When 116 of the highest-research universities in the United States put out guidance on generative AI, you’d expect a certain level of rigor. These are R1 institutions. They produce the research everyone else cites. And yet, when McDonald, Johri, Ali, and Hingle Collier (2025) analyzed the GenAI policy documents these universities published between October and November 2023, what they found was a mix of enthusiasm, contradiction, and surprisingly thin reasoning.
The study, published in Computers in Human Behavior: Artificial Humans, coded 141 documents from 116 institutions and found that 63% actively encourage the use of GenAI. That number alone isn’t surprising. What’s worth unpacking is how they encourage it and what they leave out.
Let’s start with what the guidance actually looks like. McDonald et al. report that 41% of institutions offer detailed classroom activities for integrating GenAI, 56% provide sample syllabi, and 50% supply curriculum ideas. Most institutions organize their recommendations into tiers, something like “embrace,” “limit,” or “prohibit,” giving faculty a range of options. That structure sounds reasonable on the surface. But the content within those tiers is where things start to get complicated.
The guidance leans overwhelmingly toward writing. References to coding, math, engineering, and other STEM applications show up in about half the institutions, and McDonald et al. describe those references as vague and often superficial. The discipline of engineering specifically appeared in only 7 out of 116 institutions. That’s a striking gap.
These are research universities, many of them producing cutting-edge work in computer science and data science, and their GenAI guidance barely touches the disciplines most directly affected by the technology. It’s as if the conversation got stuck in the English department.
GenAI Policies in Higher Education
One of the sharper points in the paper is about how institutions have turned GenAI into the main character of the classroom. McDonald et al. (2025) observe that:
So many institutions suggested activities that positioned GenAI as an object to be critiqued, compared, fact-checked, etc. This is, in fact,what we should ask of ourselves, absent GenAI. Institutions seem so concerned about GenAI’s entrance that they confuse acknowledging it with experimenting with it in the classroom. In our analysis, there is no reference to research backed evidence for the guidance that is given. (p. 9)
That’s a line worth pausing on. Critical thinking, source evaluation, argument analysis: none of these are new pedagogical goals. They existed long before ChatGPT. The concern is that institutions may be confusing acknowledging GenAI’s existence with wholesale experimentation that reshapes the classroom around it.
The privacy section of McDonald et al.’s analysis is where the contradictions become hardest to ignore. About 60% of institutions raised privacy concerns, particularly around sharing sensitive data with GenAI platforms and potential FERPA violations. But the authors found that privacy guidance was often vague, inconsistent, and treated like an afterthought.
Some institutions simultaneously encouraged using GenAI for grading and cautioned against sharing student data, a contradiction McDonald et al. flag directly. And only 18% explicitly mentioned the legal implications of FERPA. That’s a strikingly low number for institutions that are, in some cases, telling faculty to feed student work into commercial AI systems.
As McDonald et al. (2025) argue: “many institutions whose guidance encourages the use of GenAI express little to no concern for ethics and privacy associated with using it” (p. 9). The intellectual property angle compounds this. The paper raises the point that when students submit their own essays into ChatGPT for editing or comparison, their intellectual property is being sent into the GenAI ecosystem with no clear protections. Most institutions don’t address this at all.
To their credit, about 52% of institutions did address diversity, equity, and inclusion. The guidance covers biased output, financial barriers to paid AI tools, accessibility, and the paradox that non-native English speakers who benefit most from GenAI writing support may also face increased plagiarism accusations.
I’ve written about similar tensions in the research on multicultural students and academic writing (e.g., Hysaj et al., 2025), where the students most likely to use AI for legitimate language support are the same ones most likely to be flagged. It’s a structural problem that no syllabus statement is going to fix.
There’s also an irony McDonald et al. identify in the flip classroom trend. Flipped classrooms were originally about moving evaluation into instructional time so students would work under supervision with no incentive for dishonesty. GenAI has pushed institutions toward flip models, but the authors note that what’s actually happening isn’t what the flip model envisioned.
GenAI hasn’t just prompted in-class evaluation. It’s become the centerpiece of learning itself, where the skills being taught are primarily to query, prompt-engineer, and critique a chatbot. The classroom hasn’t just flipped. It’s been reorganized around GenAI in ways that weren’t part of the original pedagogical logic.
The detection tools conversation is worth noting too. About 44% of institutions actively discourage using GenAI detection tools, citing their unreliability. That’s a meaningful number, and it aligns with what many researchers have been arguing. I’ve covered the detection debate extensively on this blog, and the evidence consistently points in the same direction: detection is a dead end (e.g., Kalantzis & Cope,2025) and building pedagogy around it creates more problems than it solves.
But here’s what I think McDonald et al. are really getting at, and it’s a point that doesn’t get enough attention in the AI-in-education conversation: the cumulative burden these policies place on faculty. Both encouraging and discouraging GenAI use imply significant rethinking of classroom practices. Flip your classroom. Redesign your assignments. Scaffold everything. Add process documentation. Teach prompt engineering. Monitor for misuse.
And do all of this with no additional time, no additional support, and, as McDonald et al. note, no reference to research-backed evidence for any of the recommendations. “There must be considerable cognitive overhead (for both students and teachers) that accompanies turning in assignments with GenAI prompts and other process data” (p. 9). That’s not a hypothetical. That’s the daily reality for thousands of instructors who are being asked to rebuild their courses around a technology that changes every few months.
The paper also references an earlier and similar investigation by Moorhouse, Yeo, and Wan (2023), which analyzed GenAI guidance from the world’s top 50 universities. That study found a similarly strong embrace of GenAI, with most institutions focusing on academic integrity, assessment redesign, and communication with students. McDonald et al.’s study adds the US R1 context and, crucially, brings a more skeptical lens. The Moorhouse study was more descriptive. McDonald et al. are willing to say that much of this guidance is contradictory, evidence-free, and potentially harmful.
I think that willingness to name the problem is the paper’s biggest contribution. It’s easy to default to “we should embrace GenAI responsibly” and leave it there. McDonald et al. are asking a harder question: what does responsible adoption actually look like when the guidance itself is inconsistent, the evidence base is thin, and the faculty carrying the load are already stretched beyond capacity?
The authors close with a forward-looking note, arguing that:
the introduction of GenAI can be seen as a catalyst for changing assessment and evaluation practices that are more ecologically valid and grounded in principles of fairness, justice, and ethics, but these positive outcomes require, perhaps, a more thoughtful consideration of its role in the classroom. (p. 10)
I agree with that argument. GenAI could be a catalyst. But catalysts need direction. And right now, too many institutions are generating heat without knowing where the reaction is headed.
References
- Hysaj, A., Dean, B. A., & Freeman, M. (2025). Exploring the purposes and uses of generative artificial intelligence tools in academic writing for multicultural students. Higher Education Research & Development, 44(7), 1686–1700. https://doi.org/10.1080/07294360.2025.2488862
- Kalantzis, M., & Cope, B. (2025). Literacy in the time of artificial intelligence. Reading Research Quarterly, 60, e591. https://doi.org/10.1002/rrq.591
- McDonald, N., Johri, A., Ali, A., & Hingle Collier, A. (2025). Generative artificial intelligence in higher education: Evidence from an analysis of institutional policies and guidelines. Computers in Human Behavior: Artificial Humans, 3, 100121. https://doi.org/10.1016/j.chbah.2025.100121
- Moorhouse, B. L., Yeo, M. A., & Wan, Y. (2023). Generative AI tools and assessment: Guidelines of the world’s top-ranking universities. Computers and Education Open, 5, 100151. https://doi.org/10.1016/j.caeo.2023.100151