Guides
How to Check if an Image Is AI-Generated (5 Real Checks)
Check an image for AI: Content Credentials (C2PA), the IPTC digital source type, SynthID and other provider detectors, reverse search and visual clues.
To check if an image is AI-generated, look for provenance signals first: Content Credentials (C2PA), the IPTC digital source type in the metadata, and invisible watermarks such as Google's SynthID, which the Gemini app and the SynthID Detector can check. Then use reverse image search to find the original, and only last rely on visual clues. A positive signal is strong evidence; a missing one proves nothing, because metadata is easy to strip and not every AI tool marks its output.
The five checks, from strongest to weakest
| Check | What it finds | Strength | Main limit |
|---|---|---|---|
| Content Credentials (C2PA) | A signed record of the tool and edits, including AI generation | Strong when present and valid | Removed by screenshots and many uploads |
| Provider watermark detector | An invisible watermark such as SynthID | Strong for that provider | Only detects its own (and partners') watermarks |
| Metadata fields | IPTC digital source type, software name | Medium | Easy to strip or edit |
| Reverse image search | Earlier copies, the original source, fact-checks | Medium to strong | Finds nothing for new images |
| Visual clues | Odd hands, text, reflections, physics | Weak | Modern models make few such mistakes |
1. Look for Content Credentials
C2PA Content Credentials are a signed manifest attached to the file that records which device or app made it and what was done to it. AI tools from several large providers add them, and some cameras and editors do too. The manifest can say the image was created by a generative model, or that a real photo was edited with one.
- Save the original file, not a screenshot.
- Open contentcredentials.org/verify and drop the file in, or check for a CR icon on sites that display it.
- Read the result: the signer, the app or model, the date, and actions such as "created using generative AI" or a list of source images.
To see what the file declares without uploading it anywhere, our AI image checker reads the manifest, the IPTC source type and generator tags in your browser; it shows what the file says but doesn't verify signatures.
If a tool reports no credentials, that only means this copy has none. What C2PA Content Credentials are explains what the manifest records and what "valid" means.
2. Check with a provider's watermark detector
Some providers hide a watermark in the pixels that survives cropping, compression and filters better than metadata does.
- Google SynthID: in the Gemini app, upload the image and ask "Was this created with Google AI?". Google's help page says Gemini checks for SynthID and also reads Content Credentials. Google DeepMind's SynthID Detector portal checks content from Google and from partners it names, including OpenAI and NVIDIA. Gemini limits how many checks you can run per day.
- OpenAI: its help center says ChatGPT and API images carry C2PA metadata and SynthID, and it offers a verification page for its own images.
How to read the result: a detection means the provider's model created or edited all or part of the image. No detection means no watermark from that provider was found. Another tool may still have made it, and Gemini itself sometimes answers that an edit is too small to judge.
3. Read the metadata
Metadata won't survive most social uploads, but files sent by email, downloaded from a website or shared as documents often keep it.
- Open the image in our photo metadata viewer. It runs in your browser and doesn't upload the file.
- Look at Software, Make and Model. A camera make and model suggest a real capture but can be faked; a generator or editor name is a clue.
- Look for the IPTC Digital Source Type. Common values:
| Value | Meaning |
|---|---|
digitalCapture | Captured by a camera |
computationalCapture | Several real frames merged automatically, such as smartphone HDR |
trainedAlgorithmicMedia | Created by a generative AI model |
compositeWithTrainedAlgorithmicMedia | A real image edited with generative AI (inpainting, outpainting) |
compositeSynthetic | A composite that includes at least one AI element |
On iPhone, the Photos Info panel shows device, lens and settings; how to check photo metadata covers it step by step.
4. Do a reverse image search
Searching for the image itself often answers the question faster than any detector.
- Use Google Lens (in the Google app or Chrome), Bing Visual Search or TinEye.
- Sort or scan for the earliest copy. A photo of a real event usually has an origin: a news agency, a photographer, a public account.
- Check whether fact-checkers have already covered it.
- In Google results, About this image can show when Google first indexed it and, when C2PA data is present and signed by a trusted issuer, whether it was made or edited with AI.
If the image only appears in the last few days, from accounts that post AI art, that's a strong hint. To find where a real photo was taken, see how to find where a photo was taken.
5. Look for visual clues, carefully
Older models left obvious traces; newer ones rarely do. Still worth a look:
- Text: signs, labels and logos that dissolve into near-letters.
- Hands, teeth, ears, jewelry: extra or merged parts, asymmetric earrings.
- Physics: shadows pointing different ways, reflections that don't match, impossible architecture.
- Texture: skin that's too smooth, repeating patterns in hair, crowds or foliage.
- Context: a scene that would have many photographers but exists in only one picture.
None of these is proof. Real photos have odd shadows and compression artifacts; AI images can be flawless. Generic "AI detector" websites that score pixels are useful as a second opinion at most, and they misjudge heavily compressed or edited images in both directions.
Putting it together
- Signed Content Credentials say AI, or a provider detector finds its watermark: treat it as AI-generated or AI-edited.
- No signals, but a clear earlier source from a credible outlet: likely real, at least as published there.
- No signals and no source: unknown. Don't share it as fact.
Where sigil. fits
sigil. doesn't detect AI images. It's an encrypted vault for your own photos and documents, opened with a sign you draw. It strips EXIF and location metadata when you import or export, which protects your privacy; it doesn't add or remove AI labels. If someone sends you an image you need to keep as-is, such as evidence of a scam, keep the original file safe and unedited.
Bottom line
Check for Content Credentials, run the provider's watermark detector, read the metadata and search for the original. A positive result is meaningful; a blank one isn't. When in doubt, don't treat the image as proof of anything.
Sources
Frequently asked questions
Is there a reliable way to tell if an image is AI-generated?
No single check is reliable on its own. A signed Content Credential or a provider watermark detection is strong evidence; their absence proves nothing, because metadata is easily stripped and not every AI tool adds a watermark.
How do I check an image for Content Credentials?
Upload it to the free Verify tool at contentcredentials.org/verify, or look for the CR icon on sites that display it. If a manifest is present, you'll see who signed it, the tool used and any AI actions recorded.
Can Gemini tell if an image was made with AI?
The Gemini app can check whether an image carries Google's SynthID watermark and can read Content Credentials. A negative result only means it found no Google watermark; the image could still come from another AI tool.
Do AI detector websites work?
Generic detectors that guess from pixels make mistakes in both directions, especially on compressed or edited images. Treat their score as a hint, not a verdict.
Can I see AI labels in a photo's metadata?
Sometimes. Look for an IPTC Digital Source Type such as trainedAlgorithmicMedia, or a software field naming an AI tool. Many apps and platforms remove metadata, so an empty result is common.