AIQ
AIQ
Part of the Smart vs. Intelligent lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 🔧 Builder (8–10) · 💻 Hacker (11–14)
"In 1997, Deep Blue beat Garry Kasparov at chess. In 2024, ChatGPT writes essays. Are these the same kind of 'intelligence'? Let's unpack what intelligence actually means in the context of AI systems."
Ask the class to commit to a position before any instruction: same kind of intelligence, or fundamentally different? Have a few students defend each side out loud. Most will intuit that something is different but won't have the vocabulary yet — that's the gap this lesson closes.
Frame the stakes plainly: nearly every headline about AI — whether it's "taking jobs," "becoming conscious," or "about to surpass humans" — hinges on what the word "intelligence" is doing in that sentence. This lesson gives you the technical grounding to evaluate those claims rather than react to them.
This is also a useful moment to name where the class is headed: for students weighing a university path or career in computing, "narrow AI vs. general intelligence" is not academic framing — it's the actual, current technical vocabulary used in AI research and industry to describe exactly this gap. Getting comfortable with it now pays off directly in any future coursework or reading on the subject.
Move through the three scenes as a technical comparison, using them as concrete anchors for the Deep Blue vs. LLM framing from the hook.
A calculator and a stopwatch execute deterministic, hard-coded algorithms: identical input maps to identical output every time, by design, with no exposure to data required or possible. A search engine's ranking, by contrast, is learned — shaped by aggregate patterns across billions of past queries and click behaviors, and it changes as those patterns change. A chatbot is the same category of system at greater scale: a model trained on enormous text corpora to predict likely continuations of a sequence of tokens. Neither the search ranker nor the chatbot was told the "right answer" by a programmer line by line — both derived their behavior from data.
Worth pressing on: "predicting likely next tokens" sounds almost too simple to produce fluent, context-aware text — ask students why that mechanism works as well as it does in practice. The honest, current answer is that at sufficient scale (model size and training data), statistical next-token prediction captures a remarkable amount of structure in language — grammar, factual association, even some reasoning-like patterns — without that structure being explicitly programmed. Why it generalizes as well as it does across so many tasks is still an active area of research, not a fully settled question, and it's fine to say so plainly.
Shazam's audio fingerprinting and a face filter's facial-landmark detection are both narrow, well-scoped learned models: trained on a large labeled dataset (songs; annotated faces) to generalize a specific pattern to new, unseen inputs. This generalization to unseen inputs is the actual technical test that separates "learned model" from "lookup table," and it's worth stating explicitly to this age band. A radio and a flashlight, by contrast, involve no learned model at all — a radio demodulates a broadcast signal, a flashlight is a closed circuit.
These two examples are also useful for illustrating scale differences within "AI" as a category: audio fingerprinting and facial-landmark detection are comparatively small, well-defined problems that can run entirely on a phone, in contrast with a modern large language model, which requires enormous compute and typically runs on remote servers. "AI" spans an enormous range of system size and complexity — worth flagging so students don't picture every AI system as LLM-scale.
Bridge from the previous scene: Shazam and a face filter are both narrow learned models, scoped to one well-defined pattern-recognition task. Deep Blue and modern LLMs sit on the same spectrum, just at a vastly larger scale of task complexity and compute — the category "narrow AI" covers everything from a phone-sized audio fingerprinting model to a datacenter-scale language model, which is worth students holding onto before assuming "narrow" means "simple."
This is where to connect directly to the hook. Deep Blue's 1997 win used brute-force minimax search with hand-tuned position-evaluation heuristics — closer to raw, exhaustive computation guided by human-written rules than to the "learns from data" definition used elsewhere in this lesson. Modern large language models are trained differently: they learn statistical patterns over massive text datasets via a transformer architecture and generate output by predicting likely next tokens. Both, however, remain narrow AI: each is exceptional at the specific task it was built or trained for, and neither has anything resembling general reasoning across arbitrary domains, embodied experience of the physical world, or (as far as any current scientific evidence shows) subjective/conscious experience. It's worth being explicit with students that "the gap between performing intelligently and being intelligent remains vast" is not a hedge — it reflects real, open disagreement even among AI researchers and philosophers about what would even count as evidence either way.
Read the app's summary together: "Intelligence in AI is task-specific optimization, not general understanding. Deep Blue used brute-force search trees; modern LLMs use transformer architectures trained on internet-scale text. Both are narrow AI — excelling at defined tasks while lacking the general reasoning, embodied cognition, and consciousness that characterize human intelligence." Ask students to restate, unprompted, why Deep Blue and a modern LLM are placed in the same category ("narrow AI") despite being built on completely different techniques thirty years apart.
Also worth naming directly: "embodied cognition" is doing real work in that summary sentence and is easy to skim past. It refers to the idea, taken seriously in cognitive science, that a good deal of human understanding is shaped by having a body that interacts with a physical world — touch, movement, spatial navigation — not just abstract symbol or pattern processing. No current AI system has anything resembling that kind of grounding, which is one reason "it processes patterns very well" and "it understands the world" are not the same claim.
These are genuinely open questions in the field, not questions with a hidden correct answer in the back of the book — treat them that way. This age group can tell the difference between a Socratic discussion and a lecture in disguise, and the discussion works better if you're willing to say "I don't know, what do you think" on the harder ones.
The Chinese Room question in particular tends to split the room, and that's healthy — some students will argue that sufficiently sophisticated symbol manipulation eventually becomes indistinguishable from (or arguably equivalent to) understanding, others that there's an unbridgeable gap. Both are defensible philosophical positions with serious proponents; don't let the discussion collapse into a false consensus just to wrap it up on time.
Review these as a group after individual completion. At this level, the wrong options are less "obviously silly" than in younger age bands and more "commonly and confidently believed" — it's worth explicitly asking who picked each wrong option and why, rather than moving straight to the correct answer.
Return to the opening question: "So — Deep Blue and ChatGPT, same kind of intelligence? You now have the vocabulary to say no, precisely: both are narrow AI, exceptional at the task they were built or trained for, and neither has general reasoning, embodied cognition, or anything we can currently show is subjective experience. The interesting question isn't whether today's AI is 'smart' — it's what specific task it was optimized for, and where that optimization breaks down."
Extension activity (20–25 minutes): Have students pick a real AI system in the news (a chatbot, an image generator, a self-driving system, a recommendation algorithm) and write a short analysis answering three questions: What specific task was it trained/built for? What evidence would distinguish "it's a very good pattern-matcher for this task" from "it truly understands this domain"? And where does it visibly fail in a way that reveals its narrowness? This is a good bridge into a broader class discussion — or a written assignment — on how to read AI hype critically, which is arguably the single most useful transferable skill in this unit.
For a class with more time or a stronger interest in computing as a field, extend this into a short research prompt: have students look up one concrete example of a narrow AI system failing outside its trained domain — a well-documented case, not a rumor — and present the technical reason for the failure in two or three sentences. This reinforces that "narrow" isn't a vague qualifier but a specific, testable property with real, checkable consequences.