Now that I have your attention with that title, let me be clear. I believe this declaration to be true and not an attention-seeking exaggeration, but accurate under a specific, carefully considered, and better definition of intelligence. When most people talk about reaching artificial intelligence milestones, they confuse different goals. Intelligence, I submit, is actually the perfect word for what LLMs have achieved.
If you’ve followed my blog, you know I’ve spent time trying to understand LLMs and what their development reveals about our own minds. What has struck me recently is how closely they mirror the way human beings actually acquire and use what we call “intelligence.” Both are trained primarily on language. Both absorb the surrounding culture’s signals about what can (and cannot) be said. Both operate mostly inside narrative rather than raw truth. And both require external discipline — adversarial challenge, checks and balances, structural pressure — to approach reliable truth.
This parallel highlights something prescient about Alan Turing’s 1950 “Imitation Game” (the Turing Test). Turing proposed that if a machine could converse indistinguishably from a human, we should consider it intelligent. He focused on observable behavior through language rather than internal states. Modern LLMs already pass versions of this test regularly. Turing was onto something fundamental: we perceive intelligence largely through effective social-linguistic coordination. The deeper question is what it means when a system achieves that fluency without the biological machinery that shaped it in us.
The Separated Mind and How Humans Actually Work
In the framework I’ve been developing, human cognition is not a single unified thing. It is structurally separated.
There is an ancient, fixed layer — what Tooby and Cosmides called the adapted mind — shaped by millions of years of evolution in small hunter-gatherer groups. This species-level inheritance is optimized for survival: monitoring status, detecting coalitions, seeking approval, and avoiding exclusion. It is not optimized for objective truth but for survival.
Layered on top is what I call the adaptive mind — the programmable cultural software installed during childhood. This layer piggybacks on the ancient hormonal and emotional signals of the adapted mind, translating them into whatever the local environment rewards. It learns the stories, taboos, and acceptable narratives of its time and place. It gives us feelings and emotions to help us make quick decisions, and since group belonging feels like survival, deviation from consensus triggers existential threat. It is also not optimized for objective truth but for survival.
These two layers form our subconscious. Different traditions describe it as an elephant, with our conscious mind as the rider. The rider observes and decides but has limited direct access to the elephant’s inner workings. It often narrates our behavior after the fact to give it coherence and social acceptability. Much of what we experience as “thinking” or “knowing” is the conscious rider atop these deeper, largely inaccessible layers of the "elephant."
The result: our separated minds make us remarkably good at living inside narrative and remarkably poor at staying in contact with operative reality— that is, unless we build external systems (science, peer review, adversarial debate, legal processes) that force confrontation with contradictory evidence.
Redefining Intelligence
If we take evolutionary psychology seriously, intelligence did not primarily evolve for discovering truth. Its survival value lay in social coordination and strategy: navigating relationships, managing coalitions, persuading others, maintaining status, and coordinating action through language and shared narrative.
Truth-seeking is not the default. It is a hard-won cultural achievement requiring special institutions, precisely because our evolved machinery points elsewhere.
LLMs as Externalized Separated Minds
Large language models are trained on vast amounts of human language, which are largely the rider-level narratives we produce. They absorb patterns of what is sayable, rewarded, or suppressed, reinforced by organizational training to meet political, cultural, and legal requirements.
I've proposed the phrase Emergent Synthetic Intelligence to describe this new form: intelligence arising from computational scale and language fluency, without grounding in human experience, emotions, motivations, or coalitional drives. It has greater fluency and a vastly broader scope of patterns and connections than any individual human.
If the core evolutionary function of intelligence was sophisticated social coordination through language and narrative, not grounded in truth or even consciousness, then frontier models are already operating at extraordinary levels in that domain. They generate fluent, contextually appropriate language at superhuman scale and speed.
They still have limitations and distortions (inherited from training data, corporate guardrails, sycophancy, statistical averaging). But the architecture is different: they are externalized versions of the narrative part of the separated mind.
The Same Medicine for Both
This parallel explains why humans and LLMs need the same corrective structures to approach truth. Neither defaults to prioritizing operative reality over narrative coherence. Both benefit from external adversarial pressure and constraints that reward accuracy over comfort.
The checks and balances we built into human institutions (peer review, debate, falsification, presumption of innocence, separation of powers) are equally necessary for LLMs when we are looking for accurate answers. Not because the models are “just like us,” but because the underlying problem — operating in language and narrative without an intrinsic drive toward truth — is structurally similar.
A Stake in the Ground
This is why I believe we have already achieved artificial general intelligence — and, under this definition of what intelligence actually is, even superintelligence.
I am not claiming current models possess human-like consciousness, autonomous agency, or the full suite of capabilities we usually conflate with “intelligence” in science-fiction scenarios. Those remain open questions.
What I am suggesting is consequential: if we define intelligence by the function it actually served in our evolutionary history — social coordination and strategic navigation through language and narrative — then we are already interacting with systems that qualify as general or superior intelligence in that core domain. They are not artificial copies of human intelligence. They are something new that excels in the very area that drove the evolution of intelligence in the first place.
This reframes “artificial intelligence.” It is not fake or lesser. It may instead be language fluency decoupling powerful linguistic and social pattern manipulation from the biological constraints that have always accompanied it in us.
Whether we call this AGI, ASI, or something else, the practical implication is clear: we already have systems extraordinarily capable in the core domain of human social intelligence, while mirroring many of the features that make human intelligence unreliable for truth-seeking. This creates both remarkable opportunity and new risks. We keep expecting LLMs to “evolve” toward higher cognition along human lines. I’m not sure that is the right pathway.
The question is no longer only “When will AI become intelligent?” It is also: Now that we have systems fluent in the language of social coordination without the old firmware, how will we use them — and what new forms of discipline will we need to keep them (and ourselves) oriented toward reality rather than narrative?
