Telling a real photo from a generated one rests on two kinds of evidence: the pixels, which an AI image detector scores, and the file's provenance, meaning C2PA Content Credentials, IPTC labels and EXIF. These are the terms you'll meet in both.
- AI-generated image
An image made by a generative model, such as a diffusion model or a GAN, from a text prompt or another image, rather than captured by a camera. It may be fully synthetic, or a real photo with generated parts (a composite).
- AI image detector
A tool that estimates whether an image was generated. Classifier detectors score the pixels from 0 to 1; provenance checks read labels the generator attached, such as C2PA or IPTC. Scores are probabilities, not proof.
- C2PA
The Coalition for Content Provenance and Authenticity's open standard for attaching a signed record of how a file was made and edited. The record, a manifest, lives inside the image file and can state that a step used AI.
- Content Credentials
The public name for C2PA manifests, used by Adobe, OpenAI, Google, Microsoft and camera makers. They name the tool that made the image and its edit history. Most social platforms strip them on upload.
- IPTC Digital Source Type
A standard metadata field describing how an image was made. The value trainedAlgorithmicMedia means a model generated it; compositeWithTrainedAlgorithmicMedia means a real image with AI-generated parts. Meta and Google use it.
- SynthID
Google DeepMind's invisible watermark for content made with Google's AI tools. It is embedded in the pixels and only Google's tools can read it. Gemini images also declare it in their C2PA manifest, which other tools can read.
- Invisible watermark
A signal hidden in an image's pixels that survives some editing and can only be read by the vendor's own detector. Unlike metadata, it isn't removed by re-saving, but a missing watermark proves nothing about other generators.
- EXIF
Metadata a camera or phone writes into a photo: make, model, lens, exposure, capture time and often GPS location. It suggests a real camera took the picture, but it can be copied or edited, and platforms usually strip it.
- Image provenance
The verifiable history of an image: where it came from and what was done to it. In practice it means C2PA manifests, IPTC labels and EXIF, read from the file. Intact provenance is the strongest evidence of origin.
- Deepfake
Synthetic media that makes a real person appear to say or do something they didn't, usually by swapping or generating a face or voice. A photo deepfake can be checked like any AI image; video and voice need specialised tools.
- Diffusion model
The kind of model behind most current image generators, including Stable Diffusion, FLUX, Midjourney and DALL·E. It learns to turn random noise into an image step by step, guided by a text prompt.
- GAN (StyleGAN)
A generative adversarial network: two models trained against each other until one produces convincing images. StyleGAN faces, the "this person does not exist" kind, are still common in fake profiles.
- Face detection
Finding faces in an image and returning where each one is, with a confidence score. It doesn't identify anyone. It's used to require one face in a profile photo, count people, or blur faces.
- Face recognition
Working out who a face belongs to by comparing it with other faces, using a face embedding. It's regulated in many places, and it's what PhotoCheck deliberately doesn't do.
- Confidence score
A number from 0 to 1 a model gives its own output: how strongly an image reads as generated, or how sure a detector is that a box holds a face. Thresholds turn it into a decision; the right one depends on the cost of a mistake.
- Reverse image search
Searching the web for copies of an image. It finds stolen real photos, but a freshly generated face usually has no copies, which is itself a clue on a profile that claims to be a real person.