What recent research says

Contractor & Reyes (2026)

Zara Contractor & Germán Reyes. Experimental evidence on the learning impact of generative AI

Randomized experiment, IZA / arXiv working paper

Undergraduates studied an unfamiliar topic and wrote an essay with or without AI, then took unaided tests. AI access raised scores, and a smaller gain persisted a week later. Gains held for students who used AI as a tutor and faded for those who had AI write for them.

Sample and method: Randomized experiment with 211 undergraduates at one U.S. college; tests immediately and one week later.

Outcome measured: Unaided test scores immediately and one week later

Limits: One institution and one timed task. Not yet peer reviewed.

Teaching implication (author’s interpretation): Design AI use as tutoring and feedback rather than ghostwriting when retention matters.

Last verified 7 November 2026

Chen (2026)

Xinran Chen. Layer-sensitive cognitive offloading in generative AI-assisted writing

Frontiers in Psychology, 17 · quasi-experiment, 168 undergraduates, 8 weeks

Among students writing with AI, heavier offloading of ideas and reasoning was associated with lower higher-order thinking on later work done without AI. Stronger self-regulated writing weakened, but did not remove, that link.

Sample and method: Study of students writing with AI, with independent no-AI outcomes measured at week 8.

Outcome measured: Higher-order thinking on later work done without AI

Limits: Associations, not proof of cause. One context and task type.

Teaching implication (author’s interpretation): Require some independent, no-AI work so offloading does not go unnoticed.

Last verified 7 November 2026

Nakamura (2026)

Daiki Nakamura. Augmentation or erosion? Regulating cognitive offloading in generative AI-supported STEM learning

Educational Psychology Review, 38, Article 124 · integrative review

Reviews evidence showing both gains and learning penalties from generative AI, depending on how it is configured and used, and argues offloading must be regulated rather than simply allowed or banned.

Sample and method: Integrative review of research on offloading, cognitive load and self-regulated learning in STEM.

Outcome measured: None directly (review of other studies)

Limits: Synthesizes mixed evidence. Does not report new data.

Teaching implication (author’s interpretation): Decide task by task which thinking students must do themselves.

Last verified 7 November 2026

Top 200 universities worldwide (2026)

Ullah & Kamel Boulos. Assessing generative AI tools usage guidelines at the top 200 universities worldwide

Future Internet, 18(9), 492

Surveyed the websites of the top 200 universities in the 2026 QS rankings: 180 published GenAI guidelines. Most name an issuing authority and state objectives and scope.

Sample and method: Website review of guidelines at the top 200 universities in the 2026 QS rankings, April 2026. Global.

Outcome measured: Whether guidelines exist and what they contain

Limits: Counts what guidelines contain. Does not measure how they are applied or whether they improve learning.

Teaching implication (author’s interpretation): Having a guideline is common; check whether yours is specific enough to act on.

Last verified 7 November 2026

Crompton et al. (2026)

Crompton, Burke, Nickel et al.. Governing generative AI in higher education: A global Delphi study

International Journal of Educational Technology in Higher Education

A global Delphi study with experts across six continents produced a policy framework with eight core areas, beginning with academic integrity and ethical use and including equitable access, privacy and AI literacy.

Sample and method: Delphi study: rounds of expert consensus with panelists from six continents.

Outcome measured: Expert agreement on policy areas

Limits: Expert opinion, not observed practice or student outcomes.

Teaching implication (author’s interpretation): Use the eight areas as a checklist for institutional policy.

Last verified 7 November 2026

Chirikov (2026)

Igor Chirikov. How instructors regulate AI in college: Evidence from 31,000 course syllabi

UC Berkeley CSHE working paper

Analyzed 31,000 syllabi. Instructors moved from blanket bans to task-based rules. Explicit AI rules reached about 55% of syllabi by Fall 2025, so roughly 40–45% still had none, and newer policies talk more about learning and less about integrity.

Sample and method: 31,000 course syllabi, 2021–2025.

Outcome measured: Presence and type of AI rules in syllabi

Limits: Not yet peer reviewed. Syllabus text shows stated rules, not what happens in class.

Teaching implication (author’s interpretation): Name rules per task in every syllabus; silence leaves students guessing.

Last verified 7 November 2026

Li et al. (2026)

Ming Li, Qin Xie, Ariunaa Enkhtur, Shuoyang Meng, Lilan Chen, Beverley Anne Yamamoto, Fei Cheng & Masayuki Murakami. A framework for developing university policies on generative AI governance

Studies in Higher Education · 124 policy documents from 110 universities

Compared university GenAI policies in the U.S., Japan and China. Instructors act as gatekeepers in all three, and policies are built to change as tools change.

Sample and method: Comparative analysis of university GenAI policies in the U.S., Japan and China.

Outcome measured: Policy content across three countries

Limits: Three countries. Findings may not transfer to other systems or institution types.

Teaching implication (author’s interpretation): Give instructors discretion inside clear institutional guardrails.

Last verified 7 November 2026

Wu (2026)

Evelyn Wu. How institutional guidance becomes course-level policy

Frontiers in Education, 11 · 35 syllabi at one university

Analyzed 35 syllabi at a large U.S. public research university and found institutional guidance does not become student-facing until instructors translate it into course rules.

Sample and method: Qualitative analysis of 35 syllabi at one large U.S. public research university.

Outcome measured: How guidance appears in syllabi

Limits: Single institution and small sample.

Teaching implication (author’s interpretation): Translate campus guidance into course rules students actually see.

Last verified 7 November 2026

Sun, Møller & Dohn (2026)

Sun, Møller & Dohn. Policy framing and student maneuvering

Education Sciences, 16(10), 1704 · 36 universities, 8 student interviews

Students described university GenAI policies as vague, difficult to locate and insufficiently specific to their programme or assessment.

Sample and method: Cross-national analysis of policy framing and how students respond to it.

Outcome measured: Student perceptions of policy

Limits: Reports perceptions. Does not measure learning outcomes.

Teaching implication (author’s interpretation): Make policies easy to find and specific to each assessment.

Last verified 7 November 2026

HEPI Student Generative AI Survey (2026)

Stephenson & Armstrong (HEPI). Student generative AI survey 2026

Higher Education Policy Institute

Of UK undergraduates surveyed, 94% use AI for assessed work. Students remain unsure what is acceptable, and institutional support for AI skills lags behind use.

Sample and method: Online survey of UK undergraduates, 2026.

Outcome measured: Self-reported student use and attitudes

Limits: UK only. Measures whether students use AI, not how often, why or how well.

Teaching implication (author’s interpretation): Assume most students use AI; make expectations explicit and teach AI skills.

Last verified 7 November 2026

EDUCAUSE (2026)

Jenay Robert (EDUCAUSE). The impact of AI on learning assessment

The Impact of AI on Learning Assessment

Surveyed 438 faculty and staff who design assessments. Respondents want to choose when AI is used and called for clearer policies and stronger AI literacy.

Sample and method: Survey of 438 faculty and staff who design assessments.

Outcome measured: Self-reported faculty and staff views

Limits: Self-selected respondents. Reports views, not outcomes.

Teaching implication (author’s interpretation): Involve faculty who design assessments in setting AI rules.

Last verified 7 November 2026

Gu et al. (2026)

Xingjian (Lance) Gu, Wells Lucas Santo, James M. Zumel Dumlao & Barbara Ericson. To police or to guide: How CS instructors design generative AI policies

arXiv preprint · 13 instructor interviews

Interviews with U.S. computer science instructors found most course AI policies aim to AI-proof assessments. Few address how students learn.

Sample and method: Interviews with U.S. computer science instructors.

Outcome measured: Instructor-reported policy aims

Limits: Not yet peer reviewed. Small sample in one discipline.

Teaching implication (author’s interpretation): Tie course AI rules to what students should learn, not only to preventing misuse.

Last verified 7 November 2026