Teaching with Generative AI
3.1 What the Evidence Says
In Brief
- AI produced content can harm and damage student learning even if their work appears better on the surface
- Main stream use of AI damages critical thinking and brain development
- AI designed around sound pedagogy that encourages thinking, can produce learning gains.
- The difference lies in the design of the activity and how / when tools are used
Research Evidence
| Study | What was Found | Implication for Teaching |
|---|---|---|
| Bastani et al. (2025) Field experiment, c. 1,000 secondary maths students |
Unrestricted GPT-4 access improved immediate performance but students performed worse after access was removed. Tutor-style systems with guardrails reduced this effect. | Guardrails matter. Set up AI to support learning, not to replace it. |
| Kestin et al. (2025) RCT, Harvard physics |
An AI tutor built on research-based pedagogy produced learning gains more than double those of an in-class active-learning session, in less time | Well-designed AI tutoring can complement teaching, especially for preparation before class. |
| Fan et al. (2025) BJET |
Students using ChatGPT improved essay scores but not knowledge gain or transfer, and showed "metacognitive laziness". | A better product is not evidence of learning. Finding ways of assessing the process of learning is. |
| OECD Digital Education Outlook 2026 | Over-reliance lowers metacognitive engagement, "creating a misalignment between task performance and genuine learning". Tools "intentionally designed or used for pedagogical purposes show more sustained learning benefits". | Develop ways of integrating AI use that supports learning through critical engagement by the student. |
The learning science that explains these emerging trends:
Durable learning depends on effortful processing: retrieval practice (Roediger and Karpicke, 2006), "desirable difficulties" (Bjork and Bjork, 2011) and building students' capacity to judge the quality of work through evaluative judgement (Tai et al., 2018).
3.2 Designing Learning with AI
Five Design Principles
- Start from the learning outcomes. For each outcome, ask: would AI help students achieve this, or do the work for them? This is constructive alignment (Biggs, 1996) applied to a new tool.
- Protect the productive struggle. Where the struggle is the learning, for example first drafts, core problem-solving, foundational skills- keep AI out or delay it.
- Use AI to generate practice, not answers. Quizzes, worked examples to critique, counter-arguments and case variations support retrieval and self-explanation.
- Make students the judge. Ask students to verify, critique and improve AI output against sources. This builds discernment and evaluative judgement.
- Make thinking visible. Ask for drafts, decision logs, prompt histories and reflections, so that the process becomes part of the learning and the evidence.
Activity Ideas
| Activity | How it Works | Learning Purpose | Typical AIAS Level |
|---|---|---|---|
| Critique the AI Output | Students evaluate an AI-generated response against criteria or a rubric. | Evaluative judgement and disciplinary standards. | 2–3 |
| Socratic Tutor | AI asks questions and provides hints without giving answers. | Self-explanation and metacognition. | 3 |
| Retrieval Quiz Generator | Students create practice quizzes from their own notes before answering independently. | Retrieval practice. | 2 |
| Role-play and Simulation | AI plays a stakeholder, client, patient or examiner. | Transfer to professional contexts. | 3–4 |
| Compare and Justify | Students compare their own answer with AI-generated work. | Deep processing and confidence calibration. | 3 |
| Prompt-and-Reflect Log | Students document prompts, outputs and revisions. | AI literacy and academic integrity. | 3–4 |
3.3 Using GenAI in Your Own Practice
In Brief
- GenAI can save time on case studies, question banks, rubric drafts, example data and feedback templates.
- You remain responsible for accuracy, bias, copyright and data protection in anything you use.
- Tell students when AI has helped produce materials or feedback - the HEA expects this, and it models the transparency you ask of your students.
Do
- Use MTU-approved tools such as Microsoft Copilot while signed in with your MTU account.
- Check every factual claim and generated reference.
- Provide a brief disclosure statement describing how AI was used.
Don't
- Upload student work, grades, names or IDs to unapproved systems.
- Use AI to decide marks, progression or admissions. Evaluating learning outcomes is a high-risk use under the EU AI Act and needs human judgement.
- Upload copyrighted material you do not have rights to, e.g. full textbook chapters.
Example Disclosure Statement
"Practice questions drafted with Microsoft Copilot and reviewed by the module team"