How to tell if an image is AI-generated

C2PA Content Credentials are a tamper-evident manifest that some tools attach to the files they produce. It records which software created the asset and what was done to it afterwards. OpenAI attaches them to DALL·E output, Adobe to Firefly, and Google to several of its image products; Leica and Sony ship cameras that sign genuine photographs the same way.

The manifest is stored differently depending on format — an APP11 segment in JPEG, a caBX chunk in PNG, a C2PA chunk in WebP — but the effect is the same: the file names its own origin. When a manifest is present and names a generator, that is the strongest evidence you will get from any method.

One caveat worth knowing, because most tools do not mention it: detecting a manifest is not the same as verifying it. Full verification means checking the cryptographic signature against a certificate chain. A tool that only reads the manifest, as most web-based checkers do, would report a hand-crafted manifest at face value. Treat it as what the file declares about itself.

Photographs carry a distinctive trail: camera make and model, lens, aperture, shutter speed, ISO, focal length, and often GPS coordinates. Generated images almost never have any of it. That contrast is the most practical everyday signal there is.

Generators sometimes leave their own fingerprint in the same place. A `Software` or `Creator` field reading "Midjourney", "DALL·E" or "Adobe Firefly" is conclusive. So is an IPTC `DigitalSourceType` field set to `trainedAlgorithmicMedia`, which is the standards-body marker for synthetic media.

Read this evidence in one direction only. Its presence tells you a great deal; its absence tells you almost nothing, because messaging apps and social platforms strip metadata from everything you send through them.

This is the single most overlooked check, and on PNG files it is often decisive. Stable Diffusion, Automatic1111 and ComfyUI write the entire generation recipe into the image as a plain-text chunk: the prompt, the negative prompt, the sampler, the CFG scale, the seed and the model name.

Almost nobody removes it, because almost nobody knows it is there. If a PNG contains a `parameters` block reading something like `Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1234567890`, the question is settled — you are not detecting anything, you are reading the recipe.

This survives only in PNG. Saving as JPEG destroys it, because JPEG has no equivalent chunk.

If the file has been stripped of everything above, statistical detection is what remains — machine-learning classifiers trained to recognise the artefacts diffusion models leave behind.

Be realistic about them. They are trained on particular generators and lose accuracy sharply on newer ones, with no warning sign when they do. Published accuracy figures come from test sets resembling the training data; real-world files rarely do. A classifier that scores 95% in a paper may be near a coin-flip on an image from a model released after it was trained.

This is why a screenshot is so destructive. It discards every layer of real evidence and leaves only the weakest one. If someone sends you a screenshot of an image and asks whether it is AI, the honest answer is usually that it cannot be determined.

Google embeds SynthID, an invisible watermark encoded into the pixels themselves, in output from its image models. Unlike metadata it survives cropping, compression and re-encoding, which makes it far more robust than anything described above.

The catch is that only Google can read it. No third-party tool can check SynthID, and any service claiming to is not doing what it says. The same applies to most proprietary watermarking schemes.

Run the checks in order: Content Credentials, then EXIF and XMP, then PNG text chunks, then — reluctantly — a classifier. Stop as soon as one gives a clear answer, because each step down the list is substantially weaker than the one above it.

And accept that for a large share of images circulating online, the correct answer is "cannot be determined". A tool that always produces a confident percentage is not more capable than one that admits uncertainty; it is just less willing to tell you when it is guessing.