About SynthCheck

Ask the internet whether a picture was made by AI and you get handed a number. "87% likely AI-generated." It looks precise. It is usually invented.

Most detectors run an uncalibrated classifier over the pixels and dress the output up as a probability. The model was trained on last year's generators, never validated against the image in front of you, and the figure is a raw score with a decimal point added for confidence. Show it a real photograph taken in poor light and it will cheerfully call it synthetic. You cannot check the reasoning, because there is none to check.

SynthCheck does the opposite. It does not guess. It reads the provenance data actually written into the file, and shows you the evidence.

That data is real and increasingly common. C2PA Content Credentials — a signed manifest backed by Adobe, Microsoft, OpenAI and Google — record which system produced an image and what was done to it afterwards. The IPTC tag DigitalSourceType set to trainedAlgorithmicMedia is a flat declaration of AI origin; Midjourney and Meta write it. Stable Diffusion, ComfyUI, InvokeAI and Fooocus write the prompt, seed, sampler and model name straight into PNG text chunks. Cameras write EXIF: body, lens, shutter speed, ISO, sometimes GPS. Every line SynthCheck reports comes with the exact string it was read from, in an expandable panel. You can verify each claim yourself.

The honest part

Here is the position we will not move from.

A marker is strong positive evidence that an image is AI-generated. The absence of a marker proves nothing.

Metadata does not survive contact with the internet. Instagram strips it. WhatsApp strips it. A screenshot strips it. Re-saving in most editors strips it. When SynthCheck returns "inconclusive", it means this file carries no provenance information — not that the image is genuine. No honest tool can pull a verdict out of a file with nothing in it. Anyone returning a percentage anyway is filling that gap with fiction.

We state one more limit plainly: we cannot read SynthID. Google DeepMind embeds it statistically in the pixels themselves rather than the metadata, and there is no public detector for it. A Gemini image whose metadata has been stripped can still carry an intact SynthID watermark that SynthCheck will never see. Google operates its own SynthID Detector; we link to it in every result. Admitting the gap beats pretending it is not there.

Why now

Article 50 of the EU AI Act came into force on 2 August 2026. Providers of generative AI systems must mark synthetic output in a machine-readable format so it can be detected as artificially generated, with penalties reaching €15 million or 3% of worldwide annual turnover. Provenance marking is shifting from a voluntary industry gesture to a legal obligation. The markers SynthCheck reads will only get more common.

Free, no account, no upload. Your image is parsed inside your browser and never leaves your device.