they bend, look for fused or extra digits. Teeth can be too many, or eerily uniform. Ears often carry a muddled, invented internal structure that reads as almost-right until you actually study it.
- Text within the image. Signs, labels, book spines, logos. Generated text tends to dissolve into plausible-looking gibberish, or letters that change style halfway through a word. It is one of the fastest checks going.
- Symmetry and paired objects. Earrings, shoes, the arms of a pair of glasses, a repeating pattern on fabric — these frequently fail to match side to side. Backgrounds sometimes tile or repeat in ways nothing real would.
- Where objects meet. Hair against a shoulder. Glasses touching a face. Fingers gripping a cup. Boundaries are hard, and it shows: edges that melt together, or don’t quite connect.
Check the physics of light and space
Because the generator never built a real scene, its lighting can be internally incoherent. So interrogate the light.
Look at the shadows. Do they all fall in a direction consistent with the apparent light source? Are there shadows simply missing where an object should cast one? Then the reflections — mirrors, water, sunglasses, any shiny surface. A reflection that doesn’t correspond to the scene, or that isn’t there at all, is a strong signal. Scan the background too, for objects that bend, merge, or trail off into incoherence, and for a depth of field that blurs in ways no real lens produces. None of these is proof on its own. Several together start to mean something.
Notice the overall “feel”
There is also a quality experienced viewers learn to sense before they can name it. Skin that is unnaturally smooth or faintly waxy. A slight plastic sheen. A composition just a touch too perfect. Colours richer and more harmonious than an ordinary camera would deliver on an ordinary day. And the faces of people who do not exist can look oddly generic, as though averaged from a thousand others.
This is subjective. It is not evidence. But it earns its keep, because it is usually the thing that makes you stop and look harder — and the deliberate second look is where the real work happens.
The best checks aren’t about the image at all
Here is what most people skip. The strongest tests often have nothing to do with pixels. Where did this come from? Who published it first, and do they have a track record? Does any reputable outlet carry the same picture, or report the event it claims to show? A dramatic image living only on anonymous accounts, with no credible origin, deserves scepticism no matter how convincing it looks.
A reverse image search is one of the most useful tools you have. It can show that a picture is a doctored version of an older, real photograph, that it has surfaced before in a completely different context, or that fact-checkers have already taken it apart. Testing the surrounding claims against independent reporting — the discipline our technology desk applies as a matter of routine — often settles the question faster than any amount of pixel-level squinting.
Metadata and provenance: useful, not decisive
Digital images can carry embedded metadata describing how they were made, sometimes naming the software or the device. Worth knowing about. Also easy to strip and easy to fake, and many platforms wipe it automatically the moment you upload. Its absence proves nothing; its presence can be forged.
A newer approach tries to fix that weakness at the source. Provenance standards — the widely adopted C2PA framework, surfaced to the public as “Content Credentials” — attach cryptographically signed information about an image’s origin and edit history, and a growing number of cameras, editing tools and AI generators now support it. When that signed provenance is present and intact, it is far more trustworthy than ordinary metadata. But it is opt-in and still spreading, so its absence, again, tells you little. Treat every technical signal as a clue, never a verdict.
Reason about it, don’t just react
The most important habit is also the least technical. Fakes travel best when they confirm what someone already wants to believe, or when they land a hard emotional punch. So slow down precisely on the images engineered to make you furious, frightened, or triumphant — the exact moments your guard drops. Ask the plain questions. Does this actually make sense? Who benefits if I believe it? What would confirm or refute it? That pause catches more fakes than any single visual trick.
A moving target, and an honest limit
Be honest about where this is heading. Every tell described here is a snapshot of what today’s models get wrong, and the models keep getting better. The mangled hands are already rarer than they were a couple of years ago. Garbled text is being cleaned up. A checklist that works this year will quietly lose entries next year, which is exactly why the durable defences are the ones that don’t depend on the picture: provenance, sourcing, reverse search, and a sceptical pause before you share.
No single sign is proof, and no viewer catches everything. Treat these as a layered habit of doubt rather than a lie detector. The goal isn’t to become infallible. It’s to stop being the person who forwards the fake before anyone thought to ask where it came from.