AI refusal in higher education: the right to refuse, the duty to understand and the diagnostic value of non-use

Synthesis: Zagami argues that refusal deserves to be treated as evidence rather than as a problem to be corrected. Refusal in higher education is not one behavior: a student declining a chatbot for assessed writing, a course team prohibiting it for unaided reasoning, an academic refusing to delegate formative judgment to automated feedback, staff resisting automated triage in student support, and a university delaying procurement are different relations to different systems, and the paper insists they be read separately. Its central distinction is between private refusal, which should be protected, and role-based refusal, which carries a duty to justify non-use where AI materially shapes learning, teaching, assessment, research, administration or governance. Crucially, that duty is not a requirement to adopt: understanding includes the capacity to reject, restrict, audit and critique, so AI Literacy has to cover deciding when a system should not be used. Where AI is infrastructural, refusal cannot be an individual opt-out at all, and it moves to procurement, audit and contestability.

Key Findings

1. Refusal and non-use are not the same thing. Non-use may come from lack of access, confidence, time or training, while refusal is a decision that use conflicts with a value, responsibility or purpose, so a reported refusal rate conflates stances that call for different institutional responses.
2. Outright student non-use is rare in the evidence cited. A 2025 HEPI/Kortext survey of 1,041 UK undergraduates found 92% had used AI tools and 88% had used Generative AI tools for assessment-related purposes, which is near-normalization rather than mature adoption.
3. Faculty refusal is more substantial than student refusal. Shata's 2025 study of 294 full-time faculty at two mid-sized U.S. public universities found 33.6% reported not using generative AI, with reasons spanning readiness, value, professional identity and perceived threat to human intelligence.
4. The duty to understand applies to roles, not persons. A teacher educator preparing graduates for AI-mediated schools, a researcher using AI-enabled search, professional staff working inside AI-enabled support systems and leaders approving procurement all have to understand systems they may personally avoid.
5. Refusal is one of the most important forms of AI literacy. Understanding includes rejecting, restricting and auditing, so the serious question is not whether someone uses AI but whether they can justify their relation to it against the purpose, context and consequences of the work.
6. The right to refuse is unevenly distributed. Students with academic confidence can refuse without penalty while students needing language, accessibility or rapid-feedback support experience refusal as lost opportunity, and secure academics can refuse on principle while casual staff feel pressure to adopt.
7. Institutional delay is part of the same picture. Uneven governance rather than simple refusal shows up in partial adoption and incomplete policy for assessment, admissions and disciplinary processes, where the risk is treating technological intensity as educational improvement.

What refusal is diagnosing

The paper's contribution is interpretive. Where students refuse, universities are prompted to ask whether rules, assessment design or support are unclear, or whether students are defending a meaningful relation to learning. Where students use AI covertly, the questions turn to whether Assessment Validity holds, whether workload is excessive and whether policy has produced fear instead of learning. Where academics refuse, the questions concern professional agency, workload and scholarly craft; where professional staff resist, management is asked what those staff can see about procedural fairness, confidentiality and care. Refusal becomes Learner Agency exercised in a specific institutional setting rather than a technical deficit, and the diagnostic frame sits alongside the growing body of work on why students and staff decline or delay AI in ways their institutions do not anticipate.

Why the ethical field is wider than integrity

Zagami argues that the usual higher education debate, focused on academic integrity, hallucination, privacy and authorship, is necessary but not sufficient. A fuller account of AI refusal also attends to how systems are built: energy-intensive computation, data-center infrastructure, mineral and water use, and the hidden human labor of data labeling, content moderation and maintenance, along with the appropriation of creative and scholarly work into training data. Citing Crawford on AI as extractive infrastructure, Bender and colleagues on the scale of environmental and social cost, and Gray and Suri on ghost work, the paper holds that these concerns widen the field of decision rather than settling it in favor of blanket refusal. Where a university asks students or staff to use AI, it should be able to explain what is known and not known about data practices, labor dependencies, environmental cost, security risks and alternatives. This is where equity enters: those conditions fall hardest on the people with the least room to refuse.

What this means for practice

• Instructors. Provide non-AI pathways where AI use is not essential to the learning outcome, and say plainly what each task is meant to evidence, so that refusing a tool does not mean refusing the learning.
• Academic developers. Treat refusal as part of AI capability rather than a deficit: the training aim is defensible judgment about using, restricting and declining systems, not tool familiarity.
• Administrators and institutions. Preserve human review, contestability and escalation in administrative AI, and define what AI should not do rather than only how it can be scaled.
• Policymakers and program leaders. Make refusal practically possible through clear alternatives, transparent rules, secure tools, workload recognition and protection from covert productivity expectations, since a duty to understand imposed without training or infrastructure becomes hidden workload.

Limitations

• This is a Comment piece: it develops a conceptual and diagnostic category and explicitly does not measure refusal prevalence or review the literature systematically.
• The sector evidence is indicative and drawn from studies with different populations and methods, which the paper says should not be read as comparable measures of a single refusal rate.
• The vocabulary of private refusal, principled role-based non-use, bounded use and compulsory adoption is proposed rather than tested, so how well it holds up against institutional practice is an open question.

Connected Concepts

• AI Literacy
• Learner Agency
• Equity
• Academic Integrity
• AI Governance
• Educational AI Policy
• Change Management
• Assessment Validity
• AI Detection
• Privacy
• Sustainability
• Critical Thinking
• Generative AI
• Large Language Models (LLMs)
• Higher Education

Connected Articles

• Stop Writing for Me: Generative Refusal in AI Tools for Thought — refusal as a design relation between writers and AI tools for thought
• The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students — differential refusal and epistemic injustice in LLM-mediated history education
• Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity — guarded adoption as the middle ground between uptake and resistance
• Pragmatic users and skeptical nonusers: A qualitative typology of ChatGPT adoption in physics education — pragmatic users and skeptical nonusers as a typology of adoption

Citation

Zagami, J. (2026). AI refusal in higher education: The right to refuse, the duty to understand and the diagnostic value of non-use. Higher Education Research & Development, 45(7), 2287-2294.