What a detector analyzes
AI text detectors generally evaluate statistical and linguistic patterns across a passage. Depending on the model, those patterns can include predictability, variation in sentence structure, repetition, word choice, transitions, and relationships between neighboring tokens. The detector compares the observed pattern with what it learned during training.
What the score means
A percentage is best read as the model's confidence or likelihood signal under its own calibration. It is not the percentage of words written by AI, and it is not a probability that can be transferred unchanged between different detectors.
Why multiple detectors disagree
Detectors differ in training data, supported languages, minimum text length, model architecture, thresholds, and sensitivity to edited or mixed-authorship text. A short passage may not contain enough signal. A translated, formulaic, highly technical, or heavily edited passage can also fall outside the patterns a detector handles well.
Common false-positive conditions
- Short text with little stylistic variation.
- Formal templates, standardized answers, or highly constrained writing.
- Writing by language learners or text translated between languages.
- Technical definitions, summaries, and repetitive instructional language.
- Heavy grammar correction or rewriting that normalizes sentence structure.
A responsible review workflow
- Use enough continuous text for the detector to evaluate.
- Compare more than one signal instead of relying on one score.
- Read the passage and identify the actual patterns that raised concern.
- Check drafts, notes, sources, revision history, and assignment context.
- Allow the writer to explain the work before making a high-stakes judgment.
Humanization and revision
Rewriting can change rhythm, specificity, vocabulary, and sentence variation. That may change detector scores, but it does not establish authorship or guarantee that another detector will respond the same way. Revision should improve clarity and accuracy, not chase an unsupported promise of “undetectable” text.
Questions
Can an AI detector prove who wrote a document?
No. It can identify statistical patterns associated with different kinds of text, but it cannot establish authorship by itself.
Why do detectors disagree?
Models use different training data, features, thresholds, text-length requirements, and calibration methods. The same passage can therefore receive different scores.
Does humanizing text prove it was human-written?
No. Revision can change detectable patterns, but the final wording does not prove who produced the original ideas or draft.