How to check if an image is AI-generated
- Get the original file if you can — the download from the app or the sender, not a screenshot. Labels survive in originals and are lost in copies.
- Open it: drop one or more images on the tool, tap Check an image, or paste with Ctrl/Cmd+V. JPG, PNG, WebP, HEIC and AVIF work.
- Read the verdict and the list under it. Each line says what was found and where in the file it was found.
- If the result is “No evidence either way”, follow the steps further down this page — that result is common and does not mean the image is real.
How this AI image checker works — and how accurate it is
Most AI image tools now label what they make. ChatGPT, Adobe Firefly, Microsoft and Google attach Content Credentials, a signed C2PA manifest that records the app and a digitalSourceType of “trainedAlgorithmicMedia”. Many apps also write the IPTC source-type tag, and Stable Diffusion front ends save the full prompt and settings in the PNG. This page reads those blocks byte by byte in your browser and also looks for Gemini's visible corner sparkle. When a label is there, the answer is close to certain, because the generator wrote it. The limit is the opposite case: labels are metadata, and a screenshot, a crop-and-save, or an upload to most social apps strips them. So the detector never reports “real” — only “labelled as AI”, “camera data”, or “no evidence”. It does not run a pixel classifier that guesses a percentage; those tools are often wrong on compressed or edited images, and a confident wrong number is worse than an honest blank.
What the detector looks for
| Signal | Written by | What it tells you |
|---|---|---|
| Content Credentials (C2PA) with “trainedAlgorithmicMedia” | ChatGPT / DALL·E, Adobe Firefly, Microsoft Designer, Google | Generated by AI, and by which app |
| C2PA with “compositeWithTrainedAlgorithmicMedia” | Photoshop generative fill, other AI editors | A real image changed with AI |
| IPTC / XMP digital source type | Many generators and stock sites | Same meaning as above, without a signature |
| Saved prompt and settings | Stable Diffusion (A1111, Forge), ComfyUI, NovelAI, InvokeAI, Fooocus | Generated, plus the prompt, model and seed |
| Midjourney job text | Midjourney downloads | Generated by Midjourney |
| AI app name in Software / CreatorTool | Various apps | Made or saved in an AI app (weaker) |
| Visible sparkle in the corner | Gemini / Nano Banana | Generated by Gemini (pixel match) |
| Camera make, model and exposure | Phones and cameras | Probably a photo — EXIF can be faked |
Accuracy and limits — read this before trusting any detector
- A positive is strong. An AI label is written by the generator itself, so false positives are very rare. Someone could add a fake label, but there is little reason to.
- A negative is weak. Screenshots, re-saves, WhatsApp, Instagram, X and most messaging apps remove metadata. Our own test: the same ChatGPT picture goes from “Labelled as AI” to “No evidence” after one JPG re-save (try both samples above).
- Signatures are read, not verified. The page shows what the C2PA manifest claims; to check the cryptographic signature and edit history, open the file at contentcredentials.org/verify.
- Invisible watermarks are out of reach. Google's SynthID and OpenAI's invisible watermark are hidden in the pixels and can only be checked by their owners' tools.
- Camera EXIF is a hint, not proof. It can be copied from another photo or typed in with any EXIF editor.
Why there is no “87% AI” score here
Pixel-based detectors train a model to spot the texture of generated images and output a probability. They can be useful, but they are known to miss new generators, flag real photos that are heavily compressed, filtered or upscaled, and give different answers for the same picture at different sizes. A number like 87% looks precise but has no fixed meaning across images. Because this page can't measure its own error rate on your picture, it reports only evidence it can point to — and tells you plainly when there isn't any.
The result says “No evidence either way” — what now?
- Ask for the original file, or download it from where it was first posted, and check again.
- Run a reverse image search (Google Lens, Bing, TinEye) to find the earliest copy and its source.
- If you think it came from Google's tools, upload it to the Gemini app and ask whether it was made with Google AI — Gemini can check its own SynthID watermark.
- Look closely: hands, text, reflections, earrings, background lettering and repeated patterns are where generated images still slip.
- Weigh the context — who posted it, when, and whether anyone else captured the same scene.
This checker vs typical online AI detectors
| Feature | This page | Typical AI image detectors |
|---|---|---|
| Uploads your image | No — runs in your browser | Yes, to their servers |
| Method | Reads labels written by the AI tool | Pixel classifier guess |
| Shows where the evidence is in the file | Yes, line by line | Usually just a percentage |
| Names the tool and shows a saved prompt | Yes, when present | Rare |
| Says when it can't tell | Yes — never reports “real” | Always outputs a score |
| Batch, account, limits | Batch OK, no account, no limit | Often capped or behind sign-up |
Frequently asked questions
How can I tell if an image is AI-generated?
Check the original file for an AI label first — open it on this page. If it carries Content Credentials, an AI source-type tag or a saved prompt, it was made by AI. If not, use reverse image search and look for visual errors; no tool can prove an unlabelled image is real.
Is this AI image detector accurate?
When it finds a label, yes — the label was written by the generator. When it finds nothing, it can't tell, because labels are easily lost. It never guesses a percentage.
Can it detect ChatGPT images?
Yes, original ChatGPT downloads carry Content Credentials signed by OpenAI, and the page shows “made by ChatGPT”. A screenshot or re-saved copy loses that label.
Can it detect Midjourney, Stable Diffusion and Gemini images?
Stable Diffusion and ComfyUI PNGs usually keep the prompt, which the page shows. Midjourney downloads keep a job description. Gemini images are recognised by the visible sparkle; the invisible SynthID needs Google's own checker.
Does a screenshot remove the AI label?
Yes. A screenshot is a new image made by your phone or computer, so none of the original's metadata comes along. The same happens with most social apps and many re-saves.
Are my images uploaded?
No. The file is read in your browser; nothing is sent to a server and there's no account. It even works offline once the page has loaded.
Does it check SynthID?
No. SynthID is an invisible watermark that only Google can read. Upload the image to the Gemini app and ask if it was made with Google AI.
Can a real photo be flagged as AI?
Only if its file carries an AI label — for example a real photo edited with generative fill, which is reported as “Edited with AI”. The page doesn't judge pixels, so a photo is never flagged just for looking smooth or sharp.
Updated · By the imagetojpg.app team