Are AI Detectors Accurate?
We build one, so treat this with appropriate suspicion — and then read it anyway, because it is more sceptical than the marketing you will find elsewhere. Short answer: image detectors are useful and genuinely good in favourable conditions, unreliable in unfavourable ones, and dangerous when treated as proof.
What an accuracy percentage actually means
Every detector advertises a number in the high nineties. That number is measured on a test set the vendor chose, and it tells you almost nothing about your specific image.
Accuracy depends on which generator produced the image, whether it was compressed, whether it was edited afterwards, and whether it is photorealistic or stylised. A detector scoring 98% on clean Midjourney output can drop dramatically on a screenshot of a repost of an Instagram post.
The honest way to read any advertised figure: it is a ceiling measured under ideal conditions, not a promise about the file in front of you.
Where detection genuinely works well
Photorealistic images, unedited, at full resolution. This is the strongest case, and it is a common one.
Faces and portraits. Face generators train on tightly aligned datasets and reproduce that alignment in ways that are measurable.
Files that still carry metadata. When a generator signed its own output, there is little left to argue about.
Where it degrades — and by how much
Heavy compression. Social platforms re-encode everything. Much of the noise signature that forensic methods depend on is destroyed in transit.
Screenshots. A new image of an image: metadata gone, compression history rewritten. One of the worst cases.
Stylised art. Flat colour and clean linework remove texture and noise evidence. Anime and illustration are meaningfully harder than photographs, and any detector that reports the same confidence on both is not telling you the truth.
Post-processing. Grain, filters, upscaling and manual retouching all mask generation traces — sometimes deliberately.
Partial generation. A real photo with one generated element is a different problem from a fully generated image, and needs manipulation analysis rather than generation detection.
The risk nobody advertises: false positives
Everything above is about missing a fake. The opposite error does more damage in practice.
A false positive means a real photograph — or a real artist's work — gets flagged as generated. Clean digital illustration with flat colour shares statistical properties with generated art. Heavily retouched photography loses the texture that marks it as photographic.
In the AI-text world this has already produced documented harm: students accused of cheating on the basis of a detector score. Image detection carries the same risk, and the same rule applies — a score is not grounds for an accusation. If a decision affects someone, look at which methods fired, ask for the source files, and involve a human.
What to do with an inconclusive result
A good detector tells you when it does not know. Ours reports UNCERTAIN when the methods disagree, and roughly one check in eleven lands there.
That is not the tool failing. It is the tool refusing to guess, which is the correct behaviour when the evidence is genuinely ambiguous. Treat it as \u201cno answer\u201d rather than as clearance, and fall back on provenance: who provided the file, where did it come from, can you get the original.
How to evaluate a detector yourself
Do not trust reviews, including this page. Run your own test — it takes twenty minutes and settles the question.
Collect images where you already know the answer: your own photos, your own generated images. Include compressed versions, screenshots and edited copies. Run them through whichever tools you are considering.
You will learn two things: how the tool performs on your kind of material, and whether it is honest about uncertainty. The second matters more. Our comparison page lists what to check, including where we come off worse.
Key points
- Advertised accuracy is a ceiling under ideal conditions, not a promise.
- Detection is strongest on unedited photorealistic images and faces.
- It degrades on compression, screenshots, stylised art and post-processing.
- False positives are the underrated risk — a score is never grounds for accusation.
- An inconclusive result means no answer, not a clean bill of health.
- Test any detector yourself on images where you know the truth.
Frequently asked questions
Check an image yourself
Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.
More from the blog
Try it on your own image
Free account, 3 checks included, no card required.