Most institutional AI conversations are still stuck on detection. How do we tell if a student used ChatGPT. That question is a dead end. Detection tools do not work reliably. They produce false positives on neurodivergent students and non-native English speakers at rates that make them indefensible in an academic integrity hearing.
The campuses pulling ahead have moved past detection entirely. They are working on the next layer: what does it mean to teach, assess, and credential learning when a competent free tool can produce a B+ on most written assignments in thirty seconds.
This is what that next layer looks like in practice.
The four positions every syllabus should articulate
We work with course designers in a dozen disciplines. Across all of them, faculty are converging on four canonical positions. A good syllabus picks one — explicitly — and tells the student which.
Position 1 — AI-prohibited
The course assesses a skill that must be demonstrated unaided. First-semester writing courses. Foundational quantitative reasoning. Clinical-skill demonstrations. The position is defensible; the enforcement is in the assessment design, not the detector. Oral exams, in-class drafts, scaffolded assignments where intermediate artifacts are graded.
Position 2 — AI-permitted with attribution
The default for most upper-division coursework. Students can use AI tools the way they would use a calculator, a thesaurus, a research librarian. They must cite the tool, the model version, and the prompt when it materially shaped the output. The norm: same as you would cite a research assistant.
Position 3 — AI-required
The course assesses judgment over AI output, not unaided production. Examples: a senior capstone where students prompt the model, evaluate the result, identify failure modes, and publish a critique. Graduate-level case analysis where the model produces three plausible strategies and the student defends or rejects each one.
Position 4 — AI-as-tutor
The model is the student's study partner. Faculty deploy an institutionally-procured chat interface, pre-loaded with course materials and constrained to Socratic prompting. The model cannot give the answer — it can only ask the next question. We have seen this position lift completion rates in introductory STEM by 8–14 percentage points.
The hard work is not picking a position. The hard work is being honest with students about which one your course is.
What good faculty development looks like
The one-time workshop does not work. The two-hour summer seminar does not work. What does work is a three-semester arc, with a faculty community of practice underneath it.
The 100 / 40 / 15 curve is the most consistent shape we see across institutions of different sizes. The mistake is trying to push everyone to the 15% redesign tier. The 100% awareness tier is the one that breaks resistance.
The assessment redesign question
If a model can produce a B+ five-paragraph essay in thirty seconds, the five-paragraph essay is no longer a useful assessment of college-level writing. That is uncomfortable; it is also true.
The redesigns that hold up:
- In-process artifacts graded. Three rounds of revision, each graded, where the student must defend the changes between drafts.
- Oral defenses on written work. A six-minute conversation about the student's own essay catches every form of academic dishonesty, AI-assisted or otherwise.
- Authentic audiences. Writing for a real publication, a real client, a real grant proposal. The audience doesn't accept "I prompted a model" as a deliverable.
- Process portfolios. The student submits the conversation log alongside the artifact. Faculty grade the thinking, not just the output.
None of these are new pedagogies. They were always better assessments. AI has made them mandatory.
What students are telling us
We sat down with student governments at six institutions this semester. The signal is consistent.
- They want clarity. The ambiguity across courses is the worst part. "Professor A says don't use it, Professor B requires it, Professor C says don't ask."
- They want training. They want the institution to teach them what good use looks like. Almost none have received it.
- They are anxious about employability. They have heard from every recruiter that AI fluency is now table stakes. They want the institution to help them build the muscle, not just police it.
The institutions that meet that demand — clarity, training, employability — are the ones building trust. The ones that don't are accelerating the gap between what students experience on campus and what they need in the workforce.
Bottom line
The classroom AI conversation is past the detection question. The next decade is about pedagogy redesign at the course level, with institutional support for the faculty who do the redesign work.
You don't need every faculty member to be an AI expert. You need a clear syllabus position in every course, a credible community of practice, and an honest stance with students about what the institution stands for.
The campuses that get there first will look noticeably different to applicants by the 2027–28 cycle.