How do professors check for AI? Detectors, drafts and vivas

Each method below is taken from a published university page or a detector maker's own documentation, with the date it was read. None of them is a guess about what happens behind a closed office door.

The order matters. In every policy that describes a process, the detector report is at most a starting point and the conversation is where a case is decided. Cornell's teaching centre puts it plainly: instructors should use initial impressions "as a starting point for deeper inquiry, supported by objective and verifiable evidence" (Cornell Center for Teaching Innovation, AI and academic integrity, updated for Fall 2026, read 9 November 2026).

Turnitin is licensed to institutions, not to individuals, so a student never sees the report unless an instructor downloads and shares it. Turnitin states that its model scores each prose sentence between 0 and 1, labels sentences above a threshold as likely AI, and reports the share of such sentences as the document percentage (Turnitin, AI writing detection FAQ, read 9 November 2026).

Turnitin also states that it holds its false positive rate "under 1% for documents with over 20% of AI writing" and that it accepts missing some AI text to do so, read 9 November 2026. The full mechanism, including the October 2026 model update, is on Turnitin AI detection explained. Turnitin's detector is not available to the public, so this site relies on Turnitin's published statements and peer-reviewed studies and reports no Turnitin result for any tool.

Often, yes, and several universities say that is the method they prefer. Stanford's policy guidance adopted on 15 October 2026 states that "AI detection tools are not currently used at Stanford" and that instructors may ask students to explain their work in class, sections and labs (Stanford Office of Community Standards, generative AI policy guidance, read 9 November 2026). Toronto states that it "does not support the use of AI-detection software programs on student work" because they "are known to incorrectly flag instances of AI use in human-written content", and recommends meeting the student and using short in-person assessments instead (University of Toronto, generative AI FAQs, updated October 2026, read 9 November 2026).

Cornell's teaching centre lists the evidence it considers tangible, and the list is a fair description of what gives a paper away to a reader who knows the student.

Cornell adds a standard of proof: an allegation must be "clear and convincing", which it describes as above "more likely than not" and below "beyond a reasonable doubt", and it states that "AI detection scores are not necessary as evidence of academic misconduct", read 9 November 2026. Columbia's provost policy says detection "should be treated as a guideline and not a grading metric" (Columbia University, generative AI policy, read 9 November 2026).

Two kinds of number exist: what the vendors state and what independent studies found, and they are measured on different texts. Turnitin states under 1 percent false positives above the 20 percent band. GPTZero states it keeps its false positive rate "at no more than 1%" and describes a sentence-by-sentence classifier with a "Paraphraser Shield" against paraphrasing and homoglyph attacks (GPTZero, technology, read 9 November 2026).

Weber-Wulff and colleagues tested 14 detectors on 54 documents in 2023 and found no tool above 80 percent accuracy; human-written text was classified correctly 96 percent of the time, AI text with light manual edits only 42 percent and machine-paraphrased AI text 26 percent (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Liang and colleagues found seven detectors flagged 61.22 percent of 91 human-written TOEFL essays as AI in 2023 (Liang et al. 2023, Patterns, read 9 November 2026). Both studies tested 2023 models; the vendor figures are current claims on the vendors' own test sets. The guide on whether AI detectors work sets the two side by side.

No detector score should be the only evidence in a misconduct case. GPTZero, which sells a detector to schools, tells educators to ask students to show their process through edit history or in-person assessment and to look at patterns across several assignments rather than act on one scan (GPTZero, for educators, read 9 November 2026).

A process, not a verdict. Toronto states that when an instructor believes a student used an unpermitted tool, the instructor "should meet with the student as the first step" under its Code of Behaviour on Academic Matters, read 9 November 2026. Sydney states that a Turnitin AI score "would not be the only evidence relied upon" and is "considered alongside other relevant evidence" (University of Sydney, artificial intelligence and academic integrity, last updated 3 November 2026, read 9 November 2026). UCL states that a suspected student "would be likely to be invited to an investigatory viva to probe the authorship of your work" (UCL, engaging with AI in your education and assessment, read 9 November 2026).

For a student, the practical consequence is to keep what those processes ask for: drafts, notes, version history and the sources themselves. Sydney's own advice to students who are permitted to use AI is to keep copies of the outputs as evidence of the research and writing process. What to say and bring to that meeting is the subject of what to do if you are flagged for AI, and the rest of this series sits in the guides hub.