Last updated: 2026-10-06

U
Undergraduate level

How I Use AI in Teaching, and How Students Use It Responsibly

Principles, practice, and the responsibility students carry for their own work

People often ask how I use generative AI in my teaching, and how I expect students to use it. The short answer is that I do not treat AI as a separate educational problem. AI changes what can be delegated, what must stay with the learner, and how learning activities should be designed. The question is therefore not whether students should use AI, but which uses support learning and which quietly replace it.the core tension: tool or crutch?

This page draws together the arguments that run through the rest of this section. It is a starting point rather than a summary of everything, and each section links to the fuller pages where the evidence and the reasoning are set out.

AI as a Learning Tool, Not a Learning Substitute

AI can support understanding. It can explain a concept in a different way, give feedback on a draft, suggest a counter-example, and handle routine work that would otherwise crowd out thinking. It can also replace the cognitive work from which learning arises, and the two uses can produce identical submitted work.

The evidence on this is fairly specific. In a 30-month panel of 26,811 secondary students, using GenAI to complete homework raised assignment scores by 18%, while closed-book exam scores fell by 20% within six months. Students who used AI as a tutor or critic were largely unaffected, and the penalty concentrated among those whose pattern suggested they were outsourcing the work[1]. That study is a 2026 working paper from one education system, so it is strong evidence about a pattern rather than a precise prediction for a university. The mechanism it points to has a long history in memory research: effortful processing builds what easier processing does not[2]. The full argument, with four practical strategies for keeping the thinking with the student, is in Avoiding the AI Short-Circuit.

The design of the tool matters as much as its presence. A randomised study of nearly a thousand high-school mathematics students found that a general-purpose AI tutor improved practice performance but harmed later performance when access was removed. A tutor configured to offer hints largely avoided that effect[3]. A further study of developers learning an unfamiliar library found that the patterns of use mattered more than the time spent. Asking a question to understand generated code kept learning intact, while handing over the whole task or repeatedly asking for debugging did not[4]. Responsible use therefore means using AI to keep the learner thinking, which is a different thing from keeping them busy. See Teaching with GenAI for how I apply this in practice.the friction is the point

How AI Changes Module Design

When a tool can produce a competent essay or a working class in minutes, the value of an activity has to be argued for afresh. Less weight belongs on routine production, on formulaic writing, and on tasks that can be delegated without learning anything. More weight belongs on justification, reflection, evaluation, decision-making, and an account of how the work was done.

The useful question for any activity is what it is for. Procedural friction that teaches nothing can reasonably be delegated. Mental-model formation, the understanding a student must hold internally, should be protected. Independent verification, the ability to judge whether an output is right, becomes more important as outputs become cheaper. Generative AI in the Curriculum sets out this purpose-first approach, with examples from object-oriented programming and software engineering and a progression that changes the role of AI as students develop expertise.purpose-first design: protect mental models

Design also needs to change the task, not only the rules around it. A policy that tells students to use AI sensibly is a discursive change. A task that cannot be completed without AI and cannot be completed by AI alone is a structural one, and the distinction matters for how well the change holds[5]. The staff and institutional side of this is covered in What Higher Education Needs to Do About GenAI.

Responsible Student Use

I organise what I expect from students around four principles.

  • Transparency. A student should be able to say whether AI was used, how it was used, and why.
  • Accountability. The student remains responsible for correctness, evidence, decisions, and conclusions. A tool does not carry the responsibility for a submitted piece of work.
  • Critical engagement. Outputs should be evaluated rather than accepted. An AI reviewer's objection is a question to answer, not a fix to apply.
  • Learning first. Use AI where it helps the learning objective and avoid it where it replaces that objective.

Transparency is the principle that depends most on trust. Students who believe they will be punished for admitting AI use have a strong reason to hide it, and hidden use cannot be discussed. My own starting position is stated at the outset: I am in favour of using AI, but of using it well. The reflection sections I set ask about learning and productivity separately, with prompts such as "What did you hand to the AI, and what did you keep for yourself?" The reasoning behind that approach is in Teaching with GenAI: Require It, Then Ask What It Did.

The sector has largely settled on a similar pattern. Institutions typically permit AI as a disclosed aid and treat undisclosed AI-generated content as misconduct. Russell Group universities agreed shared principles in 2023 on AI literacy, staff support, adapted assessment, equitable access, and sharing practice[6]. How that plays out in disciplinary terms is covered in Academic Misconduct and GenAI.policy alone won't shift culture

Trust and the Staff–Student Relationship

I think the relationship between a teacher and a student matters more now than it did before generative AI, not less. A tool can produce a competent answer on demand. What it cannot supply is knowledge of a particular student, a record of their earlier work, and a conversation in which they are asked what they did and why. Those are relational, and they are what makes honest use of AI possible to discuss at all.

The reasoning runs as follows. Students hide AI use when they expect to be punished for admitting it. A teacher who knows a student's earlier work can notice a change in a way that reading an anonymous text cannot. Early feedback that names over-reliance only helps if the student trusts the tutor enough to act on it. The research that bears on these links is indirect. Hagenauer and Volet's review finds that teacher–student relationships at university are important and under-researched, and that they are multi-dimensional and context-bound[7]. Ryan and Deci treat relatedness, the need to feel connected to others, as one of the needs that support self-motivation[8]. Edmondson found that a shared belief that a team is safe for interpersonal risk-taking predicted whether members reported and discussed errors[11]. A survey of Australian students associated dissatisfaction with the teaching and learning environment with contract cheating, and its authors recommend teaching environments that nurture strong student–teacher relationships[10]. None of these studies is about generative AI, and none tests the relationship as a cause of honest use.

Expectation needs the same caution. Jussim and Harber's review of thirty-five years of research finds that self-fulfilling effects of teacher expectations are typically small, and that larger effects may fall selectively on students from stigmatised groups[9]. So an expectation of good work is one ingredient among several, and it has to reach every student, including the ones who never come to office hours.

There is an irony in the current situation. Generative AI can explain a concept, produce a worked example, or draft a feedback comment, at any hour and without tiring. It cannot know a student's history, notice that someone has gone quiet, or be held to account by them. The parts of teaching that remain are the ones that depend on people: showing that you expect good work from a student, admitting your own fallibility, and making it safe for a student to say "I leaned on the tool too much". My reading is that AI makes these human parts of teaching more visible, and that a module without them is easier to replace. That is my interpretation. The studies above do not test it.

The limit is that the relationship only reaches students who make contact. The engagement research that Supporting Students with GenAI draws on suggests that the students least likely to make contact are often the ones most at risk, so trust cannot be the only route. Assessment design and a clear, easy route to a conversation have to work for students who never form a relationship with their tutor, and fairness must not depend on rapport. Whether relationship-based trust adds anything beyond policy and assessment design is an open question, and the test designs in Supporting Students with GenAI show how it could be studied fairly.

Assessment in an AI World

If AI can produce much of a conventional submission, assessment has to ask for something it cannot produce on demand. Assessment should increasingly evaluate judgement, reflection, synthesis, and understanding rather than production alone.

Several designs do this in practice. A design rationale asks why a particular approach was chosen over a plausible alternative. A reflective commentary asks what the student learned and what they would change. A critique of AI output asks the student to evaluate a generated solution and say where it fails. An evidence trail records the process, including the decisions and the points at which the student checked something. Iterative submissions make the progression of the work visible, so that a finished piece cannot be the only evidence.

The AI Assessment Scale gives five levels of AI involvement that an assessment can be built around, from no AI to full AI exploration[7]. Choosing the level deliberately is more useful than a single blanket rule. The reflection section that I add to assessments is one place where the learning-and-productivity question gets asked separately, and it is most useful as material for a conversation rather than as a measurement.

When early work looks over-reliant on AI, I say so in the feedback and do not penalise it directly, unless the abuse is egregious. Noticing AI use is fallible. Detection studies have found that teachers struggle to tell AI-generated essays from student writing, and that formal processes catch only part of what markers notice[8]. A conversation costs little if the suspicion is wrong. A penalty costs a great deal. The reasoning is in Teaching with GenAI.penalty vs conversation costs differ

Rubrics have a part to play in this, but they carry a risk of their own. A rubric that specifies the criteria tightly can encourage students to satisfy the criteria rather than develop the underlying skill, and a rubric supplied at the same level of detail throughout a degree can stop students practising judgement without one. Rubrics work best when they describe the general skill rather than the task, and when their detail is reduced as students progress. The evidence and the design choices are in Evaluating Detailed Rubrics.

How I Use AI as an Educator

AI is useful for the work around teaching as well as for the students' work. Areas where it can help include generating and testing explanations, producing alternative examples, drafting feedback that a marker then rewrites, checking whether a set of comments covers every criterion, generating scenarios for practice, and reviewing a curriculum for gaps. In each case the output is raw material, and the judgement about what to keep stays with the teacher.

The limits matter as much as the uses. AI output needs human oversight. It can reflect biases in its training data, it can state wrong things fluently, and it does not have the contextual judgement of someone who knows the cohort, the module, and the students. A model that gives a plausible-sounding explanation has not thereby given a correct one.

The most demanding example I have worked on is a local pipeline that marks archived project reports against a rubric. It is a working system, but it is not ready for use. It runs only on local hardware, so no student work leaves the machine, and it has only been run on archived reports that a marker had already finished grading. The University of Reading's guidance is explicit that GenAI must not determine marks, and that bounded assistance is appropriate only after a human academic judgement has been made and all outputs have been reviewed[9]. The pipeline has so far been over-generous in some places and over-corrected in others. A model that rewards technical vocabulary without evidence is one failure, and an instruction that penalises real work for sounding generic is another. The evaluation is written up in Local AI for Marking Support: A Live Build, and Why It Isn't Ready Yet.the live build: why it is not ready yet

Looking Ahead

The tools will change, and some of the questions that matter now will be easier to answer later. The more interesting questions concern what a learning assistant should know about a student, and how that knowledge should be shared.

An assistant that adapts to a learner needs a model of that learner. I think the model should be open to the learner: visible, contestable, and clear about the difference between evidence and inference. A learning assistant that says "I have noticed this pattern, here is the evidence, and here is my uncertainty" supports agency in a way that a reassuring tone does not. The argument is developed in The Caring Learning Agent, and the learner's own model of their learning is discussed in The Self-Model as a Learning Tool.

Sharing that model should be scoped. A tutor may need to see where a student is stuck, a peer may need a shared goal, and a separate assistant may need a summary for one task. Each of those should be a separate view, with a defined purpose, audience, and lifetime, and the student should be able to withdraw it. The argument is in Networked Learning Assistants and Personal Learning Networks.

The aim throughout is to keep the learner's agency intact. Assistants that do the work for the student take away the thing the work was meant to build, however helpful they seem.agency means the learner chooses the path

Conclusion

My approach to AI in education is not built around detecting AI use or prohibiting it. It is built around delegation. Every educational activity involves deciding what the learner does, what they do with support from others, and what may be handed to a tool. Generative AI changes those boundaries, but it does not remove the need for learning. The challenge for teachers is to design activities that develop understanding, judgement, and agency, while making deliberate and responsible use of increasingly capable AI systems.

Read More

References

  1. Strömberg, D., Lei, V., & Wu, Y. (2026). The Generative AI Learning Penalty: Evidence from Chinese Secondary Education (CEPR Discussion Paper No. 21577). Centre for Economic Policy Research.
  2. Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11(6), 671–684. https://doi.org/10.1016/S0022-5371(72)80001-X
  3. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakçı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
  4. Shen, J. H., & Tamkin, A. (2026). How AI Impacts Skill Formation. arXiv:2601.20245
  5. Corbin, T., Dawson, P., & Liu, D. (2025). Talk is cheap: why structural assessment changes are needed for a time of GenAI. Assessment & Evaluation in Higher Education, 50(7), 1087–1097. https://doi.org/10.1080/02602938.2025.2503964
  6. Russell Group. (2023). Principles on the use of generative AI tools in education. https://www.russellgroup.ac.uk/policy/policy-briefings/principles-use-generative-ai-tools-education
  7. Hagenauer, G., & Volet, S. E. (2014). Teacher–student relationship at university: an important yet under-researched field. Oxford Review of Education, 40(3), 370–388. https://doi.org/10.1080/03054985.2014.921613
  8. Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68
  9. Edmondson, A. C. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383. https://doi.org/10.2307/2666999
  10. Bretag, T., Harper, R., Burton, M., Ellis, C., Newton, P., Rozenberg, P., Saddiqui, S., & van Haeringen, K. (2019). Contract cheating: a survey of Australian university students. Studies in Higher Education, 44(11), 1837–1856. https://doi.org/10.1080/03075079.2018.1462788
  11. Jussim, L., & Harber, K. D. (2005). Teacher expectations and self-fulfilling prophecies: Knowns and unknowns, resolved and unresolved controversies. Personality and Social Psychology Review, 9(2), 131–155. https://doi.org/10.1207/s15327957pspr0902_3
  12. Perkins, M., Roe, J., & Furze, L. (2025). Reimagining the Artificial Intelligence Assessment Scale (AIAS): A refined framework for educational assessment. Journal of University Teaching and Learning Practice, 22(7). https://doi.org/10.53761/rrm4y757
  13. Scarfe, P., Watcham, K., Clarke, A., & Roesch, E. (2024). A real-world test of artificial intelligence infiltration of a university examinations system: A "Turing Test" case study. PLOS ONE, 19(6), e0305354. https://doi.org/10.1371/journal.pone.0305354
  14. University of Reading, Centre for Quality Support and Development. Artificial Intelligence: Assessment and Feedback. https://www.reading.ac.uk/cqsd/artificial-intelligence/assessment_and_feedback