A useful result starts by asking what kind of detector produced it. Generic classifiers, vendor watermark detectors, stylometric analysis and provenance metadata do not measure the same thing.
| Method | What it looks for | Best interpretation |
|---|---|---|
| AI-text classifier | Patterns learned from human and model-generated examples | How AI-like the text appears to that classifier |
| Statistical watermark detector | A deliberate token-selection signal inserted by a compatible generator | Evidence of output from a watermarking system |
| Stylometry | Measurable writing characteristics | Description or comparison of style |
| File provenance | Metadata, Content Credentials, producer information | Information about file creation or editing history |
AI Doc Scan currently uses an open RoBERTa classifier for eligible English passages. The model was trained on RAID-derived examples. Its output is a classification probability, not a vendor-specific watermark verification result.
A watermark detector is only meaningful for a compatible watermarking system. A generic classifier cannot simply “see” textGrain or SynthID in the same way the provider’s matching detector can.
One detector may flag text another considers human-like because models use different training data, thresholds and signals. Technical prose, translations, heavy editing and very short passages are common sources of disagreement.
Ask which method generated the score, what text was analyzed, how much evidence was available, what the model was trained on and what conclusion the method can legitimately support.