Can Machines Think?

This course is for anyone who wants a rigorous, non-mystical way to think about claims like “this AI truly understands” or “this model is becoming conscious” — whether you encounter them in marketing copy, a philosophy seminar, or a late-night argument with a friend. No philosophy background is assumed; every argument is explained from scratch, with its original source.

Unlike most of this Academy, this course is deliberately built on decades-old, canonical sources rather than this year's headlines — the arguments don't expire when the next model ships.

In 1950, mathematician Alan Turing published “Computing Machinery and Intelligence” in the journal Mind. He opens by proposing to replace the question “can machines think?” — which he considered too ambiguous to answer — with a concrete alternative: the “imitation game.” A human judge holds text conversations with a human and a machine, without seeing either. If the judge can't reliably tell which is which, the machine passes.

The common misreading is that Turing was proposing a definition of thought: pass the test, and you've proven the machine thinks. Turing's own framing was more modest and more interesting — he was pointing out that convincing behavior is the only evidence we ever have for any mind, human or machine. You can't directly inspect another person's inner experience either; you infer it from what they say and do. The test asks whether machines can clear the same behavioral bar we already use for each other.

Thirty years later, philosopher John Searle published “Minds, Brains, and Programs” (1980), introducing the most-discussed thought experiment in the philosophy of AI. Full details are well documented in the Stanford Encyclopedia of Philosophy's Chinese Room entry. The setup: imagine Searle, who speaks no Chinese, locked in a room with a rulebook (in English) that tells him exactly how to manipulate incoming Chinese symbols to produce outgoing Chinese symbols. Follow the rulebook well enough, and the room produces answers indistinguishable from a fluent Chinese speaker — passing a Chinese-language Turing Test.

Searle's point: he still understands zero Chinese. He's doing pure syntax — symbol manipulation according to rules — with no access to semantics, to what the symbols mean. If the whole room (Searle plus rulebook) can pass the test while nobody and nothing inside understands Chinese, then passing a behavioral test cannot, by itself, prove understanding. Searle aimed this specifically at “strong AI” — the claim that running the right program is sufficient for a mind, not at AI capability in general.

In 1995, philosopher David Chalmers published “Facing Up to the Problem of Consciousness”, drawing a distinction that reframes the whole debate. The “easy problems” of consciousness — explaining how a system discriminates stimuli, integrates information, reports internal states, or focuses attention — are hard engineering problems, but ones cognitive science can in principle solve by describing mechanisms and functions.

The “hard problem” is different in kind: why is any of that accompanied by subjective experience at all? Why does processing red-wavelength light produce the felt quality of seeing red, rather than happening “in the dark,” with no experience attached? Chalmers argues that fully explaining the mechanism — the easy problems — doesn't automatically explain why there's something it's like to be that mechanism.

In 2020, computational linguists Emily Bender and Alexander Koller published “Climbing towards NLU”, updating the Chinese Room for the language-model era with the “Octopus Test.” Two people, stranded on separate islands, communicate over an undersea cable. An octopus secretly taps the cable — seeing only the signal patterns, never the world those signals describe. Bender and Koller argue the octopus could, in principle, learn to produce plausible-sounding responses purely from patterns in the signal, without ever grasping what any of it refers to.

Their core claim: a system trained only on the form of language (text, as LLMs are) has no built-in way to learn meaning — the connection between words and the world — from form alone. It's a technically specific, empirically grounded descendant of the Chinese Room, aimed directly at whether today's text-only language models can be said to understand in the fullest sense, independent of how fluent their output is.

Where does this leave large language models? Genuinely contested — Bender and Koller's critics argue that grounding can be approximated through the sheer scale and diversity of training data and multimodal input (text plus images, audio, video), while defenders maintain the form/meaning gap persists regardless of scale. That live disagreement, not a settled answer, is the actual state of the field — treat confident claims on either side with the same scrutiny you'd apply to any other unresolved philosophical debate.

Four modules of philosophy converge on one practical habit: separate three claims that get routinely collapsed into one.

1. Acts intelligently — produces outputs that look like the product of understanding, reasoning, or skill. Testable today, and current AI clearly clears this bar on many tasks.
2. Understands — has genuine semantic grasp of what symbols refer to, not just facility manipulating them (the Chinese Room / Octopus Test question). Contested, and possibly untestable by behavior alone.
3. Is conscious / has subjective experience — there is “something it is like” to be the system (the hard problem). Not resolvable by any test we currently have, for AI or arguably for anyone but yourself.

Most headline claims — “this AI truly understands you,” “this model shows signs of consciousness” — quietly use evidence for claim 1 to assert claim 2 or claim 3. Use the template below to pull these apart the next time you read one.

I'm evaluating this claim about an AI system: [paste the claim or headline here]

Walk me through it against this framework:
1. What specific evidence is being cited — a benchmark score, a demo, a self-report from the AI itself, or something else?
2. Is that evidence actually about the system acting intelligently (producing good outputs), or is it being used to support a stronger claim about genuine understanding or subjective experience?
3. If it's a claim about "understanding": does the evidence rule out the Chinese Room / Octopus Test objection (that fluent output can come from form manipulation without meaning)?
4. If it's a claim about "consciousness" or "experience": what test, even in principle, could ever settle this either way — and does the claim acknowledge that limit?

Give me a one-paragraph honest assessment of which of the three claims (acts intelligently / understands / is conscious) this evidence actually supports.