How GPTypo verification works
A verified scan means a fine-tuned neural model evaluated the writing and returned a probability. It gives you something to show when another detector flags your writing.
Every scan returns a score from 0 to 100
100 means the model is highly confident the writing is human. 0 means highly confident it is AI-generated. Most real writing lands somewhere on the spectrum, which is why we group results into three zones.
Verified vs estimate, and why the difference matters
GPTypo runs two scoring systems with different reliability. Only one carries the ✦ Verified badge. The numbers below come from automated testing against 4,133 labeled samples.
Neural verification
A fine-tuned classifier trained on a curated corpus of verified human and AI-generated text spanning 11 AI models, 8 domains, and 12 adversarial attack types. Independently evaluated on the RAID benchmark at 98.2%.
Local heuristic
A 5-signal scorer that runs in the extension itself. Useful for a quick read before submitting work, but the false positive rate is more than ten times higher than verified scans.
For shareable proof reports, only verified scans carry the badge. Estimate scans can also be shared, but the absence of the badge is itself a signal to the viewer.
Why verification exists at all
Across all major AI detectors, formal human writing is misclassified at high rates. Heuristic scoring approaches similar to what most free detectors use will misclassify human writing nearly a third of the time on a labeled corpus.
When a human writer gets caught in that error rate, they need something to point at. Verification is built to be that something. A timestamped, link-shareable scan from a separate model the accuser does not control.
What verification is not
Worth being clear about the limits, because they affect when and how a verified scan is useful.