**TL;DR** — Turnitin's AI detector examines short passages for token-probability patterns typical of large-language-model output. It flags passages with a >20% AI probability by default and reports an overall percentage to your instructor. False positives exist but are rare (<1% at the 20% threshold).
Last updated: July 2026.
What signals Turnitin uses
LLMs like GPT-4, Claude, and Gemini generate text token-by-token, always choosing high-probability continuations. Human writing has higher variance — occasional rare words, unusual constructions, inconsistent punctuation.
Turnitin's model, trained on tens of millions of AI vs human samples, computes a "perplexity signature" on 300-token sliding windows. Low variance across windows → likely AI.
What it catches well
- Unedited ChatGPT output (any GPT-4 version).
- Claude Sonnet / Opus output without paraphrasing.
- Gemini Pro output — similar signature.
- Text that is 80%+ generated even if lightly rewritten.
What it struggles with
- Text that is 30-70% generated and 30-70% rewritten in human voice.
- Paraphrased AI text run through tools like QuillBot or Undetectable.ai (still detectable at lower confidence but Turnitin's threshold may not flag).
- Very short passages (<300 words).
- Highly technical text with narrow vocabulary (medicine, law, code comments).
False-positive risk
Turnitin's published false-positive rate is under 1% at the 20% AI threshold and around 4% at 0% (any AI flag at all). Real-world false positives tend to cluster in:
- ESL students whose sentence patterns match LLM training data.
- Formulaic writing (lab reports, clinical case studies).
- Text edited heavily with Grammarly's rewrite features.
What professors actually see
A percentage bar labeled "% AI." Some institutions show only ranges (0-20%, 20-60%, 60-100%). Instructors can click into the flagged sentences.
What Turnitin does NOT see
- Which AI tool you used.
- Your ChatGPT conversation history.
- Your Google Docs revision log (that's separate).
If you're flagged
Every US institution requires additional evidence — writing history, oral defense, drafts — before pursuing academic integrity. A high AI score alone is not proof.
For genuine writing help that doesn't trigger detectors, see our [assignment help service](/services/assignment-homework-help).
**Sources**: Turnitin AI Writing Detection Technical Report (2026), Chronicle of Higher Education AI detection accuracy study (2025).