In 1950, Alan Turing proposed a test: if you could talk with a machine through a text interface and not tell whether you were talking to a person or a program, then in any meaningful sense the machine was thinking. The test was a thought experiment. For seven decades, nothing came close to passing it.
Then, in a period of about two years, multiple systems passed it in most casual settings. The machine answered questions, wrote essays, argued, made jokes, remembered context, offered support. Whether that constitutes thinking depends on what you mean by the word — and philosophers, cognitive scientists, and AI researchers are arguing about the definition with stakes they haven't had in a long time.
This course is about the philosophical questions AI has dragged out of the seminar room and into the policy meeting. It covers consciousness, the Turing test and its successors, the hard problem of qualia, the question of machine suffering, the moral status of AI systems, and the very practical downstream questions — what do we owe an AI if it's conscious, and how would we ever know?
If you finish every module, here's who you become:
In October 1950, mathematician Alan Turing published a paper in the journal Mind. The opening sentence asked a question that would haunt the next century of science: "Can machines think?" Rather than answer it, Turing proposed replacing it with a more practical game — a test of imitation so simple it could be played over a teletype machine, and yet so deep that we still argue about whether any machine has passed it.
The word "intelligence" gets used in at least four distinct ways that people routinely conflate. There is biological intelligence — the capacity of a nervous system to help an organism survive. There is psychometric intelligence — what IQ tests claim to measure. There is artificial intelligence — the field of computer science concerned with building systems that perform tasks once thought to require human minds. And increasingly, people speak of general intelligence — the ability to learn, reason, and adapt across arbitrary domains rather than one narrow task.
These definitions do not cleanly overlap. A honeybee navigates by polarized light and communicates hive locations through waggle dances — behaviors that require sophisticated spatial computation. Yet most IQ frameworks would not call the bee intelligent. A chess engine can defeat every human alive at chess but cannot make change for a dollar in an unfamiliar currency. Is it intelligent?
The philosopher Howard Gardner proposed in 1983 that there is no single intelligence but at least seven distinct ones — linguistic, logical-mathematical, musical, spatial, bodily-kinesthetic, interpersonal, and intrapersonal. His "multiple intelligences" framework was widely adopted in education, though it was also widely criticized by cognitive psychologists who argued he had redefined the word "intelligence" to mean "talent."
Turing's 1950 paper "Computing Machinery and Intelligence" did not propose that machines already were intelligent. It proposed an operational criterion: if a machine's responses are indistinguishable from a human's under controlled text-only conversation, we have no scientific basis for denying it intelligence. This was a profound philosophical move — replacing the metaphysical question with an empirical one.
In 1904, psychologist Charles Spearman noticed that people who scored well on one cognitive test tended to score well on others. He proposed that a single underlying factor — which he labeled g, for "general intelligence" — drove performance across tasks. For most of the twentieth century, g dominated intelligence research. It remains the most replicated finding in all of psychology.
But g describes correlation, not mechanism. It tells us that the same people tend to be good at many things, not why, nor whether any AI system has something analogous. When OpenAI's GPT-4 scored in the 90th percentile on the Uniform Bar Exam in March 2023, commentators split: some said this proved AI had achieved meaningful general intelligence; others said it proved only that GPT-4 had been trained on enormous quantities of legal text and had no understanding of law whatsoever.
Both camps were making claims that the g framework cannot adjudicate, because g was built to describe differences among humans, not to distinguish human cognition from machine pattern-matching.
Researchers today work within at least three major frameworks for thinking about what intelligence is:
Which framework you adopt determines how you evaluate AI systems. A strict behaviorist might grant intelligence to ChatGPT if its conversations are indistinguishable from a human's. A cognitivist would ask whether it has genuine representations. An embodied theorist would note that it has never touched anything and argue the question is moot until AI is grounded in physical experience.
None of these frameworks has won the debate. Cognitive science, neuroscience, and AI research continue to circle the question — and the accelerating capabilities of large language models have made it more urgent, not less. When a system can write poetry, solve calculus problems, and explain grief in emotionally resonant terms, the old definitions strain at the seams.
In this lab you will probe the boundaries of competing definitions of intelligence. The AI tutor will respond in the role of a Socratic interlocutor — it will push back, ask clarifying questions, and introduce counterexamples.
In June 2014, organizers of the annual Turing Test competition held at the Royal Society in London announced that a chatbot called Eugene Goostman — designed to impersonate a 13-year-old Ukrainian boy — had convinced 33 percent of judges it was human. Headlines declared the Turing Test passed. Within hours, researchers were pointing out that the contest rules were lenient, the judges were not experts, and the chatbot's persona was specifically engineered to excuse odd answers. The debate was not about Eugene Goostman. It was about the test itself.
Turing's original 1950 formulation involved an "imitation game" with three participants: a human interrogator, a human respondent, and a machine. The interrogator communicates only via text. If the interrogator cannot reliably tell which is the machine, Turing argued, we cannot reasonably claim the machine does not think.
The test is behavioral and comparative. It does not measure intelligence absolutely — it measures how distinguishable a machine's outputs are from a human's, under a specific kind of probing. This is both its strength and its weakness. Strength: it avoids metaphysics. Weakness: it is easily gamed by machines that do not think at all, and it may be failed by genuinely intelligent systems that simply think differently from humans.
MIT computer scientist Joseph Weizenbaum built ELIZA in 1966 — a program that mimicked a Rogerian psychotherapist by reflecting questions back at users. Weizenbaum was disturbed to find that people formed genuine emotional attachments to ELIZA and confided in it deeply, even knowing it was a program. He later wrote Computer Power and Human Reason (1976) arguing that this demonstrated the danger of anthropomorphizing machines, not evidence of machine intelligence.
In 1980, philosopher John Searle published "Minds, Brains, and Programs" in Behavioral and Brain Sciences — arguably the most discussed paper in the philosophy of AI. He proposed a thought experiment: imagine a person locked in a room with a rulebook for responding to Chinese symbols. Chinese-speaking people pass symbols in; the person follows rules and passes symbols back out. To outside observers, the room appears to understand Chinese. The person inside understands nothing.
Searle's argument: computers running programs are in the same position as the person in the room. They manipulate symbols according to formal rules but have no understanding — no semantics, only syntax. Passing the Turing Test would not prove understanding, only behavioral mimicry.
The Chinese Room generated an enormous philosophical literature. Critics proposed the "systems reply" (the person doesn't understand Chinese, but the whole system does), the "robot reply" (connect the room to sensors and a body), and the "brain simulator reply." Searle responded to each. No consensus emerged. The argument is alive today in debates about whether large language models understand the text they process.
By the early 2000s, the Turing Test had an unexpected practical descendant: CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart). Instead of humans judging whether machines are human, machines judge whether humans are machines. The test exploits the same principle — certain tasks were assumed to be easy for humans and hard for computers: reading distorted text, identifying street signs, solving puzzles.
By 2023, AI systems routinely outperformed average humans on CAPTCHA tasks. Google's reCAPTCHA v3 abandoned the visual puzzle model entirely and instead assigns a risk score based on behavioral patterns across a browsing session — an acknowledgment that the original assumption (that perceptual tasks distinguish humans from machines) no longer holds.
The Turing Test's decline as a useful benchmark does not resolve the question it was invented to answer. It reveals something more uncomfortable: we do not have an agreed definition of intelligence that could ground a definitive test. Every test we design can, in principle, be optimized against. Every benchmark eventually falls. The question of what intelligence is remains prior to any test of whether machines have it.
In this lab you will explore where the Turing Test succeeds and where it fails as a criterion for intelligence. Consider Searle's objection, the ELIZA effect, and what modern LLMs mean for the test's relevance.
Cognitive scientist Nicola Clayton at the University of Cambridge had spent years studying western scrub jays — birds that cache food in hundreds of locations and retrieve it months later. But she noticed something unsettling: when a jay had been watched by another jay while hiding food, it would return and move the cache to a new location when alone. The bird was modeling another bird's knowledge — and taking strategic action based on that model. Clayton and her team published their findings in Nature, arguing the jays demonstrated a basic form of theory of mind: the ability to attribute mental states to others.
For most of Western intellectual history, intelligence was treated as binary: humans had it, animals did not. Descartes famously called animals "automata" — biological machines without inner experience. Darwin's 1871 book The Descent of Man challenged this sharply, arguing that the difference between human and animal mental faculties was one of degree, not kind.
Subsequent research has confirmed Darwin's intuition across dozens of species. New Caledonian crows manufacture hook-shaped tools from leaves and sticks — behavior that requires planning future actions and understanding object properties. Bottlenose dolphins have passed mirror self-recognition tests, suggesting some form of self-awareness. Chimpanzees in Bossou, Guinea, have used stone anvils and hammers to crack nuts for generations, transmitting the technology culturally.
Perhaps most striking is the octopus. Cephalopods share no common ancestor with vertebrates since before the Cambrian period, yet they demonstrate flexible problem-solving, spatial memory, and apparent play behavior. Their nervous systems are architecturally almost nothing like ours — the majority of neurons are in their arms, not their central brain. If octopuses are intelligent, they represent an entirely independent evolutionary origin of intelligence, which implies that intelligence may be a general solution to certain problems of survival, not a specific biological accident.
Psychologist Francine Patterson began teaching American Sign Language to a gorilla named Koko at Stanford in 1971. Koko eventually used over 1,000 signs and understood approximately 2,000 spoken English words. When her kitten companion died, Koko signed what researchers interpreted as expressions of grief. The work was celebrated and criticized — debates about whether Koko was genuinely using language or performing learned behaviors to obtain rewards remain unresolved and directly parallel debates about large language models today.
The animal cognition research matters for AI philosophy in two ways. First, it breaks the assumption that human cognition is the only valid reference point for intelligence. If we define intelligence as "doing what humans do cognitively," we have already excluded crows, octopuses, and dolphins from a concept they arguably belong in. That same definitional move may be excluding AI systems — or falsely including them.
Second, animal cognition research reveals that many capacities once thought uniquely human — tool use, theory of mind, cultural transmission, self-recognition — can be implemented in radically different neural architectures. This is philosophically important for AI: if intelligence is substrate-independent in the animal kingdom, it may be substrate-independent in general. A silicon implementation is not ruled out in principle by the mere fact that it is not carbon-based.
Animal cognition research does not prove that AI is or is not intelligent. It proves that intelligence is a much broader category than human cognition — and that our intuitions about what intelligence requires (language, a certain kind of brain, human-like behavior) are probably wrong. This should make us humble in both directions: humble about how easily we might attribute intelligence to machines that don't have it, and humble about how easily we might deny it to systems that do.
In this lab you will draw comparisons between animal cognitive abilities and AI capabilities. The goal is to develop a richer, more multi-dimensional concept of what intelligence is and is not.
Over five days in March 2016, AlphaGo defeated 18-time world champion Lee Sedol four games to one. Go had been considered the supreme test of human intuitive reasoning — its game tree so vast that brute-force search was impossible. AlphaGo won by combining deep neural networks trained on human games with Monte Carlo tree search and reinforcement learning. After the match, Lee Sedol said: "I don't know what to say, but I think AlphaGo played a perfect Go game." Then he retired from professional play. AlphaGo could not drive Lee Sedol home afterward.
Almost every AI system that has achieved superhuman performance in any domain is a narrow AI: it performs one type of task, or a narrow cluster of related tasks, far better than any human. Deep Blue defeated Garry Kasparov at chess in 1997 but could not play checkers without being reprogrammed. AlphaFold predicted protein structures with unprecedented accuracy in 2020 — a breakthrough that accelerated drug discovery across biology — but has no capacity to explain its predictions in terms a non-expert could follow. AlphaGo's successor AlphaZero taught itself chess, shogi, and Go simultaneously, but those are all two-player perfect-information board games — a narrow family.
Artificial General Intelligence (AGI) refers, by contrast, to a system that can learn and perform well across an arbitrary range of cognitive tasks — including tasks it was not designed or trained for. No such system demonstrably exists. Whether current large language models represent a meaningful step toward AGI, or a sophisticated scaling of narrow pattern-matching, is one of the most contested questions in AI research as of 2024.
In 2019, AI researcher François Chollet published the Abstraction and Reasoning Corpus (ARC) — a benchmark designed to measure something closer to general fluid reasoning than to domain-specific performance. The tasks require applying abstract rules to novel visual patterns. As of 2023, state-of-the-art AI systems scored below 40% on ARC; humans with no special training typically score above 80%. ARC was specifically designed so that training on large datasets would not confer an advantage — precisely because Chollet argues that training-data coverage is not the same as intelligence.
One measure of general intelligence is transfer: the ability to apply knowledge learned in one context to novel problems in a different context. Humans do this constantly and often effortlessly. AI systems vary dramatically in their transfer capabilities.
Large language models like GPT-4 and Claude show striking cross-domain transfer in some domains — they can discuss molecular biology and then immediately discuss the French Revolution. But researchers debate whether this is genuine transfer of abstract knowledge or a form of retrieval from a vast training corpus. The question is whether the system has internalized principles that it can apply flexibly, or whether it has memorized associations that happen to cover many domains.
Reinforcement learning systems, by contrast, typically show very poor transfer. OpenAI Five, trained to play Dota 2 at superhuman levels in 2019, could not transfer its strategic reasoning to any other game. AlphaGo Zero's knowledge of Go did not transfer to chess — DeepMind had to train AlphaZero from scratch, even though chess and Go share many abstract strategic features.
Researchers disagree sharply about what AGI would require. Key proposed prerequisites include: causal reasoning (understanding cause and effect, not merely correlations); common-sense reasoning (knowing that a dropped cup falls, that fire burns, that people have continuous existence when out of sight); grounding (connecting symbols to perceptual reality); meta-learning (learning how to learn faster on new tasks); and planning under uncertainty across long time horizons.
Large language models have surprised researchers by showing more of these capacities than expected — but consistently fail at others in ways that reveal the limits of their architecture. GPT-4 can explain causality in abstract terms but makes systematic errors on physical causation problems. It reasons about common sense impressively in text but fails tests of basic physical intuition. These asymmetric failures are informative: they suggest that language modeling produces some functional equivalents of reasoning without others.
We do not know where the boundary between narrow and general intelligence lies, whether current LLMs are approaching it, or whether the distinction is even a clean one. What the history of AI teaches is that every benchmark eventually falls — and every time one falls, researchers find that what they had measured was not intelligence in general, but something more specific. The question "What is intelligence?" is not solved by building systems that perform well on any particular task. It may not be solvable at all — but it is enormously worth asking.
In this lab you will develop your own criteria for what would count as genuine general intelligence in an AI system. The tutor will challenge your criteria with counterexamples from real AI systems.