› Grades 9–12 · ages 14–18 · Socratic seminar
Can a machine understand anything?
“A large language model understands the text it produces.”
A Socratic seminar on the oldest AI question with the newest evidence. Students work from short definitions of intelligence, consciousness, and stochastic prediction, and are pushed to say what 'understand' would even mean — and whether the answer changes anything about how we should use the tools.
60 minutes · one class period · free to photocopy
Students will be able to
- Explain next-token prediction and why it can produce outputs that look like reasoning.
- Distinguish behavioral, functional, and phenomenal claims about understanding.
- Sustain a seminar that cites the text, builds on others, and tolerates an open question.
Vocabulary
- Large language model (LLM) — The kind of model behind most chatbots — a very advanced next-word guesser.
- Stochastic / nondeterministic — The same prompt can produce different answers — randomness is built in.
- Intelligence (what we mean) — Performance on tasks vs. something deeper — the word hides disagreement.
- Consciousness (in machines) — Whether a system can “experience” anything — mostly unsettled and contentious.
- AGI (artificial general intelligence) — Hypothetical AI that matches human breadth across most cognitive tasks.
- Benchmark — Standardized tests used to compare models — useful and gameable.
Agenda · 60 minutes
| Minutes | Segment | What happens |
|---|---|---|
| 0–6 | Warm-up | Memorized fact vs. understanding. Build three test questions. |
| 6–18 | Read the text | Background sheet + six vocabulary cards. Students mark: one claim they accept, one they doubt, one question. |
| 18–46 | Socratic seminar | Circle. Teacher opens, then withdraws. Teacher's only interventions: naming which of the three questions (behavioral / functional / phenomenal) a comment answers, and inviting silent students. |
| 46–54 | Turn | 'If you're right about understanding, what changes about how you'd use the tool?' Five minutes; every student writes one sentence before anyone speaks. |
| 54–60 | Written position | Exit ticket paragraph. |
Perspective — yes, in every sense that matters
Understanding is what understanding does. If it answers, explains, generalizes, and corrects itself, demanding more is demanding magic.
- We attribute understanding to each other by behavior. We have nothing else.
- To predict the next word across all of human text, you have to model what the text is about.
- Interpretability research finds internal structures that track concepts, not just word statistics.
Perspective — no, and the fluency is the trap
Prediction is not comprehension. The system has no world, no stakes, no one home — and its fluency makes us forget that.
- It fails at trivial things a child gets, which shows the 'understanding' is pattern-shaped, not concept-shaped.
- The Chinese Room: perfect symbol manipulation, zero meaning.
- Nothing is at stake for the model. Understanding without caring isn't understanding, it's lookup.
Discussion questions
- 1.openingWhat would a machine have to do for you to say it understands?
- 2.coreAre you making a claim about behavior, internal structure, or experience? Would the other two change your answer?
- 3.coreThe model passed the bar exam and then invented a case that doesn't exist. Which fact matters more?
- 4.coreIf we can't test for experience in a machine, can we test for it in each other?
- 5.closingSuppose you're right. What changes about how much you trust the tool tomorrow?
How socratic seminar runs
Students sit in a circle with a short text or definition in hand. The teacher asks an opening question and then mostly stays quiet; students talk to each other, cite the text, build on and challenge one another. No hands — a speaker finishes, then the next speaks. The goal is understanding, not winning.
In grades 9–12
Stakes are real: college, jobs, creative work, civil rights. Students can steelman, cross-examine, and write policy.
Preview the full pack
Cover, teacher guide, and every student handout. Print teacher pages once and student pages one per learner.
On a phone, open the PDF in a new tab — in-page previews are unreliable in mobile Safari.
Standards
- CCSS ELA SL.9-10.1 / SL.11-12.1 (seminar)
- CCSS ELA RI.11-12.7 (integrate and evaluate multiple sources)
- CCSS ELA W.11-12.1 (argument)
- AP Seminar / philosophy elective alignment: argument analysis
More for grades 9–12
Taught it? Tell us how it went: send classroom notes · AI concepts field guide · Try the Activities Pass free · Commission a full unit