How to tell whether something was made by AI
The advice you have been given is going out of date faster than anyone admits. Here is what still works, what stopped working, and the only thing that is likely to work in a year.
Checking the picture is a losing game, because every visual tell gets fixed. Checking the source is the durable answer - where did this come from, and who says so.
Everyone has been handed the same checklist: count the fingers, read the text in the background, look at the eyes. It is worth knowing what is actually true about that advice, because most of it is a description of 2023, and it is quietly expiring.
The tells you were taught
They were real, and some still help:
- Hands. Count fingers, look for fused or bent joints.
- Text inside the image. Generators have been notoriously bad at lettering, producing dreamlike, almost-words.
- Eyes and blinking in video. Real people blink every few seconds; some generated faces stare, or blink in a way that misses the small muscle movements around the eye.
- Physical nonsense. Jewellery that merges into skin, glasses arms that do not reach an ear, a shadow pointing the wrong way.
A useful roundup of the current cues is here.
Why I am not going to tell you that is enough
Here is the honest part, and it is the reason this piece exists.
Every one of those tells is a bug, and bugs get fixed. Hands got better. Text got better. Each generation of model closes the gaps the last one was caught on, so a checklist of visual tells has a shelf life measured in months. People who learned the 2023 list are now confidently wrong in two directions: they clear fakes that no longer have the flaws, and they accuse real photographs of being fake because a hand looks odd.
The second error is the one nobody warns you about. A checklist makes you confident, and confidence is exactly what you should not have here.
There is also serious work on signals a generator does not simulate - subtle colour changes in skin as blood moves through it, for instance - but that is detection software, not something you can do by eye.
What actually holds up
Move the question. Instead of *is this image fake*, ask where did this come from.
- Who posted it, and do they have anything to lose if it is false? A wire agency and a two-week-old account are different kinds of evidence.
- Does it exist anywhere else? A real event photographed by a real person usually has more than one frame, from more than one angle, from more than one account. A generated image often has exactly one, and no history before it appeared.
- Reverse image search it. Takes ten seconds and settles a surprising share of cases.
- Does anything else confirm it happened? Not "does it look real" but "did this event occur".
None of that requires you to out-stare a model, and none of it expires when the next one ships.
The thing that is likely to settle it
The durable fix is not detection, it is provenance: signing at the moment of capture, so a photograph carries a verifiable record of the camera that took it and what was done to it afterwards. That is what the C2PA standard is for, and camera makers and platforms have been adopting it.
That is worth being optimistic about, in the way this site tries to be optimistic - not because it is finished, but because it is the right shape of answer. It is much easier to prove where something came from than to prove where it did not. Its weakness is equally plain: it only helps where it has been adopted, and a signature stripped off is not a fake, it is just silence.
The one sentence to take away
Checking the picture is a game you will lose a bit more every year. Checking the source is not.
Found something wrong here? That is worth more to me than a compliment. Tell me and it gets corrected on the page, with the date.