How can you tell if an image is AI-generated?

Usually not by looking at it. The checks that work are in the file rather than in the picture — content credentials, metadata, and a watermark the generator left behind — and each answers a narrower question than "is this real".

Looking at it barely works any more

Six fingers, melted text, plastic skin: those were the tells of 2023, and current models get them right often enough that their absence means nothing. A visible mistake can condemn an image. A clean one clears nothing.

The detectors that score an image are classifiers. They return a probability from a model trained on the generators it has seen, and they are wrong in both directions — confidently on generators they haven't seen, and on ordinary photos that have been cropped, filtered or re-saved.

What to read in the file

• Content credentials (C2PA). A signed manifest recording which tool made the image and what was done to it since. OpenAI's and Adobe's image tools write them, and a few camera models do. A valid signature is the strongest evidence available here — but it records what the signer says, not what happened in front of the lens.
• SynthID. Google's invisible watermark, in images made or edited with Google AI. You can finally check it yourself: upload the image in the Gemini app and ask whether it was made with Google AI. It only knows the tools that embed it, so it has nothing to say about an image from ChatGPT or Midjourney.
• Ordinary metadata. Software fields, timestamps, the absence of any camera data. Worth reading and easy to edit, so treat it as a lead rather than an answer.

You can read the credentials and metadata of a file in your browser with the audit tool — nothing is uploaded.

A blank result is not a verdict

Most images carry none of this, including photographs. Platforms strip metadata when you upload, a screenshot starts a fresh file with none of the original's history, and re-encoding or editing drops credentials unless the editor knows to keep them. That is why credentials go missing far more often than they are faked, and why no mark found means no information — not no AI.