Poetry in humans and machines: Who’s great?

Hollis Robbins, LLM poetry and the "greatness" question, Anecdotal Value, Jan, 7, 2026.

I posted the following comment to Hollis’s article:

An oblique observation. To the extent that an LLM can be said to "simulate" a human brain, it would be simulating the neocortex. With only 16 B neurons the neocortex contains only 19% of 86 B neurons in the brain. Thus LLMs, no matter how many parameters, are operating with only a fraction of the capacity of a poet's brain. To be sure, the poems they're trained on were created by full-brained poets, so those poems bear the traces of full brains. But I figure a full-brained poet would be able to find new kinds of paths through the brain's full state space. An LLM's ability to do that would seem to be rather truncated.

It's not clear to me how this bears on greatness in poetry. But I would like to think that at least some poems are great because they opened up new whole brain possibilities.

The purpose of this post is to develop those ideas more carefully. First I’ll talk about the idea of the brain’s state space. Then we move to the idea of a text as tracing a path through that space. Finally we’ll take up greatness in poetry.

The phase space of a brain

The notion of phase space, as I’m using the term, derives from statistical mechanics. It’s a way of thinking about how a physical system changes over time. Change is a movement from one state of the system to another.

The phase space of a system depends on three things: 1) how many elements the system has, 2) how many states each element can take, and 3) how the individual elements are coupled. The phase space has a dimension for each element of the system. The system’s state at any given moment will be a point in the phase space.

If the system has only three elements, then its phase space has three dimensions. Its state at any moment will be a point in that space. Notice that time isn’t one of those dimensions. But as the system changes over time its position in the phase space will change as well. It is possible, however, keep track of those positions.

A phase space having only three dimensions is easy to visualize. What about a phase space with four, five, six, thirty-eight, of 999,000 dimensions? Impossible to visualize.

The human brain has roughly 86 billion neurons. A phase space with that many dimensions will have 86 billion dimensions; impossible to visualize. But the individual neuron might not be the appropriate element to be the primitive unit in the brain’s phase space. The individual synapse – a connection between one neuron and another – may be the appropriate phase space primitive. Each neuron is connected to an average of 7000 other neurons, making the number of dimensions in the brain’s phase space much higher. The exact number doesn’t matter; the point is simply that the phase space is huge.

Note that neurons do not operate independently of one another. They are coupled together. The brain is divided into a number of functional areas (NFA, my term) where the connections between neurons within an area are more numerous that the connections each neuron has to neurons in other NFAs.

At this point let me introduce a query I put to ChatGPT 5.2:

The human brain has roughly 86 billion neurons. How many different functional regions does it have? I believe that the neocortex has between 100 and 200 different functional regions, but only has 16 billion neurons or so. How many different functional regions do we have in subcortical area?

The resulting discussion was a bit messy and complicated, as these things can be, but the upshot is that that brain seems to have roughly 700 NFAs. Only half of them are in the cortex. That’s very important (recall my original comment to Hollis). Later in the conversation ChatGPT observed:

• If LLM behavior resembles anything in the brain, it most closely resembles cortical linguistic competence: high-dimensional, distributed patterning; associative completion; flexible recombination of learned material.
• But cortex is only one part of the full system. Even taking the cortex generously as “the cognitive engine,” it is not the whole engine.

Thus, “At best it’s like a subsystem of the brain, and specifically one that is downstream of bodily regulation, affect, action selection, and reinforcement gating.”

Text as trace

I tend to think of texts, any texts, as the trace of a trajectory through the brain’s state space. But, at best, an LLM models only part of that space. Thus, returning to ChatGPT:

Even when an LLM produces a convincing poem, its method is necessarily missing the full set of whole-brain constraints that shaped the poetic corpus it learned from. Therefore its exploration of possibility space is structurally truncated relative to the exploration performed by a whole-brain poet.

ChatGPT offered me this paragraph:

If we treat a text as a path through a brain’s phase space, then poetry is not simply the output of a language module. It is the trace of a whole-brain trajectory: language regions interacting with memory, interoception, affect, reward, attention, and social cognition. A large language model is trained on the textual residue of those trajectories — the fossil record, not the living process. The poems in its corpus were written by full-brained poets, so the text contains evidence of that fuller generative machinery. But the model’s own generative dynamics are constrained to what can be recovered from token sequences and their statistical regularities. However impressive the simulation, it is necessarily missing the deeper gating systems that, in humans, determine what becomes sayable, what becomes bearable, and what becomes necessary. My hunch is that some poems are great not merely because they are well-formed linguistic artifacts, but because they open new whole-brain possibilities — new stable pathways of feeling-and-thought — that alter what the mind can do afterward.

What of humans and AI’s

I then prompted ChatGPT:

The issue that’s hanging over this kind of discussion is ultimately: Are there good roles for humans in a world of intelligent machines or are we replaceable (and so will be replaced)? That’s a messy question that begs many issues as formulated and so is rather meaningless. [...]

At the moment I’m working on a book about the social and cultural implication of AI. I started working on the book last summer. I broke off work from September through about a week ago. I’m now back at it. Last year I had a title, scads of notes and drafts and two or three versions of an outline. I’ve now arrived at an outline that I don’t think I could have produced without the help of Claude and ChatGPT. Many of the chapters I was working on last year are in this new version, 6 of them in fact. But I’ve got two chapters I didn’t have last year and an overall architecture that I didn’t have. Moreover I don’t think I could have produced that architecture on my own. The AIs gave me the means of holding the whole thing in my head and working at different scales so that I’ve been able to do something new.

I’m a good thinker. I’m better working with AIs. I should think the same would be true for poets.

We had some further discussion and ChatGPT offered me this paragraph:

The question “Will AI replace us?” is too blunt to be useful. My own experience suggests something more precise: large language models can function as cognitive instruments that let a person think at scales they could not reliably manage alone. I can now hold a book-length architecture in view, revise it coherently, and move between local detail and global structure without losing the plot. I am not a poet, but I am a thinker, and I know the difference between producing text and producing understanding. The models are not merely faster typists with larger memories. They are interfaces that make new kinds of conceptual control possible. If that is true for expository thinking, it is likely true for poetry as well: not that the machine becomes the poet, but that the poet gains a new way to explore and stabilize trajectories through a much larger space of possibilities.

I’m willing to let things rest there for the moment.

(Oh, I also asked ChatGPT to produce that illustration at the head of this post.)