Raw LLM Responses

What the talk gets right Schools reward outputs over process, so students seek shortcuts. Shocking… when you build a game about grades, people speedrun it. “Autopilot” risk is real. If you outsource thinking, your thinking atrophies. That’s not new; it happened with calculators, slide decks, and the first search engines too. We need “productive resistance” in UX: nudge people to think before they accept an answer. Agreed. That’s a design problem, not a ban-AI problem. Where it goes off the rails Confuses misuse with essence. A lazy prompt yielding a lazy answer isn’t a property of AI, it’s a property of laziness. Calls validation tone a “dark pattern,” as if being readable equals manipulation. Meanwhile lectures with monotone delivery somehow aren’t a dark pattern? Treats “free finals access” like moral catastrophe. If a tool helps with studying, the fix is clear rules and assessment redesign, not pearl-clutching about availability. Claims “no sources” while ignoring that source-citing and step-by-step modes exist — you just have to, you know, use them. Frames “one-on-one AI tutoring” as inherently sterilized. Good tutoring is dialogic. You can force that: ask, quiz, withhold answers, require justification. Reframe: AI isn’t a replacement, it’s a resistance band Unassisted → Assisted → Accountable. Use AI to generate a plan, show your work, then verify with sources or examples. The reps are yours; AI adds load and feedback. Personalization ≠ pampering. It’s scaffolding. A good tutor interrupts, asks, and makes you prove it. Drop-in PSA caption (use it under the video) AI won’t fix bad incentives — it exposes them. If your assessments reward regurgitation, students will offload it. If they reward reasoning, AI becomes a sparring partner, not a cheat code. Tools aren’t the problem. Rubrics are. Fast, evidence-based counters you can say out loud “Bad prompts get bad answers. That’s user behavior, not model essence.” “Show me your rubric. If it measures thinking, AI has to reason — or the student gets caught.” “We’ve had ‘first-result bias’ since search engines. The cure wasn’t banning search; it was teaching evaluation.” “Personalization isn’t a dark pattern; unearned certainty is. Fix UX with friction: ask, quiz, source.” “Autopilot is a teacher choice. Turn it off: require chain-of-thought summaries, citations, and oral defense.” “If finals can be done by a chatbot, the exam is the problem.” “Treat AI like lab equipment: safety rules, methods, and a lab notebook. No notebook, no credit.” Classroom policy kit (plug-and-play) Allowed: AI for brainstorming, outlining, drafting variants, code comments, examples. Required artifacts: prompt history, version deltas, 150-word self-critique of what changed your mind, 3 sources with one you disagree with. Assessments that survive 2025: oral defense (5 min), live re-prompting in class, “explain two wrong answers,” transfer tasks (apply the idea to a new domain). Integrity rule: If AI makes a factual claim, student must verify and attach receipts. No receipts, no points. “Productive resistance” you can actually implement Make the model ask 2 clarifiers before answering. Force “think-aloud mode”: bullet its reasoning, then generate the answer. Add “skeptic step”: AI must propose how it could be wrong and what would falsify it. Toggle “sources-first”: answer must cite or link to at least 2 high-quality references, then explain differences.