The Coincidence Nobody Questions
Here's a linguistic accident worth pulling apart. Psychology has a g in general intelligence. Artificial intelligence has a g in artificial general intelligence. Same letter. Same word doing the heavy lifting. And yet these two Gs point at almost nothing in common.
John Nosta flagged this in a recent post on Psychology Today, calling it "a curious linguistic coincidence." I'd go further. I think it's a trap. The shared letter smuggles in an assumption that both fields are chasing the same quarry. They aren't. And conflating them will cost us clarity for years.
What g Actually Means in Psychometrics
Let me be precise about what psychologists have spent over a century studying. When Spearman introduced g back in 1904, he noticed something robust: people who do well on one type of cognitive task tend to do well on others. A child who scores high on verbal reasoning probably scores above average on spatial rotation. Someone with strong working memory typically shows it across domains. This collective cognitive pattern is what Nosta calls the positive manifold, and he's right to emphasize it simply: human cognition hangs together.
But "hangs together" is a statistical observation, not a metaphysical claim. Factor analysis extracts a common factor from correlated test scores. That factor is g. It's real in the sense that it predicts outcomes—school performance, job performance, health outcomes. It is not a little engine inside the skull. It is not a single mechanism. It is a summary of covariance.
I've spent my career pushing back against the reification of g. When you treat a statistical artifact as a causal entity, you stop looking for the actual processes that produce the correlations. And those processes—analytical thinking, creative thinking, practical thinking, what I've called the triarchic system—are heterogeneous. They don't reduce to one number without serious loss.
What "General" Means in AGI
Now flip to AI research. When researchers say "artificial general intelligence," they almost never mean a system that reproduces the positive manifold across human cognitive test batteries. They mean something closer to domain breadth. A system that can write code, then generate an image, then answer a medical question, then plan a multi-step task. The breadth is behavioral, not psychometric.
No AGI researcher is running a factor analysis on their model's outputs. Nobody is asking whether the model's "verbal scores" correlate with its "spatial scores" in a way that produces a clean first principal component. The word "general" is doing a completely different job. It means: this system isn't locked into one task. It can transfer across tasks.
That's useful. It's not g.
Why the Conflation Hurts
Here's where I get blunt. If you accept the linguistic coincidence as evidence of conceptual overlap, you import a whole metaphysics into AI evaluation that doesn't belong there. You start expecting AI systems to show the same internal structure as human cognition. You start treating failures of correlation across tasks as failures of "general intelligence" rather than as evidence that the system was never built to be general in the psychometric sense.
Nosta's word for the gap is "anti-intelligence." He describes AI's computation as stateless, pattern-based, high-dimensional—and therefore alien relative to human cognition. I'd frame it differently. AI doesn't have anti-intelligence. It has a different kind of intelligence, or rather, a different kind of competence that we haven't learned to categorize yet. The "alien" framing is honest. The "general" framing is lazy shorthand that exploits the same letter for rhetorical reach.
Decomposing the Construct
The genuinely interesting question isn't whether AI has g. It's whether human intelligence was ever one thing. Nosta puts this directly: perhaps we have mistaken the architecture of human intelligence for the definition of intelligence. AI is the first real chance to decompose this construct and learn that intelligence wasn't one thing in the first place.
I've argued something adjacent for decades. The positive manifold is contingent on the particular environment that produced it—the particular test formats, the particular educational systems, the particular demands of modern life. Remove those constraints, and the correlations might dissolve. The cluster of capacities we call general intelligence might be an artifact of a narrow slice of human experience rather than a discovery about mind's deep structure.
A 2025 paper by Stephanie Theves in the Journal of Neuroscience makes a parallel move from the neural side. She proposes that general intelligence is better understood through the neural process of analogy-making rather than through a single factor. This reframes g as a description of what certain processes produce, not a description of what the mind fundamentally is. Theves works from Max Planck Institute data on how the brain treats structurally similar problems across domains. It's a bottom-up account. It doesn't need g. It needs processes that happen to correlate.
The Letter Doesn't Carry the Meaning
So where does this leave us? Two fields share a letter. One field's g is a statistical summary of positive correlations across human test performance. The other field's g is a promise about behavioral flexibility across tasks. They rhyme. They don't mean the same thing.
I'd urge AI researchers to stop borrowing psychometric vocabulary without the psychometric machinery. If you want to claim your system is "general," show me the correlation structure. Show me whether breadth across domains comes with internal coherence the way human general intelligence does. Or drop the g entirely and call it broad AI, or multi-task AI, or whatever you like. The letter costs you a century of unresolved debate you haven't earned.
And I'd urge my own colleagues in psychology to notice what AI is actually teaching us. The fact that a system can be enormously capable across domains without producing a clean positive manifold is evidence that the manifold might be contingent. That g might describe one species' cognitive architecture, not intelligence as such. The decomposition has started. Whether we're ready for what falls out of the construct is a different question.
