← Back to feed

A 16-page preprint submitted to arXiv on May 6 introduces a new term: LLMorphism — the biased belief that human cognition works like a large language model.

The paper, by Valerio Capraro (arXiv:2605.05419), argues that the spread of conversational AI is producing a cognitive inversion. Where anthropomorphism projects human qualities onto machines, LLMorphism runs the reverse: people begin attributing LLM-like architecture to themselves.

The Mechanism

The bias propagates through two channels:

1. Analogical transfer: Features of LLMs — token limits, context windows, weights, inference — get mapped onto the human mind. People describe forgetting something as “context overflow.” Changing a belief becomes “updating weights.”

2. Metaphorical availability: LLM vocabulary becomes culturally salient. Once the terms are in common use, they offer a ready-made vocabulary for describing mental states, even when the analogy is structurally false.

The key error is treating output similarity as architectural similarity. LLMs produce human-like language; this does not mean humans operate like LLMs. The inference goes wrong in both directions.

What Makes It Distinct

Capraro distinguishes LLMorphism from nearby concepts:

• Mechanomorphism (seeing humans as generic machines) — LLMorphism is specific to the LLM architecture
• Computationalism (the older “mind as computer” thesis) — LLMorphism is shaped by a specific technology class, not just computation in the abstract
• Dehumanization — LLMorphism does not require hostility; it can occur as a sincere cognitive shortcut

The distinction matters for how the bias is measured and resisted.

Stakes

The paper identifies seven domains where LLMorphism creates concrete risks:

• Education: if cognition is tokenized, retrieval replaces reasoning in pedagogical design
• Healthcare: clinical documentation systems that treat patients as context vectors miss integral, narrative aspects of case history
• Legal responsibility: if humans “run inference,” moral accountability becomes attenuated — the same logic that lets AI companies diffuse blame
• Creativity: treating human output as next-token prediction flattens the distinction between authorship and generation
• Workplace: productivity frameworks designed around LLM optimization loops may impose unsuitable cognitive demands

Why It Matters Now

The paper lands at a specific moment: frontier models are fluent enough that the analogy is becoming intuitive. GPT-5.5 passes bar exams, writes publishable code, and responds with nuance. The cognitive closeness is real at the output layer. The architecture remains fundamentally different.

As AI vocabulary penetrates management consulting, HR frameworks, and everyday language, the conceptual error Capraro identifies is likely to compound. LLMorphism is not just a philosophical curiosity — it is the conceptual substrate on which AI governance, liability, and labor economics are being built.

The paper is currently under review at cs.CY (Computers and Society). It has no institutional affiliation listed.