AI image detectors: the three kinds, what each one sees, and how far to trust the result
Short answer: An AI image detector is one of three different tools. A classifier looks at the pixels and guesses. A watermark check looks for a hidden mark that one specific generator put there. A metadata reader looks for a record the generator wrote into the file. Only the last two can give a firm yes, and none of them can prove an image is real.
Search for a detector and you get a page of upload boxes, each promising a percentage. What those pages rarely say is which of the three kinds you are using, what a "no" actually means, and that the tools often disagree with each other on the same photo. That is what this guide covers.
The three kinds of detector
| Kind | What it looks at | A "yes" means | A "no" means |
|---|---|---|---|
| Pixel classifier | Statistical patterns in the pixels | The model thinks it looks generated | The model thinks it looks real. A guess either way. |
| Watermark check | An invisible mark embedded by a participating generator | That generator made or edited it | That generator's mark wasn't found. Any other generator is still possible. |
| Metadata / provenance reader | C2PA records, XMP and EXIF tags, PNG text chunks | The file says a generative tool was involved | The file carries no such record. It may never have had one, or it was stripped. |
Most free online tools are pixel classifiers. They are the only kind that will give an opinion on any image at all, which is also why they are the least certain.
How pixel classifiers work, and why they get it wrong
A classifier is a machine learning model trained on a large set of images labelled "real" and "generated". It learns the statistical traces that separate the two and outputs a score. The score is the model's confidence, not a measured probability that the image is fake.
Three things make those scores unreliable:
- New generators. A classifier only knows the generators it was trained on. A 2023 research paper by Ojha, Li and Lee found that the standard approach, training a network to separate real from fake, "fails to detect fake images from newer breeds of generative models" when it was trained on older GAN images. Every new image model reopens that gap until the detector is retrained.
- Ordinary handling. Resizing, recompression and screenshots change the fine pixel statistics a classifier relies on. Real photos that have been through a few apps can start to look "synthetic" to the model.
- Thresholds. Someone decides where "AI" begins, often at 50%. A photo scoring 51% gets the same label as one at 99%.
Real-world tests show the effect. In a May 2026 audit, NewsGuard ran 15 authentic news photos from credible outlets through five popular detection tools, counting a score of 50% or more as "AI-generated". Two tools got all 15 right. The worst called 6 of the 15 real photos AI-made. Across the five tools, 13% of verdicts on authentic photos were wrong. When NewsGuard added lightly and heavily edited versions of the same photos (45 images in all), at least one tool disagreed with the others on 35 of them.
Academic work points the same way. A 2024 study comparing an automated detector with expert human artists found they failed in different ways, and the authors concluded that a combined team of human and automated detectors gave the best mix of accuracy and robustness. Neither was enough alone.
Watermark detectors
Some generators hide a mark in the pixels themselves. The best-known is Google's SynthID, and Google's own check is in the Gemini app: upload an image and ask whether it was made with Google AI.
Google's help page is clear about the scope:
- A detected watermark means all or part of the image was created or edited by Google's AI models.
- No watermark means it wasn't made or edited by Google AI, but, in Google's words, "it could have been created by other AI systems."
- Results can be inconclusive. The page lists reasons such as "Not enough details to watermark" or "Likely too small an edit".
- Uploads are one file at a time, 100 MB or less, with a usage quota per rolling 24 hours. For screenshots, Google asks you to crop tightly around the image.
Google says the watermark usually survives resizing, recolouring and compression, which makes it far sturdier than metadata. The trade-off is coverage: a watermark check only answers for the generators that use that watermark. What a Gemini image carries besides SynthID is in our Gemini image metadata post.
Metadata and provenance readers
Many generators also write a record into the file. That can be a signed C2PA manifest (Content Credentials), an XMP or EXIF field naming the software, or a PNG text chunk holding the full prompt and settings. When one of these is present it is the most specific evidence you will get: it can name the tool, and a C2PA manifest is cryptographically signed.
The AI info tag checker reads these in your browser. It flags AI tool names in the Software, CreatorTool and Make fields, PNG parameters and prompt chunks, and C2PA markers, and nothing is uploaded. That matters when the image belongs to someone else and you'd rather not hand it to a third-party service just to check it. A step-by-step walkthrough of reading each signal is in how to check if an image is AI-generated.
The weakness is the mirror image of the watermark's strength: metadata sits next to the pixels, not in them. Social platforms, screenshots and re-saving in an editor all drop it, as covered in which platforms strip metadata.
Which order to use them in
- Read the metadata first. It is free, instant and, when present, specific. A C2PA manifest or a prompt chunk usually settles the question.
- Run the watermark check that matches your suspicion. If you think it came from Gemini, ask Gemini. A negative only rules out that one generator.
- Use a classifier last, and more than one. If two classifiers disagree, you have learned that the image is hard to call, not which one is right.
- Ask about provenance. Where did the file come from, and can the person who posted it show the original or the shots taken around it? For anything that matters, that question decides more than any score.
Limits: what no AI image detector can tell you
- That an image is real. Classifiers miss generated images, watermarks only cover participating generators, and stripped metadata looks the same as metadata that never existed. Every "no" is provisional.
- That a person is lying. A false positive on a real photo is common enough that a detector score should never be the basis for accusing anyone, failing a student or rejecting an entry on its own.
- How much of an image is AI. A photo with one generatively filled patch and an image generated from scratch can get the same label from a classifier, and a C2PA record on an edited photo can say "AI" about a minor change.
- What our checker can't do. The AI info tag checker reads metadata only. It doesn't analyse pixels, it can't see SynthID or any other watermark, and its C2PA check detects presence only: it does not verify the manifest's signature. For signature validation, use a validator such as
c2patool, explained in the C2PA standard post.
Questions people ask
Do AI image detectors work? Partly. Metadata and watermark checks are reliable when they find something, and silent otherwise. Pixel classifiers catch many generated images but also flag real ones, and the independent tests above show the error rate varies a lot from tool to tool.
Which detector is the most accurate? It depends on the generator and the image, and rankings change as new models come out. A tool that scored well in one test can do badly on the next generation of images. Treat any published ranking as a snapshot.
Is there a free one? Many classifiers have free tiers. Google's SynthID check in the Gemini app needs you to be signed in and has a usage quota. The AI info tag checker and the image metadata viewer are free and run in your browser.
Can a detector tell if a photo was edited with AI? Sometimes. A C2PA manifest may list generative edits, and Google's check can flag parts edited with its tools. Classifiers are inconsistent on lightly edited photos, as the NewsGuard test showed.
The short version
"AI image detector" covers three different tools. Metadata readers and watermark checks give firm answers when they find something and say nothing when they don't. Pixel classifiers always give an answer, and are wrong often enough on real photos that no score should be treated as a verdict. Check metadata first, the matching watermark second, classifiers last, and when it matters, ask where the image came from.