Use AI as a Tutor, Critic and Thinking Partner—not a Ghostwriter

Ask for algebraic, graphical, numerical or geometric comparisons when more than one method is possible. The student should still decide which method applies and explain why it works.

Avoid asking for a complete worked solution before making an attempt. That can create the appearance of understanding without developing the ability to generate a solution.

Essay writing: make AI a demanding reader

The student should retain ownership of the thesis, evidence, structure and voice. AI can serve as a demanding reader who identifies weaknesses without taking over the writing.

Give the tool your thesis and outline, then ask:

• “What is unclear, unsupported or logically inconsistent?”
• “Distinguish my claims from the evidence supporting them.”
• “Where have I confused summary with analysis?”
• “Argue against my thesis as a sceptical reader.”
• “What evidence would weaken my conclusion?”
• “Review the argument, structure, evidence, paragraph coherence and style separately. Do not rewrite the essay.”

Revise one section yourself and ask whether the revision addresses the original problem. Use AI for grammar and clarity after the ideas are settled, and keep any suggested wording only when you understand and genuinely choose the change.

Research on student use of AI reports benefits for understanding complex topics, finding evidence, revising and proofreading, but findings for critical thinking and originality remain mixed. A recent synthesis found stronger outcomes when students used AI selectively and iteratively with instructor guidance. Other research indicates that structured AI feedback can support critical writing, but feedback is most useful when students actively evaluate and apply it rather than accept it automatically.

Human teachers and peers remain important for disciplinary judgment, conceptual development and context. Students should also follow their institution’s rules on permitted use and disclosure.

Research and reading: use AI to interrogate sources

AI should deepen reading, not replace it. Before opening a chatbot, write a short summary, identify the author’s thesis and list two questions about the text.

Then ask questions such as:

• “What assumptions does this author make?”
• “What evidence supports each major claim?”
• “What would a critic say?”
• “Which terms require precise definitions?”
• “What evidence is missing?”
• “What alternative explanations fit the same evidence?”

AI can suggest search terms, related concepts and possible source leads. It should not be treated as a source database: verify every important claim against the original publication, especially quotations, references, statistics and technical details.

A comparison table can also help with two readings, but the student should fill in and check each cell against the texts. Keep a research log that separates:

• what the source explicitly says;
• what you infer from it;
• what AI suggested; and
• what you independently verified.

After the discussion, close the AI tool and write the author’s argument from memory, including one limitation or unresolved question. If you cannot do that, the reading has probably been bypassed rather than deepened.

Science: explain, expose gaps and predict

The Feynman technique asks learners to recall and explain a concept in simple language, identify gaps in that explanation, repair them and simplify again. AI can make the process interactive by acting as a curious beginner or a sceptical reviewer.

Useful prompts include:

• “I will explain photosynthesis. Ask me one question at a time as a curious beginner.”
• “Flag every undefined technical term.”
• “Give me a case where my explanation predicts the wrong outcome.”
• “What observation would distinguish these two hypotheses?”
• “Which part of my explanation describes a mechanism, and which part is only a correlation?”
• “Change one variable and ask me to predict what happens.”

The student should draw the diagram, state the causal chain and make the prediction. AI can then probe edge cases, surface misconceptions and offer counterexamples. The final check is an unaided explanation to an imaginary beginner.

Coding: review the reasoning before requesting a fix

For coding, AI works well as a reviewer, debugger and testing partner. Start with a first version of the program, then ask for help locating the problem rather than replacing the entire solution.

Try:

• “Identify the smallest failing assumption. Do not rewrite the code.”
• “Trace the variables until the output becomes incorrect.”
• “What invariant should remain true?”
• “Give me an input that breaks this algorithm.”
• “Suggest tests for a normal case, boundary case, invalid input, empty input and adversarial case.”
• “Compare the time and space complexity of these two approaches.”

Implement the fix yourself, run the tests and explain why the change works. Then alter one requirement and modify the program without asking AI to regenerate everything.

A useful standard is that students should be able to explain every submitted line, reproduce the bug, justify the fix and write at least one test that was not supplied by AI.

What Singapore’s Learning Assistant and Feynman Bot get right

Singapore’s Ministry of Education describes AI tools in the Student Learning Space as educational systems that support learning through questions, scaffolds and feedback rather than direct answers. The Learning Assistant can take different roles, including a writing-assistant role, and engage students in iterative questioning intended to sharpen understanding.

That design demonstrates an important principle: the student’s response remains the central activity. The AI creates productive friction by asking the next question, pointing to a gap or requesting clarification instead of immediately completing the task.

The guardrails do not make the system infallible. MOE’s student guidance states that Learning Assistant responses are probabilistic and may be inaccurate, and advises students to reflect on and fact-check them using other sources.

SUTD’s Feynman Bot applies a similar model to conceptual learning. Students upload learning materials, choose an appropriate level and explain the topic while the bot asks adaptive questions. Reports describe a limit of roughly three or four “I don’t know” responses before the bot suggests an approach and asks a new question based on the weakness. The related research presents the bot as an AI implementation of the Feynman technique for self-regulated learning.

The lesson is not that every AI tutor should refuse to answer indefinitely. It is that the tool should delay completion long enough for the learner to attempt, explain and confront gaps in understanding.

Rules that protect authorship and learning

Students, teachers and institutions can use simple boundaries:

• Make a first independent attempt before using AI.
• Request questions, hints, critique or diagnosis before a solution.
• Do not submit prose, proofs, code or analysis that you cannot explain.
• Finish with an unaided explanation or a parallel problem.
• Verify facts, calculations, citations, quotations and code.
• Keep drafts, prompts, feedback and revision notes when disclosure is required.
• Use teachers and peers for judgment, motivation, ethics and high-stakes evaluation.
• Treat AI disagreement as a reason to investigate, not proof that you are wrong.

The student should supply the initial thinking, make the important decisions and produce the final explanation. AI should supply questions, objections, alternative interpretations, tests and feedback. Repeated cycles of attempting, comparing, critiquing, revising and explaining are what turn AI assistance into durable learning rather than outsourced work.