How to use the AI image detector
- Drop the image on the box, click to choose it, or paste a copied image. Use the original file if you can: downloads from social networks and screenshots usually lose their metadata.
- The detector reads the file on your device. It checks Content Credentials, IPTC and XMP labels, embedded prompts and settings, software fields, camera EXIF and the image size. Nothing is uploaded.
- Read the verdict, then each piece of evidence under What we found. Use Copy text report to paste the findings into a message, fact-check note or ticket, or Share result for a link to the verdict and the kinds of evidence found (no image, file name, prompts or other text from the file).
How to read the verdict
- Strong evidence of AI generation: the file contains a marker that AI tools write and cameras don't. This could be a C2PA manifest declaring
trainedAlgorithmicMedia, an IPTC “Created using generative AI” label, a Stable Diffusion or ComfyUI prompt, or a software field naming a generator such as Midjourney or Adobe Firefly. If the only marker says generative AI edited the image (compositeWithTrainedAlgorithmicMedia, for example generative fill on a photo), the verdict reads Strong evidence of AI editing instead. - Some AI indicators: weaker hints only, such as a description that says “AI generated” or an exact output size of popular generators (1024 × 1024, 1792 × 1024, 1216 × 832…) with no camera data. Worth a closer look, but not proof.
- No AI markers found (inconclusive): nothing in the file points to AI. That is not a clean bill of health: most AI images shared online have had their metadata stripped.
Evidence that points the other way is shown too. Camera EXIF (make, model, exposure, lens) or Content Credentials declaring a digitalCapture are consistent with a real photo. EXIF is unsigned, though, and easy to copy onto any file, so it only supports a conclusion; it can't prove one.
Why there's no “87% AI” score
Pixel-analysis detectors guess from statistical patterns and produce confident-looking percentages that are frequently wrong in both directions, especially after an image has been resized, compressed or screenshotted. New generators also defeat them quickly. This tool only reports markers you can verify yourself, explains where each one was found, and says so plainly when the evidence runs out.
Which generators leave markers
Adobe Firefly, ChatGPT and DALL·E, Microsoft Designer and Google's Gemini image models embed C2PA Content Credentials; open the Content Credentials checker for the full signed history. Midjourney, Google and Meta write the IPTC digital source type label, and Midjourney also stores the prompt and job ID. Local Stable Diffusion tools (AUTOMATIC1111, Forge, ComfyUI, InvokeAI, Fooocus) and the NovelAI service save the prompt, seed, sampler and model in PNG text chunks or the EXIF comment. The EXIF viewer lists every one of these fields, and can also remove them.
Honest limits
- Metadata can be removed. Re-saving, screenshots, messaging apps and most social networks strip it. A missing marker proves nothing.
- Metadata can be faked. Only C2PA manifests are cryptographically signed; IPTC labels, EXIF and PNG text can be written by anyone.
- Invisible watermarks need the vendor's tool. SynthID sits in the pixels and survives cropping and compression. Google adds it to all its AI images, and OpenAI has added it to ChatGPT and API images since May 2026, but only the vendors can read it: ask the Gemini app “Was this created with Google AI?”, or use OpenAI's Verify tool for ChatGPT images. Other providers use their own watermarks and detectors.
- Edited photos are mixed. A real photo edited with generative fill can carry camera EXIF and an AI label at the same time. The detector shows both.
Tips when nothing is found
Run a reverse image search (Google Lens, TinEye or Bing) to find the earliest copy, which may still carry its metadata or credit an AI account. Look closely at text, hands, reflections, jewellery and background patterns. Check whether the source has a history of posting generated images. For news images, look for Content Credentials on the publisher's own copy.