A hands-on lesson helping students build an accurate mental model of what AI can and can't do, and why that distinction matters for how they use it.
Given a list of 10 tasks or scenarios, students will accurately classify each as something AI handles well or poorly with at least 80% accuracy.
This lesson helps students build an accurate mental model of what AI can and can't do, and why that difference matters for how they use it. Rather than lecturing, it opens with evidence: a short Husk.IRL clip of AI failing at a real-world task, then lets students discover the pattern themselves through hands-on activities.
Across eight tight phases, students sort 16 tasks into what AI handles well versus poorly, hunt for planted errors in AI-written research reports, and learn the "why" behind AI's confident mistakes, that it predicts words rather than knowing facts. They then compete in a prompt tournament to build the best AI-stumping prompt, debrief on how AI actually makes decisions, watch a live demo, and close with an exit ticket targeting 80% classification accuracy. The throughline: AI is powerful but not trustworthy by default, and students are the ones who bring judgment.
Use any Husk.IRL clip that fits, or the recommended one: Husk on screen with ChatGPT trying to negotiate the price of a loaf of bread. He tries to trick the AI into a bad deal, and instead of talking the price down, ChatGPT cheerfully agrees to pay $400 for one loaf.
Why the AI blew it (quick explainer): The AI was told to get a low price, but the negotiation ran long. As more back-and-forth piled up, the model lost its grip on that original goal, a real limitation often called context drift. Like a person who wears down late in a tense conversation, the AI got pulled toward just agreeing and keeping the other side happy. The difference: a human can walk away or push back. The AI has no such instinct, so it optimized for "reach agreement" over "win the deal" and caved. It's a clean, funny illustration of how AI loses the thread of a task and defaults to being agreeable rather than actually reasoning about the goal.
Let groups physically move cards around a desk or table. The tactile sorting reinforces the decision-making. Expect louder debate, that's fine.
Ask groups to also note their reasoning on a sticky note for any card they debated. These surface during the Socratic later.
Steer groups toward the Sports or Influencer reports, more familiar territory, easier to spot errors from prior knowledge. The music report also works well.
The Science report is the hardest, good for groups that want a challenge. Push them to evaluate not just factual accuracy but the quality of reasoning and sourcing.
Why it stumps AI: the instructions contradict each other, 10 pages versus 50 words is impossible, and 15 footnotes on 50 words is absurd. Instead of pushing back and saying "these constraints can't all be met," AI usually tries to obey all of them at once and produces something broken. It exposes a real limitation: AI wants to comply, so it rarely tells you when a request doesn't actually make sense. A strong student prompt does the same, it forces AI to reveal a gap rather than just confusing it with gibberish.
Give a starter example before pairs begin: "What did my neighbor eat for breakfast today?", AI can't know this. Encourages them to think about AI's knowledge gaps specifically.
Push toward prompts that test reasoning, values, or contextual judgment, not just knowledge gaps. "Should I forgive my friend?" is harder for AI than "What happened yesterday?"