What intelligence actually is
The Reason We Keep Arguing About AI Is That We've Never Agreed What Intelligence Means
Every debate about whether AI can surpass human intelligence is secretly a debate about something we've never managed to define.
The Idea
Intelligence is one of those words that feels obvious until you try to pin it down. Psychologists have been arguing about its definition for over a century, and the best they've landed on is something like 'the ability to learn from experience, adapt to new situations, and solve abstract problems.' That's a start — but notice how much it leaves out. It says nothing about creativity, embodied understanding, emotional attunement, or the capacity to know what matters in the first place. What makes this more than academic fussing is that AI has forced the question into the open. When a language model passes a bar exam, writes a sonnet, or diagnoses a rare disease more accurately than a trained physician, the instinctive response is 'but that's not really intelligence.' And maybe it isn't — but that instinct deserves scrutiny. Are we genuinely identifying a missing capacity, or are we shifting the goalposts because the machine has cleared the previous bar? The philosopher Howard Gardner proposed that intelligence isn't one thing but many — linguistic, logical-mathematical, spatial, musical, interpersonal, bodily-kinaesthetic, and more. AI systems tend to be spectacular at some of these and baffling failures at others. A model that can summarise Kant may not understand that a child crying in a photograph is in pain — not really, not in any way that connects to lived experience. Whether that gap is a temporary engineering problem or something more fundamental is the question that makes this whole conversation so genuinely unresolved.
In the World
In 1997, when IBM's Deep Blue defeated world chess champion Garry Kasparov, commentators declared it a milestone for machine intelligence. Kasparov himself disagreed — not because he was a sore loser, but because he noticed something specific: Deep Blue had no idea it was playing chess. It was performing a form of brute-force tree-search through possible moves, executing billions of calculations per second. It could not transfer any of that 'skill' to draughts, or recognise a chess piece if you showed it a photograph, or feel the pressure of a ticking clock. Its competence was genuine within an extraordinarily narrow domain, and zero everywhere else. Decades later, researchers at DeepMind built AlphaZero, which taught itself chess, shogi, and Go simultaneously — and did so by playing against itself for a few hours, starting with nothing but the rules. That felt more like something. But then came a different test: researchers found that AlphaZero, despite mastering chess to a superhuman level, could be defeated by simple adversarial inputs — illegal-looking positions that confused it in ways no human grandmaster would ever be confused. The system had learned the game without learning anything about the game — no physical intuition, no concept of what a board is for, no sense of the human drama surrounding it. Kasparov's observation from 1997 still lands: passing a performance threshold is not the same as possessing the thing the threshold was meant to measure.
Why It Matters
The reason this isn't just a philosophy seminar topic is that how we define intelligence shapes what we build, what we regulate, and what we fear. If intelligence is fundamentally about pattern recognition and prediction at scale, then current AI is already extraordinary — and the gap between human and machine may be narrower than we think, likely to close further. If intelligence is irreducibly tied to embodiment, consciousness, or genuine understanding of the world rather than statistical proxies for it, then we may be building something genuinely alien — impressive but not quite a mind. Neither answer is obviously correct, which is itself the important thing to hold. People who are certain AI will surpass human intelligence soon and people who are certain it never truly can often share the same flaw: a fixed, unexamined idea of what intelligence is. The more useful posture is to stay curious about the definition itself — because whoever controls that definition will, in a very real sense, control the conversation about one of the most consequential technologies in human history.
A Question to Ponder
If a system consistently outperforms humans on a task we considered a hallmark of intelligence, but does so through a process completely unlike how humans think, should we update our definition of intelligence — or our definition of the task?
Get a new one of these every morning.
Start learning with Thinkable