What a score settles, and what it leaves open
This tool estimates a likelihood. It does not establish a fact. That distinction matters least when nothing is riding on the answer and most when everything is, which is why it is set out here in full instead of being folded into the terms.
Never put this result in front of a tribunal
A score from this tool is not a forensic examination. It must not be submitted, cited or relied on as evidence of how a picture was made — not in litigation, arbitration or criminal proceedings, not in insurance determinations, not in immigration matters, not in academic misconduct hearings, not in employment or disciplinary decisions, and not in any other process where a person can be penalised.
What produced it is a statistical model with a published error rate, running on a device we do not control, with no chain of custody, no examiner, no case notes and nobody who could be questioned about it. Those absent things are precisely what makes image analysis admissible anywhere. If a consequential decision turns on whether a picture is genuine, instruct a qualified digital forensic examiner instead.
What the number is, precisely
The 0–100 score is the model's estimated probability that a picture contains generated pixels, calibrated so the figure maps consistently onto the five published bands. A 90 means the picture resembles the synthetic examples in the training distribution far more closely than it resembles the authentic ones. It does not mean there is a ninety percent chance the picture is fake. It does not identify which model made it. It does not identify who did.
The tile map is the same model applied independently to parts of the frame. A single hot tile is a genuinely useful pointer to where to look with your own eyes. It is not a traced boundary around an edit, and it will sometimes light up on a real part of a real photograph that happens to be smooth, out of focus, or heavily compressed.
Where it is known to be unreliable
Reliability drops — sometimes steeply — in every case below. This is not an exhaustive list, because an exhaustive list is not something a statistical model can be given.
- Pictures that have been screenshotted, re-saved by a messaging app, or pushed through several rounds of recompression. Every pass destroys more of the fine detail the model depends on.
- Very small pictures, hard crops, and frames where the subject occupies only a corner.
- Photographs with heavy in-camera computation applied — modern phone night modes, portrait modes and multi-frame merges — because 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 routinely score high with no generative model anywhere near them.
- Output from generators released after the model was trained. New architectures are the standing weakness of every pixel-based detector in existence, this one included.
- Pictures deliberately processed to defeat detection. Added grain, rescaling and re-encoding all reduce the signal, and somebody who knows that can exploit it.
Both kinds of error happen
On the benchmark used to characterise it, the model reaches 91.3 percent balanced accuracy. That figure comes from a held-out research benchmark under controlled conditions rather than from pictures arriving off the open internet, and real-world performance is lower. Even in benchmark conditions, roughly one picture in eleven lands on the wrong side of the decision line.
Read that in the direction that matters. A high score on a genuine photograph is a false accusation waiting for somebody to act on it. A low score on a generated picture is false reassurance. Neither is rare enough to wave away. A result belongs alongside the other inputs — where the picture came from, who supplied it, whether it appears elsewhere, and whether what it shows is internally consistent.
Content Credentials are the one exception
When a picture carries a valid C2PA manifest, that part of the result is cryptographic rather than statistical, and a verified manifest is strong evidence about what the signing tool recorded. Two limits still hold. An absent manifest proves nothing whatsoever, because the overwhelming majority of pictures have never carried one. And a manifest records what the signing software claimed, which is not always the same as what happened.
Claims we do not make
- No claim of any particular accuracy on your pictures.
- No claim to detect every generator, current or future.
- No claim that a result constitutes proof, evidence, certification or authentication.
- No claim of fitness for any regulated, safety-critical or evidentiary purpose.
- No certification, attestation or authentication of the provenance of any picture.
If you are about to act on one
Before a result changes anything for a real person: keep the original file rather than a screenshot of it, export the report so the exact score and model version are on the record, run the picture through a reverse image search and trace it to its earliest appearance, and ask whoever supplied it for the unedited original. If the matter could reach a court, a tribunal or a regulator, stop and instruct a qualified examiner. A free browser tool is where an enquiry starts, not where it finishes.
Questions, and wrong results
If anything here is unclear, or you believe the tool has got a specific picture wrong, write to hello@originalorai.com. Reports of failures are read and they are genuinely useful — they are the main way the evaluation set grows. See also the terms of service and the privacy policy.