AI or Authentic?
AI or Authentic?
Home › Methodology & limits
No AI detection is 100% certain - ours included. That is why we combine several independent checks, report results as probabilities rather than verdicts, and openly say “inconclusive” when the evidence is not sufficient. This page explains how the check works, what the results mean and where the limits are.
Every check runs through several independent stages whose results are combined into one overall finding:
These values come straight from our database and update themselves - no cherry-picking, no polished sample. They show how often we reached which verdict, not how often we were right.
21.7 %
of those "inconclusive"
Basis: 152 completed checks between June 24, 2026 and July 28, 2026.
The numbers are still small - the service is young, and some checks come from our own testing. Treat them as an order of magnitude, not as statistics.
You won't find a false-positive rate here - even though it would be the most interesting number. We could only state one from a lab test set, and such a value says little about how the model performs on freshly compressed social media material. As soon as we measure it properly on an independent test set, it will appear here - with date and dataset.
The honest answer: there is no serious one-size-fits-all percentage. Accuracy depends heavily on the material - an unaltered original can be assessed far more reliably than a repeatedly reposted, heavily compressed video. And errors are possible in both directions: real content can look artificial, AI content real.
We deliberately calibrate our check cautiously: rather than falsely branding content as AI, we prefer to classify it as “inconclusive”. A “likely” result is a reasoned probability statement, not proof.
“Inconclusive” is not a failure but a deliberate decision. When compression, length or quality leave no robust signals, any verdict would be a guess. We believe being honestly unsure beats accusing someone wrongly - in either direction.
Our assessment is not courtroom-grade proof, not a statement about intentions and not a judgement of people. It is a technically grounded probability statement about the material - meant as a basis for your own judgement before you believe or share something.
AI generators evolve weekly - so do our models. We regularly retrain with new generators and real everyday footage and continuously measure detection on fresh material. If you judge a result differently, use the vote on the result page - this feedback flows into further development.
A blanket percentage would be misleading because accuracy depends on the material: originals can be assessed far more reliably than heavily compressed reposts. That is why we report probabilities with a confidence level - and abstain when the evidence is not sufficient.
Because being honestly unsure beats guessing. Heavy compression, short clips or missing signals make a robust verdict impossible - then we say so openly instead of projecting false certainty. This avoids falsely branding real footage as AI.
No. The result is a technically grounded assessment with a probability, not courtroom-grade proof. For far-reaching decisions you should always check several sources and consult experts when in doubt.
Yes. Our models are continuously retrained to recognise new generators - the same footage can be assessed differently, and usually more accurately, with a newer model version. Better source material, such as the original instead of a screenshot, also improves the assessment.
Want to know right away whether a piece of media is AI or real?