Learning Alongside AI

For the first time, almost anyone can summon a patient, tireless explainer that answers at any hour, rephrases on request, and never sighs at a repeated question. Used well, an AI chatbot can compress the frustrating gaps in self-study — the half-understood paragraph, the missing example, the question you were too shy to ask.

Used badly, the same tool quietly replaces the very effort that builds understanding. The danger is not that AI is unhelpful; it is that it is helpful in a way that feels like learning while sometimes preventing it. This lesson is about capturing the upside — speed, patience, personalisation — while keeping the effortful thinking that only you can do.

AI as a Socratic tutor

The most powerful move is to make the AI ask you questions rather than answer yours. Prompt it to quiz you, to withhold the solution until you have attempted it, to ask what you already know before explaining. This turns a passive answer machine into a retrieval partner, forcing the recall and reasoning that actually build memory.

AI as explainer and practice partner

AI is genuinely good at rephrasing a hard idea five different ways, generating fresh practice problems, role-playing a conversation in a new language, or giving fast feedback on a draft. These are real accelerants. The key is that the effort stays with you: you attempt, then check; you predict, then compare. The AI supplies material and response, not the thinking.

Verification and the active thinker

Large language models generate fluent, confident text that is sometimes wrong — so-called hallucination. A plausible citation may not exist; a clean explanation may misstate a fact. So treat AI output as a knowledgeable but fallible study partner, not an oracle. Cross-check anything important against a textbook or primary source, and stay the active thinker who questions, tests, and retrieves rather than absorbs.

A student learning statistics is stuck on Bayes' theorem. Instead of asking "explain Bayes' theorem," she prompts: "Ask me questions to check what I understand, and don't give the full answer until I've tried." The AI asks what a conditional probability means; she fumbles, corrects herself, and the gap becomes visible. It then poses a worked scenario and waits. She attempts, gets feedback, and asks for two more problems. Twenty minutes later she can solve them unaided — and she double-checks one surprising result against her textbook.

Her classmate opens the same tool and types "solve these ten problems and show the steps." The answers arrive instantly and look perfect. He reads them, nods, and moves on, feeling productive. On the exam, with no AI, he cannot start a single one. He had outsourced the struggle that would have built the skill, and mistook a smooth explanation for his own understanding.

In a 1984 article in Educational Researcher, the educational psychologist Benjamin Bloom described what he called the "2 sigma problem." Summarising studies from his students, he reported that students taught one-to-one by a tutor, using mastery methods, performed on average about two standard deviations better than students in conventional classrooms — the average tutored student scoring above roughly 98% of the untutored group.

That result framed a decades-old dream: individual tutoring works extraordinarily well, but cannot be provided to everyone at scale. AI tutors are exciting precisely because they promise one-to-one attention cheaply. But honesty matters: Bloom's figure came from specific studies with particular conditions, later results have been more modest, and AI tutors are new and not yet shown to reach that bar. The promise is real; so is the need for evidence.

• "If the AI explained it clearly, I've learned it." Reading a clear explanation is not the same as being able to retrieve or apply it yourself.
• "AI answers are basically reliable." Models can state false facts and fake sources confidently; important claims need verifying.
• "Getting the answer fast is the goal." Fast answers can skip the productive struggle that builds durable skill.
• "AI tutoring is already proven to match human tutors." It is promising but new, and evidence at Bloom's level does not yet exist.

Pick one topic you are studying. Write a prompt that makes the AI quiz you, withhold answers until you attempt, and give feedback — not one that simply explains. Run a ten-minute session, then verify one factual claim it made against a trusted source. Note what the AI got wrong.

Think Like a Maester: An AI can carry the explaining and the questioning, but the understanding is built only by the effort you refuse to hand over.

1. How does prompting an AI to question you differ, for learning, from asking it to explain?
2. Why must you verify AI output, and what does "hallucination" mean here?
3. Name three legitimate roles AI can play in accelerating your learning.
4. What did Bloom's 2 sigma finding claim, and why should we be cautious applying it to AI tutors?
5. Explain how a fluent AI explanation can feel like learning while preventing it.

AI tools can be a remarkable accelerant: a Socratic tutor that quizzes you, an explainer that rephrases endlessly, a practice partner that supplies problems and feedback. The condition is that the effortful thinking stays yours — you attempt, retrieve, and verify rather than absorb. Because models can be confidently wrong, treat them as fallible partners and cross-check what matters. Bloom's 2 sigma problem shows why cheap one-to-one tutoring is such an exciting prospect, while reminding us that AI tutors are new and still unproven at that bar. Use them to learn faster, not to avoid learning.

• Learning in the Age of Information
• Thinking for Yourself in an AI World