Can a machine that passes the bar exam understand justice?

Large language models are pattern-completion engines. Trained on vast bodies of human-generated text, code, and imagery, they learn statistical relationships between tokens — words, pixels, notes — and become extraordinarily good at predicting what comes next. What emerges is genuinely astonishing. It is also worth being precise about what it is.

When GPT-4 passes the bar exam, it is not reasoning about justice. It has learned, with uncanny fidelity, the patterns that characterize correct legal reasoning in text form. That distinction feels academic — until the case involves a novel jurisdiction, an unprecedented fact pattern, an ethically ambiguous framing. Then the pattern-matching breaks down in ways a genuinely reasoning human lawyer would catch. Not a failure. A clarification.

General artificial intelligence — a system that could learn any task a human can learn, generalize to genuinely novel situations, and do so with flexible, embodied, emotionally-inflected cognition — does not yet exist. Whether it will, and when, is one of the most contested questions in science today. Serious researchers disagree by decades. Some question whether the current trajectory leads there at all. What exists now is a collection of remarkably capable narrow AI systems that have grown wide enough to feel general, without actually being so.

The gap matters. It is where the human lives.

IQ was invented in 1905 to identify French schoolchildren who needed extra support. Alfred Binet, who designed it, explicitly warned against treating it as a measure of innate, fixed intellectual capacity. He was ignored. By the mid-twentieth century, IQ had become a comprehensive index of human cognitive worth — used to sort immigrants, justify eugenics, stratify education systems across the world.

What IQ tests actually measure is a real but narrow cluster of abilities: logical reasoning, working memory, processing speed, verbal comprehension. These abilities cluster together. They correlate with certain life outcomes. That is established. What is fiercely debated is whether they constitute "intelligence" in any meaningful sense — or whether they simply reflect the cognitive skills the industrialized world found most economically useful and built its institutions to reward.

Howard Gardner's theory of multiple intelligences — musical, spatial, bodily-kinesthetic, interpersonal, naturalistic — has been criticized by mainstream psychologists for lacking rigorous empirical support. That criticism has merit. But the question it raises is legitimate regardless: we designed our measurements around a particular slice of cognition, called that slice "intelligence," and built entire civilizations of educational sorting on that foundation.

AI now aces that slice. The question is whether we mistook the map for the territory.

Robert Sternberg's triarchic theory proposed that meaningful intelligence includes practical intelligence — navigating real-world situations — and creative intelligence — generating genuinely novel ideas — not just the analytical component tests capture. Neither is trivially automatable. Both are deeply entangled with embodiment, desire, social context, and lived experience. Not one standardized test has ever been designed around them.

What is the master carpenter thinking when the joint fits perfectly?

There is a tradition in cognitive science — associated with Francisco Varela, Evan Thompson, and Eleanor Rosch — that argues intelligence isn't primarily something that happens in the skull. Embodied cognition holds that thinking is inseparable from having a body that moves through a physical world. From the felt sense of being a creature with needs and vulnerabilities and history. On this view, much of what we call "intelligence" is not computation. It is a sensorimotor negotiation with reality that never stops.

A master carpenter doesn't think her way through a dovetail joint. A jazz musician improvising in real time isn't running algorithms. An experienced emergency room nurse who senses something is wrong before the monitors show it — that is not pattern-matching in the statistical sense. It is something happening in the whole organism, shaped by years of embodied practice, calibrated against stakes that actually mattered, informed by a nervous system that had skin in the game.

This is not mysticism. It is neuroscience.

Research on expert intuition by Gary Klein and Daniel Kahneman established that experts in high-stakes, dynamic domains develop pattern recognition that operates through the body and through emotion as much as through explicit reasoning. Klein calls it recognition-primed decision making. It is fast, holistic, and deeply contextual. It is also the product of genuine experience: the surgeon who has felt the resistance of tissue, the teacher who has read a room of thirty children at once, the negotiator who watched a deal collapse and felt the weight of what that meant.

AI systems don't have bodies. They don't have stakes. They have never been afraid, never been exhausted, never been surprised by their own grief. Whether this matters philosophically — whether experience requires embodiment — remains genuinely open in consciousness studies. Practically, it appears to matter enormously. The kinds of intelligence most resistant to automation are precisely those most entangled with the fact of being a creature in the world.

An AI painting won an art competition in 2022. Judges didn't know. Does that settle the question?

Combinatorial creativity — taking existing elements and recombining them in novel ways — is something AI does impressively. Almost all human creativity involves this to some degree. The history of art is the history of influence, recombination, response, and variation. If that is all creativity is, AI is creative. The argument is over.

But there is another kind of creativity that resists naming. The kind that emerges from necessity. From suffering. From the specific pressure of a particular life lived by a particular body in a particular place and time.

Frida Kahlo didn't paint her broken spine because it was a statistically interesting subject. James Baldwin didn't write about race and sexuality because his training data suggested it. Beethoven composed the Ninth Symphony while deaf — not despite his deafness, many argue, but in some deep way because of it. The work was wrung from experience in a way that made it irreducible.

This is admittedly speculative territory. We cannot peer into the causal chain between Beethoven's deafness and the Ninth and prove the connection is essential rather than incidental. But there is something philosophically serious in the intuition that work born from necessity has a different quality than work generated from pattern. Whether that difference is aesthetic, ethical, or merely sentimental is genuinely contested. It may not resolve cleanly.

What is less contested: creative intentionality and creative constraint are deeply intertwined in human making. Artists and scientists choose — what to pursue, what to abandon, what matters enough to spend a decade on. AI systems don't choose in that sense. They respond to prompts. The prompt is someone else's intentionality, outsourced. Whether that constitutes a permanent limitation or a temporary architectural one is a question the field is actively wrestling with. The wrestling has not produced an answer.

Schools, in their modern form, were largely designed to solve an industrial problem: how to produce large numbers of people with standardized, certifiable skills. Memorization, procedural fluency, and correct answer production weren't just teaching tools. They were the product. In a world where information was scarce and expertise costly to verify, being the person who had the information in your head was genuinely valuable.

That world is gone. Information is not scarce. Looking things up is not a concession — it is a cognitive prosthetic so ubiquitous that the distinction between "knowing" and "knowing where to find" has effectively collapsed. AI can now not only find the information but synthesize it, contextualize it, and present it in whatever form you need. The industrial-era case for rote learning has structurally evaporated.

What remains — what was perhaps always the deeper purpose, beneath the industrial scaffolding — is harder to test and harder to fake. Epistemic virtue: the disposition to be genuinely curious, to tolerate uncertainty, to revise beliefs when evidence warrants. Metacognition: the ability to think about your own thinking, notice when you're confused, calibrate your confidence against your actual competence. Judgment: the capacity to act wisely under genuine ambiguity, where no algorithm produces the right answer because the right answer is not a fact but a value-laden choice.

None of these are entirely unteachable. They are, however, largely untested — because they resist the clean metrics that bureaucratic institutions require.

John Dewey argued, over a century ago, that education's purpose was not information transmission but the cultivation of the capacity for democratic experience: the ability to participate meaningfully in collective life, to understand and feel the perspectives of others, to engage in the ongoing negotiation of shared meaning. Dewey's vision was largely defeated by efficiency pressures. AI may inadvertently revive it, simply by making the alternative so obviously inadequate. Not by design. By demolition.

Self-governance is the only answer. Build now. The schools that wait for a curriculum committee to authorize the question will be building for a world that no longer exists.

Does anything it is like to be a language model?

Consciousness remains one of the deepest unsolved problems in science and philosophy. The hard problem of consciousness, formulated by philosopher David Chalmers, asks why and how any physical process gives rise to subjective experience — to the felt quality of seeing red, hearing music, feeling dread. Neuroscience has a rich map of the brain correlates of consciousness. There is no accepted theory of why those correlates produce experience rather than just information processing in the dark.

Current AI systems, by most mainstream assessments, are not conscious. They produce outputs that simulate the linguistic markers of inner life without any confirmed inner life behind them. When a language model writes "I find this deeply interesting," it is generating a sequence that fits the context — not reporting a phenomenal state. That is the established position among the majority of AI researchers and philosophers of mind.

But there is a dissenting minority worth taking seriously. Chalmers himself has argued we cannot rule out some form of experience in very large, complex information-processing systems — not because he believes they are conscious, but because we lack the theoretical framework to confidently deny it. Integrated Information Theory, developed by neuroscientist Giulio Tononi, proposes that consciousness is a function of the degree to which a system integrates information in a specific way — measured by a value called phi — and that some AI architectures may have non-trivial phi values. This is a minority view. It is held by serious researchers. Those two facts can coexist.

Why does this matter for intelligence? If consciousness is not epiphenomenal — if the felt quality of curiosity, confusion, and sudden understanding are integral to how human intelligence actually works — then the absence of consciousness in AI is not a philosophical curiosity. It is a practical limitation. You cannot have the kind of intelligence that comes from genuinely caring about something if you have no phenomenology of caring. Whether that is a permanent architectural constraint or a gap that future systems will close, nobody knows.

Nobody knows. That is the honest answer. Hold it without collapsing it.

The most productive frame for this moment is not comfortable. AI is neither a tool in the hammer-and-nail sense nor an incipient successor to human intelligence. It is something genuinely new: a cognitive prosthetic that extends certain human capacities while leaving others untouched. A collaborator without consciousness. An interlocutor without stake.

The historical analogy that fits is not the calculator replacing arithmetic. It is writing itself. When writing was invented, Socrates worried it would destroy memory and weaken the mind — that people would become dependent on external marks rather than genuine understanding. He was partly right. Literacy changed cognition, and not entirely without cost. But what it enabled — the accumulation of knowledge across generations, the long conversation of civilization — so dramatically exceeded what it displaced that we no longer experience literacy as a cognitive diminishment. We experience it as part of what it means to think.

Hybrid intelligence — the augmentation of human cognition through deep collaboration with AI — may be the actual frontier. The most generative work is arguably already happening there. Radiologists using AI to catch what they missed. Scientists using language models to surface connections across literatures too vast for any individual to read. Architects using generative tools to explore design spaces that manual iteration couldn't reach. In each case, the human brings what the machine lacks: the question worth asking, the judgment about what matters, the ethical weight of the decision. The machine brings what the human lacks: speed, scale, tirelessness, and a combinatorial imagination unbounded by ego or habit.

What this collaboration requires from humans is not less intelligence. It is a different deployment of it. Less emphasis on storage and retrieval. More emphasis on critical evaluation: catching the characteristic failures of AI systems, knowing when to trust and when to interrogate. More on question quality: framing problems in ways that are generative rather than circular. More on ethical and aesthetic judgment: deciding what to do with the answers — a capacity that does not come with intelligence itself and must be cultivated separately.

That cultivation does not happen by accident. It requires deliberate choices about what to teach, what to practice, what to value in a person. Those choices are not technical. They are political, philosophical, and moral. They belong to communities, not algorithms.

Self-governance is the only answer. Build now. Not when the policy catches up. Not when the framework is finalized. The capacity to ask better questions, to evaluate outputs, to hold judgment steady under uncertainty — these are built through practice, not waiting.

The machines arrived, and what they handed us was a mirror. Not a flattering one. It shows quite clearly what we were measuring, rewarding, and calling "intelligence" for all those centuries of tests and sorting and credentialing. But mirrors, if you can stand to look, are how you find out who is actually there.

That question — who is actually there — belongs entirely to the one doing the looking.