What this tool checks
Besides the photo or picture itself, an image file has places where the story of how it was made can be written. This tool reads the following four.
Provenance (C2PA / Content Credentials)
An international standard that embeds the history of how an image was created and edited, with a digital signature. If the signing organization is on the C2PA trust list and the image has not been altered since it was signed, the record can be verified as authentic. If that record also shows AI generation or editing, the tool shows "This image was made with AI". If the signing organization cannot be confirmed, it stops at "may have been made with AI".
IPTC digital source type
A metadata standard used for news photos and similar images has a value that means "generated by AI" (trainedAlgorithmicMedia). Because it carries no signature, it is treated as "may have been made with AI".
Generation parameters embedded in the PNG
Some image generators that run on your own computer write the prompt and settings into the PNG. If these are found, the tool shows "may have been made with AI".
Software and creator tool fields
The field that holds the name of the software that saved the image sometimes contains the name of a generator. This is also treated as "may have been made with AI".
The three kinds of results
The tool does not give a number like "X% AI". Depending on how reliable the records it finds are, it shows one of the following three.
This image was made with AI
Confidence: Certain (confirmed by signature)
Shown when the image contains a record saying it was made or edited with AI, signed by an organization that can be checked against the trust list.
This image may have been made with AI
Confidence: Possible (unverified information)
Shown when there are traces of AI, such as generator settings or a tool name, but only in information anyone can add.
We couldn't tell whether this image was made with AI
Confidence: Nothing to go on
Shown when no record we can base a judgment on is found. Most images posted on social media and most screenshots end up here.
What this tool can't tell you
- It cannot identify images whose records have been removed, or images from generators that never add records in the first place.
- Invisible watermarks (such as Google's SynthID) require dedicated detection methods and cannot be read by a browser alone.
- The information behind a "may have been made with AI" result has no signature, so people can add it themselves.
- "Couldn't tell" does not mean the image was not made with AI.
Visual clues
When no evidence is found, looking closely at the details can sometimes give you a clue. However, image generation keeps getting better, and none of these is conclusive.
- Do the number and shape of hands and fingers, or the way accessories connect, look unnatural?
- Does text on signs or clothing form readable letters?
- Do things that should match on both sides (glasses, earrings, etc.) differ from each other?
- Do shadows and reflections point in a direction consistent with the light?
- Do straight lines in the background (pillars, window frames) bend or break partway?
Privacy and safety
- Your image is not sent to a server. All reading happens inside the browser that has this page open.
- Results are not sent to analytics or anywhere else either.
- The tool does not connect to external URLs written in the image (including provenance references).
- The component that verifies provenance signatures (about 9MB) and the trust list are loaded from this site only when the image contains provenance.
- The trust list from the C2PA Conformance Program (CC BY 4.0) and the certificate usage settings from the Content Authenticity Initiative (Apache-2.0) are copied to and used from this site.
What each AI image service leaves in its images
Whether this tool can identify an image depends on what the generator leaves inside it. Below is what each major service's official documentation and official source code say (checked on 2026-10-05). "Not documented" means the vendor doesn't say; it doesn't mean nothing is there.
| Service | What's embedded | Can this tool detect it? | Official source |
|---|---|---|---|
| OpenAI Images from ChatGPT, Codex and the API | What's embedded Both C2PA (signed by OpenAI) and the invisible SynthID watermark. OpenAI says coverage can vary by product, model, export path and file type. | Can this tool detect it? C2PA: yes. Eight PNGs downloaded from ChatGPT all showed "This image was made with AI" (tested by this site on 2026-10-02). SynthID: no (check it at openai.com/verify). | Official source OpenAI Help Center |
| Google Gemini app, Gemini API, Google Cloud (including Imagen) | What's embedded Gemini app images carry C2PA and SynthID. On Google Cloud, images from supported models (such as Gemini image models) are automatically signed with C2PA by Google LLC. Imagen is not on that list of supported models. The Gemini API docs mention only SynthID. | Can this tool detect it? C2PA: yes, when present (not yet tested with real Gemini images). SynthID: no (check it in the Gemini app). | Official source Gemini Apps Help Google Cloud: Content Credentials Gemini API docs |
| Adobe Firefly Images fully generated with Firefly, such as Text to Image | What's embedded C2PA (issuer: Adobe Inc.) is applied automatically, and a copy may be stored in Adobe's cloud. Invisible watermark: not documented (as of 2026-10-05). | Can this tool detect it? C2PA: yes (not yet tested with real images). | Official source Adobe Firefly Help |
| Microsoft Copilot, Bing Image Creator, Designer | What's embedded The official Copilot FAQ says images generated in Copilot carry C2PA. For Bing Image Creator, a 2023 official blog post said all AI images carry Content Credentials, but the current FAQ doesn't mention it. Designer on its own: not documented (as of 2026-10-05). | Can this tool detect it? C2PA: yes, when present (not yet tested with real images). | Official source Microsoft Support (Copilot) Bing Blog (2023) |
| Midjourney | What's embedded Midjourney says every image contains a small "ID tag" inside the file. The format (C2PA, XMP or something else) is not disclosed (as of 2026-10-05). | Can this tool detect it? Unknown, because the format isn't public. To check, use the official midjourney.com/verify (free, no account needed). | Official source Midjourney Docs |
| AUTOMATIC1111 / ComfyUI / InvokeAI UIs you run yourself for Stable Diffusion, FLUX and others | What's embedded No C2PA. They write the prompt and settings into PNG text fields (AUTOMATIC1111: parameters; ComfyUI: prompt and workflow; InvokeAI: invokeai_metadata and others). AUTOMATIC1111 writes to EXIF for JPEG and WebP. This can be turned off in settings. | Can this tool detect it? Yes. Because nothing is signed, it shows "This image may have been made with AI". | Official source GitHub: AUTOMATIC1111 GitHub: ComfyUI GitHub: InvokeAI |
| FLUX (Black Forest Labs) BFL API and Playground / official code you run yourself | What's embedded Images from the BFL API and Playground carry C2PA (signed by BFL); BFL says a watermark is embedded optionally. The official reference code, when you run it yourself, writes "AI generated;txt2img;flux" or similar to the EXIF Software field and embeds a watermark with the invisible-watermark library. | Can this tool detect it? C2PA: yes (whether the signer can be confirmed on the trust list is untested). The Software field from the official code is also read and shows "This image may have been made with AI". The watermark is not read. | Official source BFL Help Center GitHub: black-forest-labs/flux |
- All of this applies to the original downloaded file. Posting to social media, taking a screenshot, editing or converting the format can remove metadata, including C2PA.
- Even when C2PA is present, if the signing organization can't be confirmed against the C2PA trust list, the tool only says "This image may have been made with AI".
- This tool can't read invisible watermarks (such as SynthID). Use each vendor's own checker.
FAQ
I got "We couldn't tell whether this image was made with AI". Does that mean it isn't AI?
No. It only means no records were found. Images posted on social media and screenshots have almost always lost their records.
If it says "This image was made with AI", is that certain?
It is certain that a signed record states AI generation or editing and that it has not been altered since signing. However, how much AI was involved (whether the whole image was generated or only part of it was edited) depends on what the record says. Check the original text in the evidence section.
Is my image stored anywhere?
No. Your image is not sent to this site's server or anywhere else.
Does it work on a phone?
Yes. Only for images that contain provenance (C2PA), it loads a component of about 9MB the first time.
Can I check an image by entering its URL?
No, that isn't supported. Having the server fetch external images poses security problems, so please save the image first and then add it.
Related articles
- ChatGPT, Codex, and Claude Code Image Generation: How They Differ, Pro Usage Limits, and a Practical Workflow
- Does an AI-Written Article Need an "AI-Generated" Label? Reading EU AI Act Article 50 in Practice
- Getting Started with AI Image Generation — How It Works, the 4 Steps, the Image-Prompt Anatomy, and Rights