How to Tell If an Image Is AI-Generated

The five checks, in the order to try them

1. Read the labels in the file. Content Credentials and metadata can say outright that AI made or edited an image. Valid ones are your best evidence. Missing ones tell you nothing.
2. Ask the company that made it. Google and OpenAI each let you check whether an image carries their invisible watermark or credentials.
3. Trace the image to its source. A reverse image search often finds the first copy and the story behind it.
4. Look closely, but don’t lean on it. Hands, text and reflections show you where to dig, and they prove very little either way.
5. Run a detector, and read its answer as an estimate. It works when the labels are gone, and it can be wrong.

Then weigh the evidence together. Each check answers a different question, and they are strongest in combination.

1. Read the labels in the file

Many AI tools label what they make. The labels live inside the file, in two layers that differ a great deal in how far you can trust them.

• Content Credentials (C2PA). An open standard for signed provenance records. A tool that follows it adds a record to the file saying which tool made it, when, and what has happened to it since. The record is cryptographically signed, so a change to the file or to the record makes the check fail. Adobe says it applies them automatically to assets whose pixels are all generated with Firefly, and OpenAI documents them on its images too, though not on every one. Some cameras and phones sign photos the same way.
• Standard metadata. The IPTC DigitalSourceType field has values for this: trainedAlgorithmicMedia (“Created using Generative AI”) and compositeWithTrainedAlgorithmicMedia (“Edited using Generative AI”, for example inpainting or outpainting). The EXIF Software field sometimes names a generator, and Stable Diffusion tools such as AUTOMATIC1111 and ComfyUI usually write the prompt and settings into the PNG. None of this is signed: anyone can add, change or delete it.

To look, drop the image into a checker. Our AI image detector lists the metadata it finds and whether the file holds Content Credentials. The Content Authenticity Initiative’s Verify tool checks the credentials and shows the full record.

How to read what you find:

• Valid credentials that declare AI. The signer is telling you that AI made or edited the image. That is strong evidence, and only as good as the signer, so look at who signed.
• Valid credentials that declare a camera. The signer says a camera captured it. That supports authenticity without proving it: a signed photo can still show a staged scene, or a screen.
• Credentials that fail. The file was changed after it was signed, or the credentials were written wrongly. Set aside what they claim.
• Nothing at all. This tells you nothing. Screenshots, re-saves and most social networks strip metadata, so most images reach you without any. Some issuers, Adobe among them, also keep a copy of the credentials online so that they can be recovered after a file has been stripped.

2. Ask the company that made it

Labels in a file are easy to lose. Invisible watermarks are harder to lose, because they are woven into the pixels themselves. Only the company that added a watermark can read it, and two of them now let you check.

• Google. Google says its SynthID watermark is embedded in the images, audio, text and video made by its generative AI consumer products, and that people can’t see it. You can upload an image to Gemini and ask whether it was created or edited by Google AI, and Gemini checks for the watermark.
• OpenAI. OpenAI’s documentation describes a web tool at openai.com/verify that checks an image for OpenAI’s Content Credentials and for its SynthID watermark.

These checks are strong when they hit and weak when they miss. Each covers only its own maker’s output. OpenAI says its tool doesn’t detect content from another company’s models, and that a miss doesn’t rule out its own: metadata can be stripped, a watermark can be degraded, and older models didn’t add either. FreeAIDetector doesn’t read invisible watermarks at all.

3. Trace the image to its source

Where an image came from often settles the question faster than anything in it. A reverse image search (Google Lens, TinEye and Bing Visual Search all offer one) finds other copies of the picture. If the tool lets you, sort the results by date, and look for the earliest copy, a larger original, and whoever posted it first.

• A first appearance in a gallery of AI art, or on a generator’s share page. That answers it.
• A first appearance earlier than the event the image is said to show. Then it can’t show that event, whether or not AI was involved.
• No other copies at all, for a supposedly big news event. Real events produce many photos and videos from many angles. One striking image and nothing else is a warning sign.
• A credible original source. A photographer, an agency or an artist can tell you how it was made. Ask.

The SIFT method, from the digital-literacy researcher Mike Caulfield, is a good checklist for this: Stop, Investigate the source, Find better coverage, and Trace claims, quotes and media to their original context. It asks you to judge the claim before you judge the pixels.

If the file still holds EXIF data, the camera model, the capture time and the GPS position are worth a look. Treat them as a lead, not as proof: EXIF is easy to edit, and it is usually stripped before an image reaches you.

4. Look closely, but don’t lean on it

People are worse at this than they think. In a study of faces made with an older generator, StyleGAN2, 315 participants sorted real from synthetic faces with an average accuracy of 48.2%, against 50% for guessing, and the synthetic faces were rated 7.7% more trustworthy on average (Nightingale and Farid, 2022). A second group given training and feedback reached 59%. A meta-analysis of 56 papers and 86,155 participants found that people spotted deepfake images with an average accuracy of 53.16%, and the 95% confidence interval (42% to 65%) includes guessing (Diel and colleagues, 2024).

These are the things people still check, and what to check for:

• Hands and fingers. Extra, missing or fused fingers, and joints that bend the wrong way.
• Text and signs. Letters that almost spell words, or lettering that changes from one line to the next.
• Repeats and symmetry. Identical faces in a crowd, tiled textures, earrings or glasses that don’t match.
• Light and reflections. Shadows that disagree about where the sun is, and reflections that show a different scene.
• Edges and backgrounds. Hair that melts into the background, objects that merge, architecture that doesn’t add up.
• Surface. Skin and fabric that look unnaturally smooth or waxy.

Use the list to decide where to zoom in, not to decide the answer. These tells fade with every generation of models, so a clean image proves little. Real photos have odd artifacts of their own: compression, panorama stitching, HDR merging and beauty filters. Once you suspect an image, you will find flaws in almost any picture. And a real photo can be edited with AI in a single spot, which no overall impression will reveal.

5. Run a detector, and read its answer correctly

An AI image detector looks at the pixels. A model trained on real and generated images picks up statistical traces that people can’t see, such as patterns of noise and texture, and returns an estimate. Some detectors, ours included, read the file’s labels first.

It is the only check here that still works when every label has been stripped. It is also the least certain, and it fails in known ways:

• Compression and resizing. Screenshots, heavy JPEG compression, upscaling and re-encoding by social networks all smear the traces a model looks for.
• New generators. A model recognizes what it was trained on. Images from a generator released after its training can score as real.
• Heavy retouching. Beauty filters, phone night modes and AI upscalers change a real photo’s pixels in ways that resemble generation.
• Edited rather than generated. One object added or removed with generative AI leaves most of the image untouched.
• Deliberate evasion. People can add noise, re-render or paint over an image to get past a detector.

So read the answer as an estimate. A middle score is an answer: “uncertain” means the evidence doesn’t decide it. Two tools that disagree are telling you something. Prefer a tool that shows the evidence behind its number, as our AI image detector does, to one that gives only a percentage, and never accuse a person or a publication on the strength of a score. How do AI detectors work? explains what they measure and why they fail.

Putting the evidence together

Each check answers a different question. This is what a set of findings lets you say.

Mistakes to avoid

• Treating a missing label as proof of a real photo. Or as proof of AI. It is neither.
• Trusting one number. A score without evidence, or one that ignores “uncertain”, invites overconfidence.
• Reading “no visible flaws” as “real”. Modern generators often leave none.
• Confusing “AI-edited” with “AI-generated”. A real photo with an object removed, or improved by a phone’s AI, is not an AI image. The IPTC vocabulary keeps the two apart, and so should you.
• Doubting real images. Once fakes are common, a genuine image can be dismissed as “AI” to escape an inconvenient truth. The same standard of evidence applies in both directions.
• Accusing a person. A check tells you about a file. It doesn’t tell you who made it, or why.

What we saw on real files

Content Credentials leave the least room for opinion, so we tested how they read in practice. On September 30, 2026 we ran an open-source C2PA reader over 16 public sample files: 14 from the C2PA project’s public test files and 2 demo images from the Content Authenticity Initiative’s Verify site. These are test cases, some of them broken on purpose, so the counts show what each kind of result looks like, not how common credentials are.

Three things stand out. A valid signature is not a recognized signer: the two Adobe images pass every check and still come back as “signer not recognized”. A failed check is not proof of fakery: it means the file and its credentials no longer agree, and an edit, an expired certificate and a faulty writer all look alike. And the AI label inside a credential is the signer’s statement, worth what the signer is worth.

Sources

• C2PA Technical Specification 2.2. What a Content Credentials record contains, and how it is signed and validated.
• IPTC NewsCodes: Digital Source Type. The standard values that say whether media was captured, edited or created with generative AI.
• Adobe: Content Credentials overview (Firefly). Adobe applies Content Credentials automatically to assets generated entirely with Firefly, and can store a copy online.
• OpenAI: Content provenance. OpenAI’s checks for Content Credentials and SynthID, and their limits.
• Google DeepMind: SynthID. What SynthID is, and how to check an image for it in Gemini.
• Content Credentials: Verify. The Content Authenticity Initiative’s tool for inspecting credentials.
• Nightingale and Farid (2022), PNAS. People’s accuracy at telling real faces from synthetic ones, and how trustworthy they rated them.
• Diel and colleagues (2024), Computers in Human Behavior Reports. A meta-analysis of 56 papers on human detection of deepfakes.
• Mike Caulfield: SIFT (The Four Moves). A short method for judging a claim before judging the pixels.
• C2PA public test files. Fourteen of the sample files behind our table of results.
• Content Authenticity Initiative: Verify site. The two Adobe demo images behind our table of results.

Facts and figures were checked against these sources on September 30, 2026. If something is out of date or wrong, tell us.