How Instructors Regulate AI in College: Evidence from 31,000 Course Syllabi

How instructors regulate AI in college — a large-scale longitudinal study by Igor Chirikov (2026) analyzing 31,000+ course syllabi (2021–2025) from the full course universe of a large public research university. Building a task-based framework (adapted from labor economics) of how AI reshapes skill formation via displacement, augmentation, and reinstatement, the paper finds instructors are warming toward AI: Regulation grew to 55% of courses by Fall 2025, but shifted from restrictive toward permissive, differentiated by task type, and increasingly framed around learning rather than academic integrity — with substantial disciplinary variation.

Overview

Generative AI substitution for the cognitively demanding tasks students must practice raises a core tension: if AI substitutes for practice, students may fail to develop skills; if AI is restricted entirely, students miss tools expected in the labor market. Chirikov adapts task-based models of technological change (Autor et al. 2003; Acemoglu & Restrepo 2019) from production to education, where students develop skills by practicing tasks. This yields three mechanisms — task displacement (AI performs tasks instead of students → skill erosion), task augmentation (AI supports practice without displacing essential effort → enhancement), and task reinstatement (AI enables new tasks → new AI-based skills) — and predicts how instructors respond: restricting AI for displaced tasks, permitting it for augmented ones, and introducing new AI-based tasks.

Method

Data: Over 31,000 course syllabi representing the full universe of courses at a large public research university in Texas, 2021–2025, tracking the same courses and instructors over time.
Design: Longitudinal; computational extraction of required course tasks and AI policies from each syllabus, with LLM-assisted classification validated at 96% agreement with human coding.
Analysis: Within-course-instructor fixed-effects estimates plus repeated cross-sections; placebo tests (e.g., oral presentation tasks where AI capabilities are weaker).

Key findings

Regulation grew rapidly but shifted toward permissiveness. Explicit AI regulation rose from near zero before ChatGPT to 55% of courses by Fall 2025. Within-course-instructor estimates show regulation increasing 15.3 pp/year, while the share of fully restrictive policies fell ~5 pp/year and a permissiveness index rose 0.12 points/year — individual instructors moved toward more permissive policies over time.
Task composition predicts regulation. Each additional writing/coding task required in 2021–22 (pre-ChatGPT) is associated with a ~3 pp higher likelihood of an AI policy by Fall 2025; a course requiring five such tasks is ~15 pp more likely to regulate. A placebo using oral-presentation tasks finds no effect.
Instructors differentiate by task type. Policies increasingly permit AI for some tasks and restrict it for others (task differentiation +4.5 pp/year). In Fall 2025, instructors most commonly restrict drafting/revising (79%) and reasoning/Problem Solving (65%), and most commonly permit editing/proofreading (83%), study support/synthesis (80%), and coding (75%). Ideation/planning is most contested (46% permit / 54% restrict).
New AI-based tasks are emerging modestly. Adoption of AI-integrated assignments, prompting exercises, and output-verification activities reached 11% of courses by Fall 2025 (Business leads at 27%; Humanities lowest at 5%).
Framing shifted from integrity to learning. Academic-integrity mentions fell from 63% (Spring 2023) to 49% (Fall 2025); references to AI's impact on learning rose from 1% to 29%; attribution requirements rose from 16% to 43%.
Disciplinary variation is pronounced. Humanities remained most restrictive; Business moved most rapidly toward permissive policies and new AI-based tasks. Larger pre-AI task bundles predict differentiated policies (each additional task → +1.3 pp).

Implications

For instructors: A task-level lens offers a useful framework for AI policy design — distinguishing tasks facing displacement risk (where AI should be restricted) from those offering augmentation potential (where AI can be permitted or encouraged), rather than adopting blanket rules.
For institutions: The substantial disciplinary and task variation argues for flexible policy frameworks that accommodate disciplinary differences and grant instructors autonomy within their domains, rather than one-size-fits-all mandates.
For labor markets: The framework highlights a potential feedback loop — if AI displaces skill-building tasks, students may graduate with weaker skills precisely where AI is strongest, further shifting comparative advantage toward AI and accelerating automation. Reinstating new learning tasks may be essential to maintaining human comparative advantage.
For research: Studies should measure task-level AI use rather than aggregate use — the distinction between displacement and augmentation is central to understanding AI's effects on learning, and reconciling apparently contradictory findings of skill gains vs. erosion.

Connected Concepts

Connected Articles

Chirikov AI Grade Inflation 2026 — Companion paper: AI task displacement as a mechanism of grade inflation (Chirikov 2026)
GenAI Policies Higher Ed Computing — How universities develop and communicate AI policies
Adarkwah GenAI Unesco Policy 2026 — UNESCO guidance on GenAI in education policy
Crompton Governing GenAI Higher Ed Delphi 2026 — Global Delphi on governing generative AI in higher education
Zuo Instructor Power GenAI Writing 2026 — Instructors' perceived power relations with GenAI in writing
Stromberg Generative AI Learning Penalty Secondary 2026 — Learning losses from homework outsourcing in secondary education

Citation

Chirikov, I. (2026). How Instructors Regulate AI in College: Evidence from 31,000 Course Syllabi. CSHE Higher Education Working Paper Series, 26(1).

Source History Edit Chat