Responsible AI in Education: Learning With, Not Instead Of, AI

As an educator, I see generative AI changing classrooms faster than any technology before it. Students can ask an AI tutor to explain backpropagation at midnight, practise conversation in a new language, or get feedback on code. They can also outsource their thinking entirely — and learn nothing. How we use these tools will shape a generation's education. This lecture offers guidance for students and teachers alike.

The opportunity: personalised learning at scale#

Benjamin Bloom's famous "2 sigma problem" (1984) reported that students tutored one-to-one performed about two standard deviations better than students in conventional classrooms — but individual tutoring is too expensive to provide for everyone. AI tutors raise the possibility of affordable, personalised support:

• Explanations on demand, at the right level, in the learner's own language.
• Practice and feedback: generating exercises, checking reasoning, giving hints rather than answers.
• Accessibility: support for learners with disabilities (reading aids, speech interfaces) and for those studying in a second language.
• Teacher support: drafting lesson materials, differentiated exercises and rubrics, freeing time for human interaction.

Early studies of carefully designed AI tutoring have shown promising learning gains in some settings, particularly when the tutor is designed to guide rather than give answers. Evidence is still developing, and design matters enormously.

The risks#

1. Outsourcing thinking: learning requires effortful practice ("desirable difficulties"). If AI does the struggle, students may produce good answers while learning little — an illusion of competence. A randomised study with high-school maths students (Bastani et al., 2024) found that unrestricted access to a chatbot improved practice performance but reduced later performance on exams without AI, while a tutor-style version with guardrails largely mitigated the harm.
2. Hallucinations: confident errors in explanations, citations or code can mislead learners who cannot yet spot them.
3. Academic integrity: submitting AI-generated work as one's own.
4. Equity: unequal access to devices, connectivity and paid tools can widen gaps; models often perform worse in lower-resource languages.
5. Privacy: student data entered into AI tools may be stored or used for training.
6. Homogenisation: essays and ideas converging towards generic AI styles.

Why AI-text detectors are not the answer#

Guidelines for students#

Guidelines for educators#

• Write a clear AI policy for each course: what is allowed, what must be disclosed, and why.
• Design assessments for learning: in-class work, oral explanations, drafts and reflections, projects grounded in local contexts and personal experience, and tasks that require critique of AI outputs.
• Teach AI literacy: how models work, where they fail, how to verify, ethical and privacy issues.
• Use AI to support, not replace, feedback and relationships — students still need human mentors.
• Choose tools responsibly: data protection, accessibility, language coverage and cost for students.
• Model good practice by being transparent about your own use.

UNESCO has published guidance on generative AI in education and research, recommending, among other things, age-appropriate use, data protection and building educators' capacity.

An example: a tutor-style prompt#

You are a patient tutor for a university machine learning course.
- Never give the final answer to an exercise directly.
- Ask me what I have tried, then give one hint at a time.
- When I make an error, ask a question that helps me find it.
- Use simple language and check my understanding with a short question at the end.
- If you are not sure about a fact, say so.

This shifts the AI from answer machine to Socratic partner — the design principle behind the most promising educational uses.