Teaching With AI in 2026: What Instructors Are Actually Doing
The faculty development session I sat in on last fall had the air of a group that had moved past the crisis stage and into something harder: ordinary practice. No one was arguing about whether to permit AI. That argument had been lost, or won, depending on whom you asked, sometime in 2024. What the twenty or so instructors in the room wanted to know was narrower and less philosophical. How do you grade a class of eighty first-year students when most of them have used a model on the draft. How do you design a weekly discussion question that’s still worth the assignment. How do you give feedback that doesn’t get ignored because a chatbot can generate a cheerful-sounding substitute in twelve seconds. The room was tired. It was also, I thought, more honest than any panel discussion on AI in education I’d attended in the previous three years.
The practices that have emerged in the two years since most institutions moved from policing AI to integrating it are scattered, local, and more modest than the press coverage suggests. There is no AI-enabled pedagogy sweeping the discipline. There is a patchwork of small adjustments, some of which are working and many of which aren’t. Worth describing the ones that seem to be sticking.
Grading assistance is the most common use and the one that took the longest to become acceptable. A political-science professor at a midsize liberal arts college described her current workflow: she reads every final paper herself, but she uses a model trained on her rubric and five calibration samples to produce a first-pass comment sheet that she then edits. The model catches obvious structural issues: a thesis that shifts between the introduction and conclusion, evidence that doesn’t match the claim. She catches the ideas. She estimates it’s cut her grading time by about 30 percent on a stack of forty papers, without changing the grades she assigns. What matters, she said, is that the student still gets a human-written comment at the top. The model’s output is never the final word. It’s a scaffold that lets her spend her attention where it counts.
The failure mode of grading assistance is well known and worth naming. Instructors who use the model to generate final comments, not scaffolding, tend to produce feedback that reads as generic, and students notice. A calculus instructor told me, with some embarrassment, that his first semester of AI-assisted grading produced the worst student evaluations he’d had in a decade; students complained that the comments felt formulaic, which they were, because he’d stopped editing them. He now treats the model’s output as a draft he has to substantially revise. The students’ evaluations recovered. The time savings are smaller. That’s the trade.
In-class Socratic prompting has emerged as a second use that’s less discussed but more interesting. A history instructor at a community college runs her discussion sections with a projector showing a live prompt window. When the conversation stalls, a seminar dynamic that used to require the instructor to generate a new question on the fly, she types a short description of where the class has gotten to and asks a model to produce three follow-up questions, each harder than the last. She picks one and asks it. The students don’t see the model’s output directly; they see her pose a question that turns out to be well-calibrated to what the class just said. Her argument for the practice is modest. She’s not a natural improviser. The model helps her produce the kind of third question that a more experienced seminar leader would have produced naturally. The class runs better. Whether this counts as teaching with AI or teaching with a prep tool is a definitional question she’s not very interested in.
AI-assisted feedback on student writing, as opposed to grading, is the use case that most instructors I’ve spoken to have settled on as the highest value and the trickiest to get right. The shape is straightforward: a student submits a draft, the model produces comments focused on specific dimensions the instructor has flagged (argument clarity, evidence use, citation accuracy), and the student revises before submitting for human feedback. Done well, this multiplies the number of feedback rounds a student gets. Done poorly, it trains students to optimize for the model’s preferences, which are not always the instructor’s preferences. The writing programs that have navigated this best tend to be explicit with students about what the model is good at flagging (mechanics, structure) and what it’s bad at (argument strength, voice, originality). The alternative, treating the model’s feedback as if it’s equivalent to the instructor’s, produces flatter, more formulaic student writing over a semester. This pattern has been reported by writing-program directors at several institutions and rhymes with the research summarized in the overview of AI writing assistants in academic contexts.
Worth noting: the instructors who have gotten the most out of AI integration are, almost without exception, instructors who were already good teachers. A model doesn’t turn a bad lecture into a good one, and it doesn’t rescue an assignment designed to produce a specific product rather than a specific understanding. The teachers who’ve adapted well are the ones who already knew what they wanted students to learn, and who treat the model as one more instrument in the room. Teachers who were coasting (and every department has a few) are coasting harder now, with AI-generated lecture slides and AI-generated comments that no one reads. That’s a staffing problem, not a technology problem, but the technology has made it more visible.
What instructors have learned that surprised them: students often use AI badly when left alone. The naive assumption in 2023 was that students would become cheating geniuses; the reality in most classrooms is that students use models the way they used to use study guides from friends, for the bits that felt too hard, often at the worst moments in their study cycle, and frequently without understanding what they’ve produced. The instructors who’ve been most effective have incorporated explicit instruction in how to use AI well into the first two weeks of their course, drawing on the kind of pattern-level guidance we walked through in the piece on what active study looks like compared to passive review. Students who learn to use the models as tutors and critics do better than students who learn to use them as ghostwriters, and the difference is larger than the difference between strong and weak students at baseline.
The second surprise: the in-class portions of courses have become more valuable, not less. Discussion, in-class writing, oral presentations, small-group problem sets, the parts of a class that can’t be outsourced to a model, have quietly moved from supplementary to central in many syllabi. A chemistry instructor I spoke to has redesigned her course so that roughly 60 percent of the grade comes from in-class work, up from about 30 percent two years ago. Her lectures are shorter. Her problem sessions are longer. The students complain more, because in-class work is harder than take-home work, and they do better on the cumulative final than previous cohorts, which she credits to the simple fact that they’re now forced into more retrieval practice than the take-home model allowed.
What’s not working: wholesale AI integration as a marketing pitch. Universities that have branded themselves around AI-enabled learning have, in most cases I’ve seen, shipped glossy programs whose actual classroom practices look much like the classrooms next door. The useful work is happening in ordinary syllabi, taught by ordinary instructors, adjusting their assessments one quarter at a time. That’s the boring truth about teaching with AI in 2026, and the best indicator, I think, that the technology has become something teachers use rather than something that uses them.
Photo via Unsplash.