How to read and evaluate AI detector scores
A detector score is one input to a review, not definitive proof.
Updated October 7, 2026
What the score means
The score is Sapling's estimate of how likely the checked text is to be AI-generated. It is not the share of words written by AI. Sentence highlights are scored separately, so they won't always agree with the overall score. Formulaic human writing can be flagged, and edited AI writing can be missed.
Check how much text was scored. Free checks cover the first 2,000 characters, so on a longer document the score says nothing about the rest. For PDFs and Word files, look over the extracted text too: extraction can change the reading order.
Evaluate the detector on your own documents
- Decide what a flag leads to. For example, it might lead to a closer human review or a request for more context. Don't base a consequential decision on the score alone.
- Gather examples that match your workflow. Collect human-written and AI-generated text whose origin you know, reflecting the lengths, languages, topics, and level of editing you actually see. For generated text, note the model and the date.
- Measure both kinds of error. Count human writing that gets flagged (false positives) and AI writing that gets missed (false negatives). Break results down by length and document type, and report counts alongside rates.
- Record the setup. Note the date, detector version, threshold, extraction method, sample sizes, and any truncation. Re-run the evaluation when the model or your workflow changes.
Metrics to report
| Metric | Calculation | What it tells you |
|---|---|---|
| False-positive rate | Flagged human examples ÷ all human examples | How often human writing gets flagged |
| Detection rate (recall) | Flagged AI examples ÷ all AI examples | How much AI writing your threshold catches |
| Precision | Flagged AI examples ÷ all flagged examples | How often a flag is right, given your mix of examples |
Precision depends on what share of your examples are AI-generated. A benchmark with equal numbers of human and AI examples can overstate precision for a workflow where AI submissions are rare. Small samples make every estimate less certain.
Reviewing a flagged document
Read the flagged passages and check their sources. Look at drafts or revision history if you have them, and ask the author how they wrote it. Record the outcome of your review separately from the detector's score.
Further reading
The detector API documentation covers response fields and integration, and the general evaluation guide goes deeper on test data and metrics. For methodology and reported results, request the technical whitepaper. Before relying on its numbers, check that its dataset and model version match what you use today.