Guardrails: Age, Rules, and Being Wrong Confidently

The honest summary of the research is not 'AI is good for learning' or 'AI is bad for learning'. It's this: whether AI helps or harms depends on whether you delegate or engage. There's real evidence on both sides. A meta-analysis of thirty-five experiments found a moderately positive effect on learning outcomes, and a Harvard trial found a well-designed AI tutor produced better learning in less time than in-class active learning. But a 2025 survey of 666 people found frequent AI use correlated with weaker critical thinking, and the mechanism was cognitive offloading — letting the tool do the thinking. Younger users were hit hardest. Better outputs, worse thinking. So the single rule this whole class rests on is: the AI may question you, hint at things, and check your work, but you do the thinking.

AI can be wrong. Not obviously wrong — fluently, confidently, plausibly wrong, in the same tone it uses when it's right. The advice 'always verify important facts' is correct but carries a trap: the person learning a topic is exactly the person least equipped to spot a confident error in it. If you could reliably tell, you wouldn't need to be studying it. This bites hardest in one specific place, which is letting AI write practice questions and mark your answers, because a wrong answer key delivered confidently teaches you a misconception and then rewards you for repeating it. So: AI-generated answer keys get checked against your textbook, your notes or your teacher, every time. The safe version is to let AI ask the questions and mark them yourself against a source you trust.

Wrong answers come in three recognisable shapes. A REAL-SOUNDING SPECIFIC: exact dates, precise percentages, named studies — invented but entirely plausible; check the number against your source rather than against your sense of whether it sounds right. A CORRECT METHOD WITH A WRONG STEP: the working looks right end to end and one line has a sign error or a swap; redo the arithmetic yourself, computing rather than scanning. And a RIGHT ANSWER WITH A WRONG REASON: the conclusion matches your book but the explanation underneath doesn't, which is the most dangerous kind because the thing you checked was correct. The way to build this skill is to practise it on a topic you already know well — catching AI being wrong about something familiar is what builds the habit for topics where you can't tell.

Four rules that aren't optional. AGE AND CONSENT: ChatGPT's minimum age is thirteen, and between thirteen and eighteen you need a parent's permission; under thirteen, an adult runs the session. That's the tool's own rule, not ours. YOUR SCHOOL'S POLICY: many NSW schools forbid generative AI on schoolwork unless a teacher directs it, so ask what's allowed before you use it, not after. WHAT YOU UPLOAD: uploading your work sends it to a commercial service, so never upload other people's work, personal details, or anything marked confidential. And ASSESSMENT IS DIFFERENT: study help and assessment help are not the same thing — if it's being marked, the default answer is no unless your teacher has said otherwise. When you're unsure, ask first.

Remember that every technique in the FAT Teacher framework works with a pen. Free recall on a blank page, writing your own questions, spaced review and the three piles, the error log, Feynman on paper, teaching a real person — none of those need AI, and the last one is better without it. AI is an accessory to the framework, never a replacement for any part of it. And there are three moments where you should close the laptop: when you're stuck but haven't genuinely tried for fifteen minutes, when you're about to ask for an answer rather than a hint, and when you're tired and just want the work over with. Those are the three moments where delegation wins, and recognising them is most of the discipline this class is asking for.