Detectors infer patterns in text. They do not observe who wrote it. PACT uses process evidence and conversation to understand learning.
In a 2023 test, Ars Technica documented detector outputs classifying US Constitution text and a passage from Genesis as AI-generated. These are examples of false positives, not a benchmark for every detector today.
Ars Technica · July 2023 ↗A 2023 Stanford study tested seven detectors on 91 human-written TOEFL essays. The average false-positive rate was about 61%. At least one detector flagged almost 98% of the essays; all seven agreed on roughly 20%.
Stanford HAI · May 2023 ↗When Vanderbilt disabled Turnitin’s detector in 2023, it illustrated the stakes: a claimed 1% false-positive rate across 75,000 papers could mean about 750 incorrect labels. This was a hypothetical calculation, not a count of observed cases.
Vanderbilt University · August 2023 ↗Turnitin’s guidance acknowledges that its model can misidentify human, AI-generated, and AI-paraphrased writing. It says the result should not be the sole basis for adverse action against a student.
Turnitin Guides · checked October 2026 ↗State the permitted level for each stage before work begins.
Read what the student says they did and how AI contributed.
Consider drafts, versions, and ordinary classroom evidence.
Explain the question and give the student a chance to respond.
Ask the student to justify choices or extend their own work.
Keep authorship and integrity decisions with people.
PACT does not claim that detectors are always wrong, that self-report proves honesty, or that an app can catch cheating. A detector score is not proof, so PACT does not use one to judge authorship.
Read the full authorship policy →