AIQ
AIQ
Part of the Who's Responsible? lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · ⚡ Architect (15–18)
Hacker mode students can handle real complexity and appreciate being told a problem is genuinely unsolved, rather than having it oversimplified for them. Lean into that.
Say: "When AI causes harm, who is responsible? The programmer? The company? The user? Let's examine the accountability gap and emerging regulatory responses." Use the phrase "accountability gap" deliberately and write it on the board — it's the real term for the problem this lesson is about, and naming it precisely signals that this is a serious question, not a rhetorical one.
Open with a framing question rather than an answer: "When a human driver crashes a car, we know how to assign blame — we look at what that one person chose to do. What breaks about that process when there's no human driver at all, just code?" Let a few students take a swing at it. Likely answers will circle around "there are more people involved" or "you can't ask the code why it did that" — both are exactly right, and you can tell them so before moving into the activity.
A good way to sharpen the warm-up further: ask, "A single self-driving car's software might run in thousands of identical cars at once. If a human driver makes one bad decision, that's one mistake by one person. If that same flaw exists in the AI's code, what's different about the scale of the problem?" Students usually land on "it could cause the same mistake everywhere, all at once" — a genuinely important point that also explains why regulators treat software-driven systems differently from individual human error: the same defect can replicate across every deployment simultaneously, in a way one distracted driver's mistake never could.
Have students read through the three scenes individually or in pairs, then run the debrief as a structured discussion that goes one level deeper than the surface facts — the goal at this age is understanding the mechanisms behind each example, not just recalling that they exist.
Scene 1 — The Blame Problem. The four cases: a self-driving car crash, an AI loan denial, the "scales of justice" framing, and an AI misdiagnosis a doctor acted on. Push past "who's at fault" into the structural reason it's hard: harm from an AI system is rarely the result of one decision by one actor — it's the downstream effect of a chain of decisions (what data the model was trained on, what the company chose to test for before deployment, how the deploying organization configured it, how the end user behaved) made by different people at different times, none of whom necessarily foresaw this specific outcome. That mismatch — between how AI harm actually arises and how liability law is built to assign blame to one foreseeable actor — is the "accountability gap" from the hook.
Scene 2 — Transparency. Define the black box problem precisely: many AI systems, especially large ones, make predictions by combining millions of learned internal weightings in ways that don't reduce to a simple causal story, even for their own engineers. Introduce Explainable AI (XAI) as the research response: rather than reading the model's "true" internal reasoning (which usually isn't a coherent narrative to begin with), XAI techniques approximate an explanation after the fact — for example, estimating which input factors most influenced a specific output. Be precise with students about the limits here: an XAI explanation is a useful approximation of the model's behavior, not a literal transcript of its reasoning, and that distinction matters when a company claims its AI is "explainable."
Scene 3 — Rules & Laws. This is where to introduce the EU AI Act's actual structure, since the base lesson only names it: it classifies AI systems into risk tiers — unacceptable risk (a short list of practices banned outright, such as government social-scoring systems), high-risk (systems used in areas like hiring, credit, and medical devices, which must meet documentation, testing, and human-oversight requirements before deployment), limited risk (systems like chatbots that must disclose they're AI), and minimal risk (no special obligation). Contrast this with the US's historically lighter, sector-specific approach — relying on existing regulators like the FTC or FDA plus non-binding executive guidance rather than one comprehensive law — and UNESCO's 2021 Recommendation on the Ethics of AI, adopted by all 193 member states as shared principles, but not enforceable law in any of them.
Introduce two governance concepts the base quiz will test: an algorithmic impact assessment is a formal review, done before a system is deployed, documenting its likely harms, fairness implications, and societal effects — some governments already require these for their own automated systems. The distributed responsibility problem is the more precise name for what Scene 1 illustrated: accountability becomes genuinely unclear when harm results from the interaction of multiple actors' decisions, none of which alone would have caused the outcome.
A good stretch question for a strong class: "If an algorithmic impact assessment is done honestly and still says 'this system might cause harm,' but a company deploys it anyway because the assessment isn't legally binding — has the accountability gap actually been closed, or just documented?" There's no clean answer; the point is noticing that documentation and enforcement are two different things.
It's worth pausing on one more subtlety before moving to discussion: "explainable" and "accurate" are not the same property, and a system can have either without the other. A very simple model — one with only a handful of clear, human-readable rules — is usually easy to explain but may be less accurate than a complex one. A large, high-performing model is often much harder to explain precisely because its accuracy comes from combining an enormous number of small factors in ways that don't reduce to a short human story. This tradeoff is a real, ongoing tension in the field: regulators asking for "explainable AI" in high-stakes settings are sometimes implicitly asking companies to accept a less accurate model in exchange for one that's easier to audit and justify — and reasonable people disagree about when that trade is worth making.
This quiz checks whether students can use the vocabulary from the activity correctly, not just recognize it — if a student gets a question right without being able to explain the term in their own words, treat that as a cue to revisit the relevant scene.
Close by returning to the opening framing: "We started by asking what breaks when there's no single human driver to blame. What breaks is the whole assumption our usual idea of 'fault' relies on — one person, one choice, one consequence. AI accountability is hard precisely because that assumption doesn't hold anymore, and the tools we looked at today — explainability research, risk-based laws, impact assessments — are all attempts to rebuild accountability for a world where it doesn't."
Extension activity (20–25 minutes): In small groups, have students design a simple algorithmic impact assessment for a hypothetical AI system — for example, an AI that decides which students get flagged for extra tutoring support at their school. Groups should identify: who could be harmed if the system is wrong, what data it might be trained on and how that data could be biased, and what human oversight step should exist before the system's recommendation actually affects a student. Have each group present their assessment and let the class question it — this mirrors, in miniature, the real process regulators are asking companies to go through.
A shorter alternative: assign each group one real-world AI accountability case (a self-driving car incident, an AI hiring tool later found to be biased, an AI content-moderation error) to research briefly and present as a "who's responsible" case file, naming every party involved and where they think the accountability gap showed up.
If your class has extra time and enjoys debate, a good closer is a quick two-minute "gut check" vote before the extension activity begins: ask students to stand on one side of the room if they think a strict, EU-style risk-based law is the better model for other countries to copy, and the other side if they think a lighter, US-style sector-specific approach is better. Have a representative from each side give one sentence defending their position, then let students move if they've changed their mind. This works well as a quick temperature-check before students dig into the deeper case-file or impact-assessment work above, since it surfaces their starting assumptions before the research complicates them.