How to tell if a photo is AI generated

What to look for inside the file

The checker above runs in your browser and reports everything it finds. Here is what it is reading, and where else each thing can be read.

• C2PA Content Credentials: a cryptographically signed manifest naming the tool that made the image and the edits since. The Content Credentials Verify page shows the same manifest in full detail.
• IPTC digital source type: one metadata field. The value trainedAlgorithmicMedia is the software stating outright that a generative model produced this.
• XMP creator tool: the application that last wrote the file, as it named itself. Several generators fill it in and most people never look.
• PNG text chunks: Stable Diffusion front ends, ComfyUI and InvokeAI write the prompt, seed, sampler and model straight into the file.
• EXIF capture data: camera make, model, lens, shutter and aperture. This points the other way, toward a physical device.
• SynthID: Google embeds an invisible watermark in output from Imagen, Gemini and Veo. Only Google can read it, through the SynthID Detector portal, or by uploading the picture to the Gemini app and asking whether Google AI made it.

Your eye is close to chance, and somebody measured it

Sophie Nightingale and Hany Farid published the test in PNAS in 2022. Asked to sort synthetic faces from photographs, participants averaged 48.2%. Chance is 50%. A second group of 219 people was trained on the artefacts to watch for and given feedback after every single image, and they reached 59.0%.

Both numbers describe faces made by a 2022 generator. Diffusion models have moved a long way since, and the artefacts those people were trained on are the ones that got fixed first. So read 59% as the ceiling for a coached observer under laboratory conditions, not as a description of you scrolling a feed.

The visual signs that still hold up

• Written text. Signage, book spines, shirt logos and shopfronts still come out warped or misspelled in a lot of output.
• Light that does not agree with itself. Shadows falling two ways, a window reflection showing the wrong room, a catchlight in one eye and not the other.
• Objects where they meet other objects. A bag strap merging into a shoulder, spectacle arms that never arrive at the ears, fingers that change shape behind a cup.
• Repetition in the background. The same face twice in a crowd, brickwork that tiles, leaves with no variation between them.
• Surfaces with no history. Skin with no pores, a kitchen with no crumbs, a jacket with no wear at the cuffs.

Questions people ask before they check anything

When the file says nothing at all

This is where most people actually are, and it is not a dead end. It just moves the question from the pixels to the history of the picture.

• Run a reverse image search and sort for the earliest copy, not the best match.
• Ask whoever sent it for the file as it came off their camera roll, sent as a file rather than through a messenger, which recompresses and strips.
• Look at the account that posted it. An account with years of ordinary posts behind it is a different proposition from one created in March.
• Look for the same moment from another angle. A real event photographed by one person and nobody else is worth a second thought.