How our analysis works

The AI-assistance score is a likelihood, not a verdict. A result of “78% likelihood” means the model sees patterns it associates with AI-assisted writing - it does not prove how the text was written. We always show an estimated range and a confidence level alongside the number.

We deliberately avoid binary “AI / Human” labels. Genuine writing varies, and so does AI output; honest reporting means showing uncertainty.

Your text is split into sections and scored by a fine-tuned language model trained on paired human and AI examples. We also compute transparent writing signals - sentence rhythm (burstiness), vocabulary diversity, repetition, and reading-level consistency - and show you exactly which patterns contributed to the estimate.

Every report includes a section-by-section breakdown so you can see where the model reacted, not just an overall number.

A false accusation is more damaging than a missed detection. We tune our thresholds to keep the false-positive rate low, and we make it easy to challenge a result. If you believe your own writing was wrongly flagged, use “Report false positive” on any report - every report is reviewed and feeds back into model improvement.

• No detector is perfect - treat results as one signal among many, not proof.
• Very short texts give less reliable estimates than longer passages.
• Heavily edited AI text, or AI trained to mimic a person, is harder to detect.
• Non-native English writing can share surface patterns with AI text; we account for this but it remains a known risk.
• Results should never be the sole basis for an academic or disciplinary decision.