How to tell if an image is AI
No free method can prove that a picture was, or was not, made by an AI image generator. You can read the evidence a file carries, learn how strong each kind is, and ask the right company's detector about what only it can read. This guide covers the five kinds of signal and says where each one stops.
To check a file now, use the New Watermark image checker. It reads the image in your browser and never uploads it. It is not a pixel-based AI image detector; those guess. It reads the signals a file carries.
The five kinds of signal
| Signal | What it can show | What it cannot show | Checked here? |
|---|---|---|---|
| Signed Content Credentials (C2PA) | Who signed a record, what it says, and whether the image changed since signing. | That the statement is true. Anything once the record is removed. | Yes, against two dated trust lists. |
| Unsigned metadata (IPTC, XMP, EXIF, PNG text) | A label a tool wrote, such as a source type or generation settings. | Anything reliable. Anyone can add, edit or delete it. | Yes, as declarations. |
| Invisible watermarks | That a mark from one system is in the pixels. | That nothing was added or removed. Editing can destroy a mark. | Only the Stable Diffusion reference watermark. |
| Visible marks | That a tool or person put a label on the picture. | Anything else. A mark can be cropped out or added to any picture. | No. A guide only. |
| Pixel-based detectors | An estimate from how the picture looks. | Proof. An estimate can be wrong either way. | No. |
Signed Content Credentials
Content Credentials are signed statements stored inside the file. A checker can test whether the signature is intact, whether the image still matches what was signed, and whether the signer's certificate is on a trust list. This site says "trusted" only when all three pass against the C2PA Trust List, and even then the record shows who made a statement, not that it is true. See What are Content Credentials?
Unsigned metadata
Software writes fields into images. The IPTC and XMP digital source type value trainedAlgorithmicMedia is the standard way to say that generative AI created the media. Automatic1111 stores generation settings in a PNG text chunk named "parameters", and ComfyUI stores "prompt" and "workflow" data unless that is switched off. Midjourney is reported to put the prompt and job ID in description fields. All of it is as easy to delete as to read, so we report these fields as declarations.
Invisible watermarks
A mark hidden in the pixels can outlast metadata, but only its owner's decoder can read most of them. Google says its detector can find SynthID in media from Google and partners including OpenAI, NVIDIA and Kakao. We cannot run that check for you; What is SynthID? explains how to. Meta has its own mark, and Adobe's TrustMark carries an identifier that must be looked up in the issuer's records.
The one watermark we decode is the one Stable Diffusion's open reference software adds. The Hugging Face diffusers library adds the SDXL version by default when the invisible-watermark package is installed. A match means a pattern is present, not who made the image, and a real photograph upscaled with that software would carry it too. The project's README says it survives JPEG compression, but our own tests found it fragile: it survived lossless files (7 of 7) and high-quality JPEGs saved without colour subsampling, while the ordinary JPEG compression most sites and apps apply removed it every time (0 of 14), as did resizing and cropping. So for an image that has been shared, not finding it means very little.
Visible marks
A label on the picture, such as Meta's "Imagined with AI" or the Gemini sparkle, shows that a tool or platform added it, or that someone drew it there. It says nothing when it is missing. See our guide to visible AI watermarks.
Pixel-based detectors
These classify the picture itself and ignore provenance. Sightengine's documentation says its AI image detection ignores metadata, EXIF tags, C2PA records and invisible watermarks. We run no pixel-based detector and cannot give an error rate for one, so treat any answer as a lead.
Why finding nothing proves nothing
- Signals get lost. Most images lose their metadata and records when they are shared, screenshotted or re-saved.
- Many tools write no signal. Midjourney is reported to attach no Content Credentials and to have no known invisible watermark. We found no documented signal for Ideogram, Leonardo or Canva. Open-weight Flux models add none, and ComfyUI can be set to write no metadata.
- Watermarks break. In our tests the Stable Diffusion watermark did not survive ordinary JPEG compression, resizing or cropping, and reports say Meta's preview detector missed cropped images.
- Visible marks are optional. It is reported that since August 14, 2026 the Gemini app lets people turn its visible mark off.
Our own small test agrees. On October 27, 2026 we ran 31 sample images that Wikimedia Commons contributors or vendors had labelled as made by tools including ChatGPT, Gemini, Midjourney and Adobe Firefly. Eight carried a Content Credentials record and 23 carried none, including every sample labelled Gemini, Imagen or Nano Banana. It is a sample, not a survey, and we cannot say why any file lacked a record.
"Nothing found" does not mean an image is real or human-made. It means these checks found no supported signal in this file.
Checks only the makers can run
Each link opens another company's website. You may need to sign in, your image is uploaded to them and not to us, and we are not affiliated with any of them.
- Google SynthID Detector checks for Google's watermark, which Google says also covers partner media.
- OpenAI Verify checks Content Credentials and SynthID on content from ChatGPT, OpenAI's API and Codex only.
- Adobe Inspect reads Content Credentials. Adobe says credentials can be backed up in its public cloud and recovered with it.
- TikTok's AI content checker is a public preview for marks from TikTok, CapCut, Dreamina, Hypic and Pippit only.
- Meta's identification tool was reported in July 2026 as a preview for Meta's own watermark on output from its newest image model.
A sensible order
- Find the original file. A screenshot, or a copy saved from a social platform, has usually lost its signals.
- Drop it into the checker. Read the headline, the findings under it, and the "not checked" line.
- If a record, a metadata field or a visible label names a company, try that company's detector.
- Run SynthID Detector as well. It is Google's own detector for its watermark, which we cannot check for you.
- Weigh the outcome with where the image came from. One signal is one piece of evidence.
Frequently asked questions
- Can anyone tell for certain whether an image is AI-generated?
- Not from the file alone. A trusted record shows who made a claim, and a watermark match shows a pattern is present. Nothing shows that AI was absent. Our checker reports evidence, never a conclusion.
- Does "no signals found" mean the image is real?
- No. Most images lose these signals when shared, and many tools never add them.
- Does this site check SynthID?
- No. See What is SynthID?
- Why does a file made with ChatGPT say "signer trust not established"?
- OpenAI's signing certificates are on neither list we check, so the record is well formed but not trusted. It also means we cannot tell a genuine OpenAI file from one signed with a certificate someone made and named after OpenAI. Privacy and method explains this in full.
- Is my image uploaded?
- No. The checks run in your browser and the page contacts only our own domain. Anything you upload to another company's detector goes to them. See Privacy and method.