What a score proves, and what it does not
This tool estimates a likelihood. It does not establish a fact. The difference matters most exactly when the stakes are highest, so it is set out here in full rather than buried in the terms.
Do not use this result as evidence in a legal proceeding
A score from this tool is not a forensic examination and must not be submitted, cited or relied upon as evidence of how an image was created in litigation, arbitration, criminal proceedings, insurance determinations, immigration matters, academic misconduct hearings, employment or disciplinary decisions, or any other process where a person can be penalised.
It is produced by a statistical model with a published error rate, on a device we do not control, with no chain of custody, no examiner, no case notes and no ability to be cross-examined. Those are the things that make image analysis admissible, and this tool has none of them. If a decision with consequences turns on whether an image is genuine, engage a qualified digital forensic examiner.
What the number actually means
The 0–100 score is the model's estimated probability that an image contains generated pixels, calibrated so that the number maps consistently onto the five published bands. A score of 90 means the image resembles the synthetic examples in the training distribution far more closely than the authentic ones. It does not mean there is a 90 percent chance the image is fake, it does not identify which model produced it, and it does not identify who produced it.
The region map is the same model applied to tiles of the frame independently. A single hot tile is a useful pointer to where to look with your own eyes. It is not a boundary tracing of an edit, and it will sometimes light up on a genuine part of a photograph that happens to be smooth, blurred or heavily compressed.
Where it is known to be unreliable
Reliability drops, sometimes sharply, in all of the following cases. This is not a complete list, because a complete list is not possible for a statistical model.
- Images that have been screenshotted, re-saved by a messaging app, or passed through several rounds of recompression. Each pass destroys the fine detail the model reads.
- Very small images, heavy crops, and pictures where the subject occupies a small part of the frame.
- Photographs that have had strong computational processing applied in-camera, including modern phone night modes, portrait modes and multi-frame merges. These share texture characteristics with generated pixels.
- Illustration, 3D renders, CGI, product mock-ups, heavily retouched studio work and stock photography with aggressive noise reduction. These regularly score high without any generative model involved.
- Output from generators released after the model was trained. New architectures are the standing weakness of every pixel-based detector, including this one.
- Images deliberately processed to defeat detection. Adding noise, rescaling and re-encoding all reduce the signal the model depends on.
False positives and false negatives both occur
On the benchmark used to characterise it, the model reaches 91.3 percent balanced accuracy. That figure is from a held-out research benchmark under controlled conditions, not from images arriving off the open internet, and real-world performance is lower. Roughly one image in eleven is scored on the wrong side of the decision line in benchmark conditions alone.
Read in the direction that matters: a high score on a genuine photograph is a false accusation waiting to happen, and a low score on a generated image is false reassurance. Neither outcome is rare enough to ignore. A result is one input among several, alongside where the image came from, who supplied it, whether it appears elsewhere, and whether the content is internally consistent.
Content Credentials are the exception
When an image carries a valid C2PA manifest, that part of the result is cryptographic rather than statistical. A verified manifest is strong evidence about what the signing tool recorded. Two limits still apply: an absent manifest proves nothing at all, because most images have never had one, and a manifest describes what the signing software claimed, not necessarily what happened.
What we do not claim
- We do not claim any specific accuracy on your images.
- We do not claim to detect every generator, present or future.
- We do not claim a result is proof, evidence, certification or authentication.
- We do not claim the tool is fit for any regulated, safety-critical or evidentiary purpose.
- We do not certify, attest to, or authenticate the provenance of any image.
If you are about to act on a result
Before a result changes anything for a real person, keep the original file rather than a screenshot, export the report so the exact score and the model version are recorded, check the image against a reverse image search and its original source, and ask the person who supplied it for the unedited original. If the matter could reach a tribunal, a court or a regulator, stop and instruct a qualified examiner. A free browser tool is the beginning of an enquiry, not the end of one.
Questions
If something on this page is unclear, or you believe the tool has produced a wrong result on a specific image, write to hello@bestaiimagedetector.com. Reports of failures are genuinely useful and are read. See also the terms of service and the privacy policy.