Someone sends you a photo and asks: is this real? You squint at the hands. You count the fingers. You zoom into the background text to see if it's gibberish. Two years ago that worked. It mostly doesn't anymore.
Here's the thing almost nobody tells you: the most reliable way to check an image isn't to look at it at all. It's to read the file.
The one-line version
Ask for the original file, not a screenshot. Original files often carry metadata that says outright what made them. Screenshots and social downloads carry nothing, because every platform strips it.
Why looking at the picture stopped working
The visual tells everyone learned in 2023 — six fingers, melted ears, garbled signage, plastic skin — were artifacts of a specific generation of models. They got fixed. Current image models render hands correctly, spell words correctly, and produce skin texture that survives a zoom.
What's left is a vibe: an image that's a bit too well-lit, a bit too symmetrical, faces that are slightly too even. That instinct is worth something, but it isn't evidence, and it fails badly on professional photography, which is also well-lit and symmetrical. Plenty of real studio portraits get called AI. Plenty of generated images sail past.
What the file knows
Image files carry more than pixels. Depending on what made them, they can also carry a record of how they were made. There are three kinds worth knowing about.
1. C2PA, or Content Credentials
This is the serious one. C2PA is a cryptographically signed provenance standard backed by Adobe, OpenAI, Microsoft, Leica and Sony. When a file carries it, the manifest records the chain: what device or model created the image, and what was done to it afterwards. Because it's signed, it's tamper-evident — you can't quietly edit it the way you can edit EXIF.
It's still far from universal. But when it's there, it's the strongest answer you'll get from a file.
2. The IPTC digital source type
A dull-sounding metadata field that does something very specific: it declares how an image originated. The value trainedAlgorithmicMedia means, in plain terms, “a generative model made this.” News agencies and stock libraries use it to label AI imagery. If you find it, you're done — the file is telling you directly.
3. Generator fingerprints
This is my favourite, because it's almost comically direct. Stable Diffusion writes the entire prompt into the PNG. Not a hint, not a flag — the actual text someone typed, along with the sampler and step count. Midjourney, DALL路E, Firefly and NovelAI tend to leave their names in the Software or CreatorTool field.
If you've never looked, it's worth doing once just to see it. Take a PNG straight out of an image generator, open it in any metadata viewer, and read the prompt back.
The catch, and it's a big one
All of this only survives in the original file. Upload an image to Instagram, X, WhatsApp, Discord or Slack and the metadata is stripped on the way in. Take a screenshot and you've created a brand-new file with none of the history. Re-save it in an editor and most of it is gone too.
So in practice: if a friend forwards you a screenshot, metadata will tell you nothing. If you can get the file as it came out of the camera or the generator, it will often tell you everything. That's the whole game — ask for the original.
Reading it without installing anything
You can paste a direct image link into our AI image checker and it will read the C2PA manifest, the IPTC source type, generator fingerprints and camera EXIF, then show you exactly which of those it found. If a generation prompt is embedded, it shows you the prompt.
It's deliberately built to give one of four answers, and one of them is “I don't know”. If a file has no metadata, it says so plainly rather than converting silence into a confident percentage. I'd rather hand you nothing than hand you a number I can't defend.
What about the tools that score the pixels?
They exist, and some are genuinely decent. But they share a failure mode that matters more than their headline accuracy: when they're wrong, they're wrong confidently, and there's no way to audit the reasoning. You get 87% and a shrug.
Metadata is different in kind. It's not a probability — it's a claim, attached to the file, that you can inspect. That's why we built the metadata half first. Not because pixel detection is worthless, but because a verifiable answer and a plausible guess are different products, and mixing them makes both less trustworthy.
A practical order of operations
- Ask for the original file. Not a screenshot, not a forward. This single step decides whether any of the rest is possible.
- Read the metadata. C2PA, IPTC source type, generator name, embedded prompt. A hit here is close to conclusive.
- Check the source, not the file. Where was it first posted? Does the account have history? Reverse image search it. Provenance in the ordinary sense usually beats forensics.
- Then, and only then, look at the picture. Use your instinct as a tiebreaker, not as the verdict.
The honest ending
There is no tool right now — ours included — that can look at an arbitrary screenshot and reliably tell you whether a machine made it. Anyone selling that certainty is overselling. What you can do is check whether the file admits it, and get much better at asking for files that still can.
That's a less satisfying answer than a big red AI DETECTED banner. It also happens to be true, and on this particular question the difference matters.
