Is the LLM an anti-poet?
To be better at poetry, LLMs must be programmed with rewards for useful unusualness. They are usually built to minimise surprise (or perplexity) by predicting the most likely words. Poets target meaningful surprise – triggering a neural spike & asking us to see the world afresh.
Poets can make the familiar strange & the strange familiar: “The evening is spread out against the sky”. An LLM predicts.. “…like a blanket of stars.”. A poet writes “like a patient etherized upon a table”.
To write poetry, AI must prioritize the shock of the new over the safety of the likely (hence why it is a ‘mean’ machine).
Cognitive science suggests that metaphor is embodied. We speak of love as hot partly because, as infants, we learned that to be held is to be warm. An AI knows that ‘warm’ correlates statistically with ‘love,’ but it has never shivered from the cold. Without a body, its use of metaphors is ungrounded.
An LLM is a map without a territory. It possesses the syntax but is exiled from the experience (the feel). Writing poetry is the interaction of subconscious feeling and experience with conscious craft. It is human ‘technology’ (meaning ‘art/craft’ + ‘word/speech’ etymologically).
Great poetry thrives on ambiguity and pregnant possibilities – it asks us to live with seeming contradictory truths (Keat’s negative capability). A great poem is like a puzzle which surprises into a truth (not the ‘Truth’). But truth is not the answer to the puzzle it is the shape of the question.
For most LLMs a problem that could give rise to a poem is a dataset to be compressed—a noisy signal that must be resolved into a probable output. It seeks the ‘Truth’ as the solution, it looks for the most probable path across this linguistic landscape – a poet knows the truth is the landscape.
John Keats called this world “The Vale of Soul-Making,” where suffering forges a Soul. AI has no real ‘skin in the game’. Without mortality or the capacity to suffer, how can it truly feel the comedy and tragedy of life and loss? How can it be hurt into poetry? (Auden)
Current AI is usually, by design, a mean-seeking machine—an engine that averages the collective output of human language into a smooth shallow valley. But great art is the inventive outlier. It scales mountains, dives deep into the sea. A poet’s catch is often submerged – they fish for our souls.
C.S. Lewis warned of the risk of our becoming”Men Without Chests”— people with greed and intellect but without moral sentiment. Ironically we risk becoming ever more like AI as we converge. Is the real danger: not that AI will never write a great poem but that we will lose our ability to hear or read one?
“…People are not going
To dream of baboons and periwinkles.
Only, here and there, an old sailor,
Drunk and asleep in his boots,
Catches tigers
In red weather.
(‘Disillusionment of Ten O’Clock’, Wallace Stevens)
Compossible - that which can live together 
Leave a comment