Search for an AI image detector and you get dozens of pages that all look the same: an upload box, a spinner, and a percentage. The percentages disagree with each other, none of them explain themselves, and none would survive being questioned by an adjuster, an editor or a lawyer.
Rather than rank tools that change every month, here are the criteria that actually determine whether a detector is useful when a wrong answer costs money — and how to test any tool against them in ten minutes.
1. Does it show evidence, or only a score?
A bare "87% AI" is unusable. You cannot act on it, explain it to a colleague or defend it in a dispute. A useful result names the signals: metadata present or stripped, compression anomalies and where they are, provenance manifest found or not, and what a model observed in the picture. Evidence lets you overrule a wrong score. A score alone leaves you stuck with it.
2. Does it distinguish generated from edited?
These are completely different problems with completely different consequences. A fully synthetic photo of a car crash is fabricated evidence. A real photo of a real crash with the damage enlarged is inflated evidence. A screenshot of a genuine receipt is neither — it is just a screenshot.
Any tool that collapses all three into one number will fail you on the most common real-world case, which is a genuine photo with one altered region. Look for a tool that reports separate outcomes: authentic capture, fully generated, partially manipulated, metadata tampered.
3. Does it localise the change?
When something has been edited, the only answer worth having is where. An error level analysis heatmap with boxes around the suspect regions turns a verdict into something a non-specialist can see for themselves. It is also what makes a report persuasive to someone who does not trust detectors.
4. Does it read provenance as well as pixels?
Content Credentials, invisible watermarks and embedded generator strings give definitive answers when present. A detector that ignores them throws away the strongest evidence available and guesses instead. Ask whether the tool inspects C2PA manifests at all — many do not. Our provenance explainer covers why this matters and where it stops.
5. Can you prove what you checked?
This is the criterion almost no consumer detector meets, and the one that matters most in a claim, a court or an editorial dispute. You need a SHA-256 hash of the exact file, computed on receipt; an append-only record of everything done to it; and an exportable certificate that a third party can verify independently. Without that, your result is a screenshot of a website — and the other side will say so.
6. Is it honest about uncertainty?
Treat any advertised accuracy figure with suspicion. Accuracy depends entirely on the test set, and a tool measured on pristine generator output will look superb and then fail on a twice-compressed social download. Good tools say "inconclusive, needs human review" on borderline files. Tools that are always confident are confidently wrong some of the time, and you will not know which time.
7. Does it fit into your workflow?
One file at a time in a browser is fine for curiosity. For actual work you need batch upload, a shared queue your team can review, configurable thresholds (your risk tolerance is not the vendor's default), an API, and a way to collect files from other people without emailing them around.
A ten-minute test you can run on any tool
- Photograph something with your own phone and upload the original. A tool that flags it is unusable.
- Screenshot that same photo and upload the screenshot. The right answer is "metadata stripped, no evidence of generation" — not "AI-generated".
- Open the original, paste in an object, export, upload. The tool should localise the paste, not just raise a score.
- Generate an image with any current model and upload it. Now check whether the tool names the engine or only says "AI".
- Take the same generated image, screenshot it, and upload that. This removes the metadata shortcut and tests whether the tool can actually see.
Most tools pass steps 1 and 4 and fail steps 2, 3 and 5. That pattern tells you the product is reading metadata and calling it detection.
Where VerifyAI sits
We built VerifyAI around the criteria above rather than around a single score: four independent layers (provenance, physical forensics, model review and chain of custody), separate verdicts for generated versus manipulated versus metadata-tampered, a heatmap that boxes the altered regions, engine attribution when the signature traits match, and a court-ready certificate carrying the file hash and a QR code anyone can verify.
You can run the whole thing on your own file below without an account, or start with the AI image detector.