How Sahihly scores text
Most detectors hand you a number and ask for trust. We'd rather show our work. Here is exactly what the engine measures, how the final score is produced, and where it can be wrong.
Last reviewed: July 2026
1. Deterministic style metrics
Before any AI is involved, we compute four measurable properties of your text. Rhythm variety (burstiness) is the coefficient of variation of sentence lengths — human writing alternates short and long, machine writing is flat. Vocabulary richness is the ratio of unique words to total words. Average sentence length captures pacing. And we count machine-typical stock phrases, with a separate phrase list for Arabic and English. These numbers are reproducible: the same text always yields the same metrics.
2. A hybrid ensemble, not a single opinion
Those metrics are then passed into the language model as evidence alongside the text itself, so the model reasons with data rather than vibes. The final score blends the model's judgment (65%) with the deterministic statistical score (35%). Two independent views cross-check each other, which reduces the erratic swings you get from a single-model detector.
3. Long texts are chunked, not truncated
Passages above roughly 600 words are split at sentence boundaries and analyzed in parallel, then merged with a word-weighted average. This keeps sentence-level detail sharp on long documents instead of degrading or silently cutting your text short.
4. An honest confidence rating
Every result carries a confidence level. Under ~40 words the signal is physically weak, so confidence is low. If only one engine ran, confidence is capped at medium. And when the statistical and AI views disagree by more than 30 points, we demote confidence — because a genuine disagreement means the case is ambiguous, and you deserve to know that.
5. Where we can be wrong
Two error types matter. A false positive flags genuinely human writing as AI — most likely with very formal, highly edited, or non-native prose. A false negative misses AI text, most likely when it has been heavily rewritten by a person. No detector on the market escapes this trade-off, and any tool claiming 99% certainty is selling confidence it cannot have. That is why we present scores as guidance for revision, never as evidence for an accusation.