AI Image Detector — Metadata Inspector for ChatGPT / Midjourney / SD / ComfyUI
Drop an image to determine from metadata whether it was AI-generated. Analyzes embedded metadata (C2PA Content Credentials, PNG textChunks, EXIF Software, XMP) from major AI tools including ChatGPT (DALL·E 3), Sora, Adobe Firefly, Midjourney v7, Google Imagen 3, Microsoft Designer, Stable Diffusion, ComfyUI, NovelAI, InvokeAI, Leonardo.AI, Krea, Flux, and Apple Image Playground. Fully processed in-browser; images are never sent to a server.
Drop an image or click to select
PNG / JPEG / WebP / AVIF · 100% browser-only
📡 Detected signals
▶ 📄 All metadata (raw)
▶ 📚 Supported AI tools (16+)
📖 Where people get stuck
Reads the metadata embedded in an image — C2PA Content Credentials, PNG text chunks, the EXIF Software field, XMP — and reports whether any major generator left a trace. Everything runs in the browser and the image is not transmitted. What this establishes is only whether a generator left a signature or a record — it is not a determination of whether the image is AI-generated. That distinction is decisive: metadata can be stripped with one command, and it disappears simply by taking a screenshot. Detecting nothing is not evidence of anything.
| Case | What happens | What to do |
|---|---|---|
| Detecting nothing does not prove it is not AI | Metadata is like a sticky note attached alongside the image content, and it is lost astonishingly easily. Taking a screenshot, uploading to social media, opening in an editor and re-saving, sending through a messaging app — every one of those drops metadata without anyone intending it. In other words, a single pass through an ordinary, innocent sharing route erases the trace. Removing it deliberately is, naturally, even easier: one command line. So when this tool reports nothing found, there are two possibilities: the image genuinely is not AI-generated, or it is and the metadata is gone — and there is no way to tell those apart from here. | Read the result as found or unknown, never as found or absent. A positive detection is close to certain information, whereas a negative yields no information at all — and that asymmetry should carry through into your reports and your interface. A tool that displays the words "not AI" is wrong at that moment. The fitting use is as a pre-filter over a large batch of images — process the detected ones as settled and put the rest in front of a human, and the weakness of the judgement stops mattering. Conversely it is unsuited to reaching a conclusion about a single image — what you need there is not detection but provenance, which is the next item. |
| Metadata can also be faked | The error runs in the other direction too. The EXIF Software field and PNG text chunks are plain text fields — anyone can write Software: Midjourney into a genuine photograph. Equally, anyone can write a DSLR model name and exposure parameters into an AI-generated image. Unsigned metadata therefore never rises above self-declaration. C2PA alone is different in kind — it carries a cryptographic signature, so tampering is detectable by verification. That too has a condition, though: it means something only when you actually verify the signature and trust the issuer — the presence of fields that look like C2PA and a successfully verified C2PA signature are entirely different claims. |
Treat signed and unsigned metadata differently. Regard text chunks and EXIF statements as hints, and never reach a conclusion on that basis alone. When C2PA is present, run it through an official verifier, such as the Content Credentials verify site, and confirm both that the signature is valid and who issued it — only at that point can you say it has been verified. One further practical note: pay attention to images whose fields contradict each other — EXIF asserting a camera model while XMP still holds the name of a generator is a meaningful trace of editing. Consistency between fields carries more information than any single field. |
| Trying to judge by visual tells | The number of fingers, garbled text, unnatural symmetry — these visual tells weaken with every model release. What was common knowledge a few years ago no longer holds, and the trend only moves further in that direction. More serious is the error in the other direction: concluding that something is AI because it looks unnatural harms people by declaring real photographs and real faces to be fake — camera shake, compression artefacts, an unusual composition, a retouched commercial photograph all look artificial. Automated AI-detection services have the same problem — the accuracy of published detectors falls sharply once the distribution of inputs changes, and the cost of a false positive can be very high depending on context. | Verify the provenance rather than the image itself. Who published it first, how long it has existed online, whether an older version exists elsewhere — reverse image search is far more dependable for this purpose than any detector. For a news or incident image, whether you can trace the photographer, the date and the place is the centre of the judgement — if you cannot, it is unusable regardless of whether it is AI. And where the outcome is consequential — journalism, a legal determination, suspending an account, a hiring decision — never let any automated detector be the sole basis, this tool included. If several independent lines of evidence do not line up, recording that you cannot determine it is the correct conclusion — that is not a failure; it is an accurate report. |
C2PA is not a mechanism for indicating whether something is AI; it is a mechanism for recording a history of who did what. It attaches to photographs from supporting cameras too, and passing through an editor accumulates who modified it, when and how — its purpose is provenance, not AI detection. Which is exactly why the absence of C2PA does not mean something is suspicious: use equipment or software that does not support it and a perfectly legitimate photograph carries none either. It is a technology still spreading, so the normal state, for a while yet, is not having it. As for where it earns its keep in practice: if you are the one publishing, it is worth configuring things to preserve C2PA — because it becomes material for showing later that you took or made this. Finally, this page only reads metadata and analyses no pixels at all — understand it as a different kind of tool from a detector that judges by looking at the image. Those two differ both in what they are good at and in how they fail.
📖 How to Use
-
1
Drop an imageFully in-browser. No server upload
-
2
Check detected signalsFacts from C2PA / PNG textChunks / EXIF / XMP
-
3
See prompts, models, seedsSD / ComfyUI fully retain prompt / model / seed
❓ Frequently Asked Questions
When detection fails?
100% accurate?
Privacy?
What is C2PA?
Can I see SD / ComfyUI prompts?
🐛 Found a bug or issue with this tool?
Free to use, no signup. Even just the steps to reproduce are helpful. Reports go directly to the operator and help us fix issues.
Thanks for your report!
Your report has been delivered to the operator and will be used to improve the tool.