The most sophisticated language model in the world has never felt the weight of a bad decision at 3 a.m., never experienced the particular dread of a phone ringing at an unusual hour, never known the relief of a cool drink on a hot day. This is not a bug to be patched in version 4.0. It is a defining feature of what these systems are — and a clue to their permanent limitations.

The AI industry's preferred framing presents current constraints as temporary inconveniences on the road to artificial general intelligence. Give it more data, more compute, more clever architecture, and the machine will eventually understand everything humans understand. This narrative is useful for fundraising. It is also almost certainly wrong in ways that matter for how we should actually deploy these tools.

The body problem

Philosophers have argued for decades that human cognition is not a process that happens to occur inside a skull but is fundamentally shaped by having a body that moves through space, gets tired, feels pain, and will eventually die. The technical term is "embodied cognition," but the intuition is simpler: you cannot truly understand "heavy" without having lifted things, "bitter" without having tasted, "exhausted" without having pushed past your limits.

Language models learn statistical patterns across billions of words. They learn that "exhausted" often appears near "sleep" and "work" and "finally." They can deploy the word appropriately in context. But the deployment is mimicry of form, not comprehension of substance. The system has no referent for the feeling, because feelings require a body capable of having them.

This is not mysticism. It is an observation about what kind of understanding emerges from what kind of learning process. A model trained on chess games understands chess. A model trained on text understands text — including text about exhaustion, grief, desire, and fear. Understanding the text is not the same as understanding the experiences the text describes.

Why the gap persists

The standard rebuttal is that humans also learn about many things indirectly. You have probably never experienced war, but you understand something about it from books and films. True enough. But your indirect understanding is anchored by direct experience of related phenomena — fear, pain, loss, loyalty — that you have felt in your own body. The indirect knowledge plugs into a vast network of embodied experience. For a language model, there is no such network. Every concept is equally indirect, equally floating free of physical reality.

Some researchers are working on robotic embodiment, giving AI systems bodies that interact with the physical world. This is promising work, but it does not solve the problem as cleanly as it might appear. A robot that learns to navigate a warehouse is not thereby learning what it feels like to be tired, to be hungry, to be afraid of death. It is learning physics, not phenomenology.

What follows from this

None of this means language models are useless. They are extraordinarily useful for tasks that involve manipulating language, recognizing patterns in data, and generating text that conforms to specified constraints. These are valuable capabilities. The mistake is assuming they generalize to capabilities that require something language models structurally cannot have.

When an AI system gives advice about grief, it is pattern-matching against text written by people who have grieved. When it helps draft a medical diagnosis, it is synthesizing patterns from clinical literature. These are legitimate applications — but they are fundamentally different from the understanding a human brings to the same tasks. The human has a body that can get sick, that will die, that has loved people who have died.

Our take

The AI industry's incentives push toward overpromising. Investors want to fund the path to superintelligence, not the path to very good autocomplete. But intellectual honesty requires acknowledging that some gaps may not close with more scale. The absence of embodiment is not a missing feature to be added later; it is a structural property of systems that learn from text. This does not make such systems less useful. It makes them useful for different things than their most enthusiastic promoters suggest. The sooner we accept this, the sooner we can deploy AI where it genuinely helps and stop waiting for it to become something it cannot be.