Last updated: 2026-10-05
Generative AI in the Curriculum: Delegation, Collaboration, and Educational Purpose
Asking what an activity is for before deciding what role AI should play in it
Generative AI changes a curriculum whether or not anyone redesigns it. A student can now produce a working class, a summary of a statute, or a first draft of an architecture document in minutes. An activity that once took hours of practice may no longer need doing by hand, or may need doing by hand for a reason that is no longer obvious. Either way, the module team has to decide.
The useful question is not "Can students do this without AI?" It is "Why should students do this, and what role should AI play?" The first question asks whether a task is still possible. The second asks what the task is for, and it is the one that a team can answer for each activity in a programme.the 'why' is the pedagogy the 'what' is the tool
1. Start from Educational Purpose FoundationalKnowledge that endures for decades — core principles
An activity is best evaluated by the educational work it does. The same output, a piece of code or a written analysis, can serve very different purposes. A student who writes a sorting routine to learn how loops and indices interact is doing something different from a student who writes one because the module needs a sorting routine. The artefact may be identical. The learning is not.the map matters more than the territory
The table below sets out the purposes that recur across curriculum design, with the role AI would normally play for each. These are starting points for discussion, not rules.
| Category | Educational purpose | Typical AI role |
|---|---|---|
| Procedural friction | Removes effort that teaches nothing | Delegate |
| Authentic professional practice | Reflects how the work is done in practice | Integrate |
| Necessary practice | Builds competence through repetition | Restrict or scaffold |
| Mental model formation | Develops a conceptual understanding that the student must hold internally | Limited support |
| Independent verification | Ensures students can judge whether an output is right | Challenge and check |
| Delegation after competence | Allows efficiency once the underlying understanding exists | Delegate selectively |
| Collaborative amplification | Improves outcomes through a wider range of perspectives | Collaborate |
| Dialogic understanding | Builds understanding through questioning and discussion | Tutor and discussion partner |
A single activity usually serves more than one purpose, and the purposes can change over a programme. A group project in the final year may be mainly about professional practice, while the same skill in the first year is mainly about building a mental model. The team should record the primary purpose of each activity and revisit it as students progress.
2. Beyond Delegation and Prohibition Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Curriculum debates often assume a choice between two things: students do the work themselves, or they hand it to an AI. That choice leaves out most of professional practice. Software is reviewed by more than one engineer. Policy is drafted by people with different expertise who argue about it. Research is checked by colleagues who did not write it. In each case the final judgement is better because several perspectives met.
Generative AI can take a place in such a process as another contributor. This shifts the question. The useful one becomes: when does working with an AI produce something that neither the student nor the AI would produce alone?
Examples where this is plausible include design review, security analysis, debugging, requirements analysis, ethical evaluation, research synthesis, generating test cases, and finding edge cases. In each, the AI is not doing the student's thinking. It is widening the set of possibilities the student considers before deciding.
The many-eyes principle Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Engineering practice has long relied on review by more than one person, because a single author tends to miss what they already assume. Formal inspection of design and code, set out by Fagan in 1976, is one established form of this review[3]. An AI reviewer can play a similar role. It may point to a vulnerability the author did not see, suggest an alternative reading of a requirement, make a hidden assumption visible, produce a counter-example, or flag an edge case. Its suggestions are not automatically correct, and that is the point: they are things to evaluate.
The student brings what the AI cannot: knowledge of the context, awareness of stakeholders, professional judgement about what matters, and responsibility for the result. The analysis that comes out of both is often better than either would have produced alone. The learning is in the evaluation.
Complementary strengths Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Human and AI contributions differ in ways that make the combination useful. Humans bring contextual understanding, domain expertise, value judgements, awareness of organisational constraints, empathy for the people affected, and accountability for decisions. An AI brings fast retrieval across many examples, a wide spread of suggestions, recall of rare cases, persistence in exploring an option, and consistency. Neither set covers every situation. Educational value often lies in the gap between them, which is also where students learn to notice what each side misses.the skill is in the gap, not the output
3. Modes of Collaboration Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
The role of the AI changes depending on who leads the task. Four modes are useful in curriculum design, along with one where the teacher leads and the AI responds.
| Mode | Human role | AI role | Main educational value |
|---|---|---|---|
| AI as tool | Directs the work | Executes it | Efficiency |
| AI as tutor | Learns | Explains | Understanding |
| AI as reviewer | Produces the work | Critiques it | Verification |
| AI as collaborator | Co-develops | Co-develops | Joint reasoning |
| AI as generator | Evaluates | Produces options | Oversight and judgement |
| Human as mentor | Guides the inquiry | Responds | Reflection |
An activity should state which mode is intended. "Use AI as you see fit" leaves students guessing about what the task is for, and leaves assessors guessing about what good work looks like. A brief that says "the AI is your reviewer for this design; you must decide which of its objections to accept and say why" is clear about both.the brief is the contract with the tool
4. Dialogue as a Learning Mechanism Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Most earlier educational technologies answered questions or delivered content. Generative AI supports dialogue, which lets a student ask a follow-up, test a partial understanding, request another explanation, compare two ideas, challenge a response, or explore a hypothetical. The educational value may lie less in the answer than in the conversation that produces it.
The distinction matters when judging whether AI use helps or undermines learning. The same tool can be used to retrieve an answer, complete a task, or develop an understanding, and each produces a different outcome. Chi and Wylie's ICAP framework proposes that learning increases as students move from passive engagement through active and constructive to interactive engagement, where they build on each other's ideas in dialogue[1].
The evidence on unrestricted use is a warning. In a study of nearly a thousand high-school mathematics students, access to a general-purpose generative AI tutor improved performance while students practised with it, but students who lost access performed worse than students who had never had it. A tutor designed with guardrails, which steered students towards hints and explanations rather than answers, largely avoided that effect[2].
| Activity | Primary outcome |
|---|---|
| Answer retrieval | Information acquired |
| Delegation | Task completed |
| Dialogue | Understanding developed |
| Critique | Evaluation capability developed |
| Collaboration | Judgement enhanced |
A student who asks "Why is composition preferable to inheritance in this design?" and then argues with the reply is doing something quite different from a student who asks for a finished class diagram. Both may use the same tool for the same length of time. Only the first is likely to build a model the student can use later.arguing with the reply is productive struggle
5. Worked Example: Object-Oriented Programming Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
In an introductory object-oriented module, the primary purposes are mental-model formation, necessary practice, and independent verification. Students need an internal model of objects, classes, inheritance, polymorphism, abstraction, and how objects interact with one another. The goal is conceptual understanding. Producing code quickly is secondary.
| Activity | Educational purpose | Appropriate AI role |
|---|---|---|
| Writing simple classes | Mental-model formation | Limited support |
| Creating object relationships | Model formation | Scaffold only |
| Explaining inheritance | Conceptual understanding | Dialogue partner |
| Generating boilerplate | Procedural friction | Delegate |
| Debugging interactions between objects | Verification and understanding | Collaborate |
| Refactoring a design | Judgement, after competence | Collaborate |
Consider a library-management exercise. The AI, asked to review the student's code, identifies duplicated logic, weak encapsulation, and a field that is never used. The student identifies a requirement the code does not meet, an abstraction that suits the code but not the domain, and a concern about how loans are recorded for members who have left the library. The two sets of observations together give the student a better review than either would have produced alone, and each observation is a question to answer rather than a fix to accept.the friction is the point
In dialogue, a student might ask why composition would be preferable to inheritance for loan status. A useful reply compares the two designs, gives an example in which inheritance breaks down, and offers a counter-example in which it would be the right choice. The student then has to decide which design fits the requirements and say why.
6. Worked Example: Software Engineering Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
A software engineering module has a particular position in this discussion. Collaborative development is already normal professional practice, so AI use here is closer to a real working environment than in most other subjects. The primary purposes are authentic practice, collaborative amplification, verification, and professional judgement.
| Activity | Educational purpose | Appropriate AI role |
|---|---|---|
| Requirements analysis | Human judgement | Limited assistance |
| Writing user stories | Professional practice | Collaborate |
| Design exploration | Creative reasoning | Collaborate |
| Routine code generation | Tool use | Delegate selectively |
| Code review | Collaborative amplification | Collaborate |
| Security review | Many-eyes analysis | Collaborate |
| Test generation | Professional efficiency | Delegate, after competence |
| Architecture selection | Professional judgement | Human-led collaboration |
Architecture work shows both collaboration modes clearly. In a human-led version, students design a system and the AI critiques it for scalability problems, security weaknesses, and technical risks. The students then weigh those critiques against stakeholder needs, organisational constraints, and maintainability, and decide which to act on. In an AI-led version, the AI proposes several architectural alternatives. The students compare them, justify their choice, reject suggestions that do not fit, and explain the trade-offs. Both versions develop judgement, and both mirror what a practising engineer would do.AI proposes, humans judge
7. Worked Example: Data Integrity and Ethics Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Ethics and law modules show the limits of delegation most clearly. Their primary purposes are ethical reasoning, legal interpretation, critical evaluation, verification, and reflective judgement. A central point is that knowledge and responsibility are different things. An AI can contribute analysis, but it cannot take responsibility for a decision. That stays with a person.
| Activity | Educational purpose | Appropriate AI role |
|---|---|---|
| Learning legal concepts | Understanding | Tutor |
| Summarising legislation | Procedural support | Delegate, with checking |
| Comparing ethical frameworks | Reflection | Collaborate |
| Evaluating a dilemma | Human judgement | Discussion partner |
| Checking compliance | Verification support | Collaborate |
| Deciding on accountability | Professional responsibility | Human only |
Take an educational analytics system. An AI may point to possible concerns about consent, retention periods, transparency obligations, and the lawful basis for processing. Students may point to things the AI cannot see: the institutional context, the effect on learners who are already struggling, and the competing interests of students, staff, and the institution. Put together, the two lists give a richer evaluation than either.
A dialogue might begin with a student asking why a utilitarian and a deontologist would disagree about the use of student data. The AI can set out both positions, press each on its weak points, and offer examples that test them. The aim is ethical understanding. A student who ends the conversation with a sharper argument of their own has learnt more than one who ends with a summary.the friction is where the learning lives
8. Programme-Level Progression Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
The role of AI should change as students develop expertise. A programme that allows delegation before understanding exists produces graduates who cannot check what they receive. A programme that prohibits AI where professional practice requires it produces graduates who have never worked the way their profession works. A progression avoids both problems.
| Stage | Dominant educational purpose | Typical AI role |
|---|---|---|
| Early study | Mental-model formation | Tutor and explainer |
| Developing competence | Practice and feedback | Scaffold and reviewer |
| Intermediate study | Verification and critique | Reviewer and challenger |
| Advanced study | Professional judgement | Collaborator |
| Professional level | Augmentation and productivity | Integrated team member |
Teams can use this to check a single module against its place in the programme. A first-year activity that asks students to delegate their core practice is probably in the wrong place. A final-year activity that forbids the tools a professional would use is probably in the wrong place too.
9. What Assessment Should Check Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
If AI use is part of an activity, assessing whether the student avoided it tells us little. Assessment should focus on whether the student exercised sound judgement, evaluated alternatives, justified decisions, and used AI critically and responsibly. The brief for each activity should say what those things look like for that activity, so that students know what is being judged.
This also changes how a team can read an artefact. A finished piece of code, a report, or a design says little about the process that produced it. Where process matters, the activity can ask for the record of the dialogue, the critiques the student accepted and rejected, and the reasons. These are more informative than the artefact alone, and they are harder to produce by delegation.
Conclusion FoundationalKnowledge that endures for decades — core principles
Curriculum evaluation should not ask only whether an AI can perform an activity. It should ask what educational purpose the activity serves. Some activities remain essential because they build mental models, competence, or independent judgement. Some may legitimately be delegated. Others gain educational value through dialogue, critique, and collaboration between humans and AI. The question is therefore not whether the human or the AI should perform a task, but how responsibility, understanding, and expertise develop through the form of interaction that best serves the purpose.
Related Topics
- Pedagogy for Networked and Agentic Learning — the design of learning when students work with agents and other people at once.
- The Caring Learning Agent — an assistant that keeps a visible, contestable model of the learner, which makes the dialogue in section 4 traceable.
- Networked Learning Assistants and Personal Learning Networks — the permissions that govern what an assistant may see while it collaborates.
- Contradiction and the Productive Struggle of Learning — why a reviewer's objection can be the most useful thing a student receives.
- Rubrics in HE — how assessment criteria can reward judgement about AI output rather than its absence.
References
- M. T. H. Chi and R. Wylie, "The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes," Educational Psychologist, 49(4), 2014, pp. 219–243. https://doi.org/10.1080/00461520.2014.965823
- H. Bastani, O. Bastani, A. Sungu, H. Ge, Ö. Kabakcı and R. Mariman, "Generative AI without guardrails can harm learning: Evidence from high school mathematics," Proceedings of the National Academy of Sciences, 122(26), 2025, e2422633122. https://doi.org/10.1073/pnas.2422633122
- M. E. Fagan, "Design and code inspections to reduce errors in program development," IBM Systems Journal, 15(3), 1976, pp. 182–211. https://doi.org/10.1147/sj.153.0182