Do AI Detectors Actually Work? What Professors Should Do Instead in 2026

Every fall, some fraction of integrity cases get built on a number from an AI detector. This is the year to stop, not because misuse is not happening, the data says it clearly is, but because the detector evidence will not survive contact with an appeals panel, and there is a better way to run the whole problem. Here is what the current research actually says, and the framework that holds up.

Do AI detectors actually work?

Not reliably enough to carry an accusation, and the record is public. OpenAI shut down its own AI-text classifier in 2023 for a low rate of accuracy; it caught roughly a quarter of AI-written text. Stanford researchers found GPT detectors flagged more than 61 percent of TOEFL essays by non-native English speakers as AI-generated, essays written by real people, with at least one detector flagging nearly all of them. Turnitin claims under 1 percent false positives at the document level, and at roughly 100 million papers a year even that claim implies on the order of a million wrongly flagged documents. The Washington Post's independent testing caught the tool flagging authentic student writing. Vanderbilt disabled Turnitin's AI detector entirely, citing accuracy concerns it could not validate.

Weigh the error costs and the conclusion writes itself. A false negative means one assignment slipped through. A false positive means you formally accused an innocent student, and the Stanford data says that harm lands hardest on international students. A detector score is, at most, a reason to look closer. It is never proof.

How many students are actually using AI on coursework?

Nearly all of them. The Higher Education Policy Institute's 2026 survey found 94 percent of undergraduates using generative AI to support assessed work, up from 88 percent the year before and 53 percent two years ago. Meanwhile Tyton Partners' Time for Class 2026 study of more than 3,000 students, instructors, and administrators found only 32 percent of administrators reporting an institution-wide AI policy, and just 22 percent of faculty who consider their institution's policy effective. That gap between universal use and absent policy lands on one document: your syllabus. In most classrooms this fall, the operative AI policy is whatever the course syllabus says, which makes writing it well the highest-leverage prep hour of the summer.

Should I just ban AI in my course?

Only where you can actually enforce it, and the data suggests that is narrower than it feels. Among faculty at institutions that ban AI outright, 13 percent say the policy is effective. Where AI is integrated into assignments, that figure is 30 percent. Neither number is a triumph, but a 2.3x gap is a finding, and it matches the mechanism you would predict: rules that acknowledge reality get followed, rules that deny it get routed around. The honest framework is a four-tier choice made per course, and ideally per assignment: fully prohibited, permitted for specified tasks, permitted with disclosure, or required and assessed. The right tier follows from one question: what must students do unassisted for your course to mean anything? Protect that core with in-class and oral assessment where a ban is real, and write looser, clearer rules everywhere else.

I need to rewrite my syllabus to include AI usage. Where do I start?

Work five steps in order: decide, write, tag, teach, and name the process. If that sentence is exactly what you typed into a chatbot this week, you are in good company, and here is the checklist you were looking for:

1. Decide the protected core. Name the two or three things a passing student must be able to do unassisted. Everything in your policy follows from this fifteen-minute decision, so make it on paper, not on vibes.
2. Write the statement with all six parts. Your stance in one bolded sentence, the learning reason behind it, concrete permitted and prohibited examples from this course's real assignments, the disclosure mechanism if any, one honest sentence about what happens when something looks off, and an invitation to ask first without penalty.
3. Tag every assignment. Add a small grid to the syllabus: each major assessment, its AI rule, and a one-line reason. Students consult the grid at 11 p.m., not the prose. Per-assignment rules get followed far better than one blanket rule.
4. Add the carve-outs. Accessibility tools required by an accommodation are always permitted, and a dated right-to-update clause keeps you agile when the tools change mid-semester, never retroactively.
5. Name the process. Say in the syllabus what happens when a submission concerns you: a walk-me-through-your-work conversation, drafts and version history requested, and your institution's formal process by name. This sentence is what makes the whole policy enforceable in week eleven.

With ready-made statement templates, the whole rewrite is about a two-hour job. Without them, budget a weekend and a lot of second-guessing.

What should I do when I suspect a student used AI?

Gather first, then have a documented conversation, and never open with the detector. Before any contact: annotate the specific concerns, pull the student's earlier work, collect the process artifacts your assignment already requires, and check the citations, because fabricated sources are the most objective and least arguable evidence in this entire space. Then ask the student to walk you through how they built the piece, from blank page to final draft. Real work narrates with texture: the dead ends, the restructure at 2 a.m., the source they hated. Its absence is visible without any tool. Document the meeting, match the outcome to your syllabus exactly, and refer through your institution's process when the evidence is objective. Cases built that way get sustained. Cases built on a probability score get overturned, sometimes publicly.

Want the complete kit?

The whole system is packaged as The Syllabus AI Policy Kit: Write Your Course AI Policy Before Week One: a 43-page guide with the cited 2026 data and charts, twelve editable syllabus statements across all four tiers, the assignment redesign workbook with eleven patterns, discipline playbooks for writing, STEM, coding, seminar, and online courses, and the printable integrity conversation protocol. Fast path from download to a defensible policy: about two hours.

→ Get The Syllabus AI Policy Kit ($49)

Planning lessons rather than policy? The classroom-planning side is covered in The Teaching Context System, which includes a full higher-ed chapter.

Frequently asked questions

Can I fail a student based on a Turnitin AI score?
Acting on a detector score alone is the fastest way to lose the case and harm an innocent student. Detector output is a screening signal at most. Build on process evidence, citation checks, and a documented conversation, inside your institution's formal process.

What AI policy should I put in my syllabus?
The one that matches what your course certifies. Choose a tier per assignment, from prohibited to required, state it with concrete examples, give the reason, and name the process you will follow when something looks wrong. A policy without examples and a process is a rumor.

I teach online. Can any of this be enforced asynchronously?
Yes, by design rather than surveillance: version history on everything, staged submissions so work exists across time, short recorded explain-your-work videos, and one scheduled ten-minute conversation per student per term. That certifies more than a semester of lockdown-browser quizzes.

Are AI-written assignments detectable at all?
Not reliably by software, and light paraphrasing defeats most detectors. What is detectable is the absence of process: no version history, no drafts, sources that do not exist, and an author who cannot explain their own argument.

Does banning AI in a course ever make sense?
Yes, where the process is the certified skill, like foundational writing or early math sequences. But an enforceable ban moves assessment in-class or oral. A take-home ban with no process evidence is decoration, and students learn exactly that from it.