AI Image Detector: Is This Image AI?
Is this image AI or a real photo? Upload a picture and we check it three ways: provenance metadata (C2PA, IPTC, generation parameters), an ML model trained on recent generators (it runs in your browser) and a checklist for your own eyes. You get a probability and the signals found, not a verdict.
ML model (in your browser)
Provenance metadata
Check with your own eyes: a checklist
These are hints, not an algorithm: recent generators make fewer mistakes, and real photos can look odd too. Look for several signs at once.
- Hands and fingers — extra or fused fingers, odd joints, objects melting into the palm.
- Text — signs, labels and plates with non-existent letters or words, mixed fonts in one line.
- Reflections and shadows — a reflection in a mirror or glasses does not match the scene, shadows point in different directions.
- Background — mushy blobs instead of people and cars, lines that break off, repeating windows and tiles.
- Skin, hair, small details — plastic-looking skin without pores, hair blending into the background, mismatched earrings, asymmetric glasses.
- Context — run a reverse image search: where did it first appear, and are there other shots of the same event?
AI Image Detector: How to Tell if an Image Is AI-Generated
The tool combines two independent ways to check whether an image is AI-generated. The first is provenance metadata: a C2PA (Content Credentials) manifest with its signer and source type, the IPTC DigitalSourceType field, XMP and EXIF, plus PNG text chunks where Stable Diffusion WebUI and ComfyUI store generation parameters. When a generator leaves such a record, it is the strongest signal. The second is the Community Forensics ML classifier, fine-tuned on images from GPT Image, DALL·E 3, FLUX, Imagen, Seedream, Midjourney and other generators: it scores the pixels themselves and works even when metadata is gone. The model runs in your browser, so the image is not uploaded for this step. You get a probability and a list of signals, not proof.
How to Use
- Upload the image — ideally the original, not a screenshot or a copy from a messenger
- Click “Check the image”: metadata is checked right away, and the ML model loads once on a button click and then comes from the cache
- Weigh the summary: the metadata record, the model score and your own look using the checklist
FAQ
Is this image AI? A step-by-step check
The most reliable way to tell whether an image is AI-generated is not one detector but several independent signals. This AI image detector gives you two of them — metadata and an ML model score — and the third is up to you: the source and context of the picture. A practical order:
- Find the original. A reverse image search (Google Lens, TinEye) often leads to the first publication, with the author’s caption, a higher resolution and sometimes metadata that got lost in reposts.
- Check the metadata. If the generator wrote a C2PA manifest or an IPTC label, that is the weightiest evidence. For a detailed manifest breakdown use the C2PA Checker.
- Run the ML model. It looks at the pixels and works when metadata is missing.
- Look for yourself using the checklist on this page: hands, lettering, reflections, background.
- Weigh the context: who posted it, whether other shots of the same event exist, whether weather, signs and clothing fit the place and date.
No single step gives a yes-or-no answer. Confidence comes when several signals point the same way.
What an AI generated image checker actually sees
Metadata: a record of origin
Some generators label their files. The C2PA standard (Content Credentials) stores a signed manifest inside the file: who created the image and what was done to it. For generative AI the manifest uses the source type trainedAlgorithmicMedia, and for composites with AI parts compositeWithTrainedAlgorithmicMedia. The same IPTC vocabulary is also used without C2PA, in the XMP field DigitalSourceType. Local generators write their own records: Stable Diffusion WebUI saves the prompt and settings in a PNG text chunk called parameters, and ComfyUI uses prompt and workflow chunks.
Keep two properties of metadata in mind. Its presence is a strong signal, because such a record does not appear by accident. Its absence means nothing: the blocks disappear on re-save, and anyone can strip them, for example with an EXIF remover. We show what the manifest claims but do not verify the signature cryptographically; the official service for that is contentcredentials.org/verify.
Pixels: what the classifier learns
An ML detector does not “understand” the scene. It is trained on hundreds of thousands of real and generated pictures and picks up statistical traces of generation: characteristic noise, regularities left by upsampling, texture quirks. The model on this page classifies the whole frame: the image is resized to 440 px on the short side, a 384×384 square is taken from the center, and the output is a single number that we turn into a probability. The “likely AI” threshold is 65%, the value the model’s author chose during calibration.
Invisible watermarks
A separate class of signals lives in the pixels themselves. The best known is Google’s SynthID: it is embedded in images from Google models and designed to survive compression and cropping. Only Google’s detector can read it, so this page does not detect it. How to check SynthID in Gemini is covered in a separate guide.
Where AI image detectors get it wrong
Knowing the weak spots is more useful than remembering accuracy percentages:
- Heavily processed real photos. HDR, phone night modes, AI denoising, upscalers and beauty filters leave traces that resemble generation. A high score on such a shot is a common false alarm.
- Partial edits. If AI repainted the sky or removed a passer-by and the rest is a real photo, a whole-frame classifier can miss the insert.
- Small and recompressed images. Thumbnails, avatars and messenger copies lose the fine detail the model relies on. For images under 256 px on the short side the tool shows a separate warning.
- Screenshots and photos of screens. They add frames, UI, moiré and rescaling, which blurs the signal.
- New generators. The model knows what it was trained on. A generator released after training may produce images it misses.
- Drawings, 3D and CGI. Digital art made by a person is not a photograph, and the model may score it unpredictably.
Why different detectors disagree
Check the same picture in three services and you can easily get 12%, 58% and 97%. That does not necessarily mean one of them is broken. Each model is trained on its own dataset: one has seen plenty of Midjourney and little FLUX, another the reverse. Each uses its own threshold: somewhere “AI” starts at 50%, here at 65%. And one service’s percentage is not another’s: it is the model’s internal scale, not a shared unit.
So compare conclusions rather than numbers, and only on the same file. Even re-saving a JPEG at a different quality or trimming the edges can shift a score by tens of percentage points, especially near the threshold. If the result changes depending on which copy you test, that is information in itself: the image sits where models are unsure, and the call should rest on metadata, the source and a manual look.
How to read the result
| Metadata | ML model | What it usually means |
|---|---|---|
| AI signals found (C2PA, IPTC, generation parameters) | Any score | The file is labeled by a generator. A low model score does not cancel that: the file record is the stronger signal |
| None, or no AI signals | 65% or higher | The model sees generation traces. Look for the original and use the checklist, especially if the photo is heavily processed |
| None, or no AI signals | 35–65% | Borderline: the model is unsure. Context and a manual look decide |
| None, or no AI signals | Below 35% | No clear signs. Not a guarantee: partial edits and recompressed AI images slip through |
| Camera or editor C2PA manifest with no AI actions | Low | A good sign of authenticity, but validate the signature at contentcredentials.org/verify |
When in doubt, check the best copy you can get: a full-resolution file from the original source almost always gives a more meaningful result than an image saved from a feed.
Practical checklist before you conclude
- Does the image have an author, date and place you can verify?
- Can you find an earlier publication or a higher-resolution version?
- What is in the metadata: C2PA,
DigitalSourceType, generation parameters, camera EXIF? Remember that EXIF is easy to fake. - What score does the model give on the original rather than on a screenshot?
- Do the small things add up: fingers, text on signs, reflections, shadows, matching earrings?
- Was the shot processed in a way that explains a high score (HDR, retouching, upscaling)?
For a full file audit — hidden labels, LSB anomalies and metadata cleanup — use the image watermark detector.
RankBotAI — AI SEO Optimization for WordPress
RankBotAI AI plugin automatically finds and fixes SEO issues in WordPress products and posts — no manual editing.