Is this image AI-generated? A verification checklist, from credentials to detectors

Someone sends you a photo of a flooded street, a politician at a rally, or a product that looks a little too perfect, and you want to know whether it is AI-generated. The honest answer is that no single check settles it. What you can do is run the checks in order of strength, stop when one gives you real evidence, and be clear about what each result can and cannot show.

This checklist follows the same evidence ladder used across our guides to verifying images, documents and text: signed records first, labels and marks next, searches and visual inspection after that, and detector scores last. Each step below says what to do, what a positive result means, and what an empty result means.

The order of the checklist, and why it matters

The order is by strength of evidence. A signed credential can be checked cryptographically, so a valid one is hard to fake. A metadata label or a watermark is useful when present but easy to lose. A detector score is a statistical guess, and it should only add to evidence you already have.

Step 1: Check for content credentials

Content credentials are signed records attached to an image under the C2PA standard. The core of the record is a SHA-256 hash of the image's bytes plus a digital signature from whoever issued it, such as a camera maker, an editing app or an AI image service. When an AI image service attaches a credential, the record can name the generator, which is the strongest positive answer you can get.

Open the file in a credentials viewer; our walkthrough on checking content credentials in a browser shows each step with no command line. Look for three things: the signer, the list of actions (created, edited, generated), and the validation result.

The C2PA specification defines what that result means. A manifest can be Well-Formed, Valid or Trusted, and the specification's validation states say the image itself counts as Valid only if the content covered by the hash is unmodified and the manifest is Valid or Trusted. A credential that fails validation means the file changed after signing, which is worth knowing in itself.

What it cannot prove: a valid credential tells you who signed the record and what their tool reported. It does not tell you whether the scene in front of the camera was staged, and it says nothing about images that never had a credential.

Step 2: Read the metadata for an AI source label

Many generators and editors write a label into the file's metadata. The IPTC photo metadata standard has a field for this, Digital Source Type, and its controlled vocabulary includes trainedAlgorithmicMedia, which the IPTC Digital Source Type list defines as "Created using Generative AI". Related values cover composites that include AI elements and purely algorithmic images.

To see the label, open the file in a full metadata reader. ExifTool prints every tag, grouped by where it lives in the file:

exiftool -a -u -g1 photo.jpg

Look for a Digital Source Type entry, a software or creator tool field naming an AI service, and the dates. Our guide to reading an image's metadata on Windows, Mac or with ExifTool explains the other fields and what the built-in viewers show.

What it cannot prove: metadata is unsigned text inside the file. Anyone can delete it, and anyone can write a false value into it. A present AI label is good evidence that the creating software declared AI; a missing label proves nothing, because uploads, converters and screenshots routinely drop it.

Step 3: Look for an invisible watermark

Some providers embed a watermark in the pixels themselves, where metadata tools cannot see it. Google DeepMind describes SynthID as an imperceptible watermark for AI-generated images, audio, text and video, and says the image and video watermark is designed to survive edits such as cropping.

Checking for a pixel watermark needs the provider's own detection tool, so the answer is always about one provider. A found watermark is strong evidence that that provider's model generated or edited the image. The difference between a watermark, a file hash and a signed credential is laid out in our comparison of watermarks, hashes and signatures.

What it cannot prove: a negative result only tells you that this provider's mark was not detected. The image could come from a generator that does not watermark, or the mark could have been weakened by heavy editing.

Step 4: Search for earlier copies

A reverse image search often settles the question faster than any forensic tool. Run the image through two or three engines (Google Lens, Bing Visual Search and TinEye index different parts of the web) and sort by date where the engine allows it.

You are looking for the earliest copy and its context. If the same photo appears in a news agency archive from 2019, it was not made by a 2025 image model, though it may have been reused out of context. If the earliest copy sits on an AI art account, or carries a caption naming the prompt, that is strong context.

Search engines match images that look alike, even after resizing and recompression, using similarity fingerprints instead of exact hashes. That is the same idea explained in our page on perceptual hashes and SHA-256, and it is why a cropped or resized copy can still be found.

What it cannot prove: finding nothing means the image is new to the index, which is true of every genuine photo taken today. An earlier copy shows the image existed by that date; it does not show how it was made.

Step 5: Look at the image, carefully and with doubt

Visual inspection is where most people start, and it belongs near the end. Zoom in on the places generators have historically struggled with: hands and teeth, text on signs and labels, jewellery and earrings that do not match, reflections and shadows that disagree with the light, and repeated patterns in crowds, foliage or brickwork.

Treat anything you find as a reason to dig further. Newer models fix many of the classic tells, and real photos have their own oddities from lens distortion, motion blur, phone processing and compression. A clean-looking image is not evidence of a real one.

What it cannot prove: visual tells are judgement calls. They can point you to where to look; they cannot establish origin for you or for anyone you need to convince.

Step 6: Detectors, last

An AI image detector is a classifier trained on examples of real and generated images. It returns a score, and that score depends on how close your image is to what the detector was trained on. Images from newer generators, and images that have been compressed, resized or screenshotted, are exactly the cases where that match is weakest.

Researchers study this gap directly. A 2022 paper by Corvi and colleagues examined how detectors built for older GAN-generated images perform on images from diffusion models, with particular attention to the compression and resizing that social networks apply. Read any detector's own published error rates, and the base they were measured on, before you rely on a score.

If you use a detector, run more than one, keep the scores alongside the evidence from steps 1 to 5, and write down the tool names and the date. A high score with no other evidence is a hint. A low score does not clear an image.

What it cannot prove: a score is a probability with an error rate in both directions. It cannot name the generator, and it cannot show where the image came from.

What counts as proof

Only the top of the ladder produces evidence that someone else can check independently. A valid credential, a provider's confirmed watermark, or an earlier dated copy with clear context can each support a firm statement. Metadata labels support a softer one, and visual tells and detector scores support only "worth investigating".

Context matters as much as the image. Ask who first posted it, whether they have a history of genuine photos from that place, and whether anyone else captured the same moment from another angle.

When you report a finding, list each check, what it returned, and the date you ran it. If nothing came back, say the origin is unknown. That sentence is accurate, and it is more defensible than a confident guess in either direction.

Sources

1. C2PA Technical Specification 2.5, validation states
2. IPTC NewsCodes, Digital Source Type vocabulary
3. Google DeepMind, SynthID
4. Corvi et al., On the detection of synthetic images generated by diffusion models (arXiv 2211.00680)

Questions

• Can an AI image detector prove a picture is fake?+

  No. A detector gives a probability based on patterns it learned from training images, and it can be wrong in both directions. Use its score as a reason to look harder, and rely on signed credentials, labels and earlier copies for anything you have to defend.

• The image has no content credentials. Does that mean it is AI-generated?+

  No, it means nothing either way. Most cameras and apps still do not attach credentials, and many websites and messaging apps strip them when a file is uploaded. An absent credential simply moves you to the next step on the checklist.

• Can someone remove the AI label from an image's metadata?+

  Yes, easily. Metadata is plain data inside the file, and any editor, converter or screenshot can drop it. A label that is present is useful evidence; a label that is missing tells you very little.

• Is a screenshot of an image good enough to check?+

  Rarely. A screenshot is a new file made from pixels, so it carries none of the original's credentials or metadata. Ask for the original file, or search for the earliest copy online, before you draw a conclusion.

• What should I do if every check comes back empty?+

  Say that the origin is unknown. Report what you checked and what you found, and avoid calling the image real or fake on the strength of visual tells or a single detector score.