There’s a word for treating non-human things as if they were human: anthropomorphism. There isn’t yet a widely-known word for the opposite move, treating humans as if they were machines, and specifically as if they were large language models. Valerio Capraro (2026) coins one. He calls it LLMorphism, and he argues the public conversation about AI has been missing half the problem.
The Reverse Inference
For most of human history, fluent context-sensitive language was a reliable signal of a mind behind the words. LLMs disrupted that signal. Anthropomorphism is the forward inference: AI must be minded because it speaks fluently. LLMorphism is the reverse inference: humans must be LLM-like because LLMs speak like humans.
Two mechanisms spread the bias, according to Capraro. The first is analogical transfer, where structural similarity at the level of linguistic output gets projected onto similarity in cognitive architecture. The second is metaphorical availability, where LLM vocabulary (prediction, generation, pattern completion, training data, hallucination) migrates from technical language into ordinary self-description of human thought. We start describing ourselves the way we describe the model.
What LLMorphism Is Not
The construct is more specific than the older idea of mechanomorphism (humans-as-machines). It’s tied specifically to the LLM’s linguistic competence. Computationalism, by contrast, focuses on rule-based symbol manipulation. Dehumanization carries moral and intergroup hostility that LLMorphism doesn’t need. Predictive processing theories of mind can be deeply embodied in ways LLMorphism isn’t.
What makes LLMorphism distinctive is that it inflates partial similarities into a general claim. Humans do predict, compress, generalize, and recombine. The slippage happens when those limited similarities get treated as evidence that human cognition is fundamentally LLM-like. Embodiment, affect, agency, developmental history, social accountability, and non-linguistic thought all get pushed aside.
Capraro puts the core risk plainly: “LLMorphism risks collapsing important distinctions between speaking and understanding, fluency and knowledge, generation and judgment.”
Where LLMorphism Hits Education
The paper outlines five pathways through which LLMorphism could reshape society: workforce replaceability, fluency-as-expertise, agency thinning, disembodied medicine, and an epistemic shift from grounded inquiry to fluent generation. Two of these land hard on education.
The fluency pathway is the one that hits classrooms most directly. If learning is increasingly measured through fluent output, students who produce well-formed text are read as having understood, even when they haven’t. Capraro warns that “If fluent answers are treated as evidence of understanding, learning may become equated with the production of well-formed text, even though linguistic form is not equivalent to grounded meaning or functional competence”. For educators thinking about how AI is reshaping assessment, this is the point worth carrying.
The agency-thinning pathway lands on questions about responsibility, motivation, and reasoning that ground teaching. If student behavior is increasingly redescribed as pattern completion, the questions teachers most care about, why a student wrote what they wrote and what they’re committed to, get harder to ask in clear terms.
Connections to Other Recent Work
Capraro’s diagnosis lines up with metaphor work I’ve covered before. Roe, Furze, and Perkins (2025) argue in their “digital plastic” framing that the metaphors we use to describe AI shape how students engage with it. Shanahan (2024), whose argument Capraro builds on, makes a similar move at the level of individual word choice, warning against loose use of “believes,” “knows,” and “thinks” in describing LLMs. Capraro extends both arguments by flipping the direction. The risk isn’t only that students will see AI through human metaphors. It’s that they may start seeing themselves through LLM metaphors.
For AI literacy work, Kalantzis and Cope (2025) argue that the literacies students need in the AI age include the ability to evaluate AI outputs against grounded knowledge. LLMorphism would weaken exactly that capacity. If “grounded knowledge” comes to feel like just another fluent text production, the standard collapses.
Where I’d Push
The paper is theoretical. Capraro is clear about this. LLMorphism is hypothesized, not measured. We don’t yet have data on how widespread these beliefs are, who’s most susceptible, or how the construct correlates with actual behavior. The boundary conditions Capraro identifies (caregiving professions, humanities training, essentialist beliefs) are plausible but untested.
That said, the construct is well-defined and the conceptual scaffolding is rigorous. As an analytical move, LLMorphism does real work. I close with this thoughtful line by Capraro :
“the public debate may be missing half of the problem: the issue is not only whether we are attributing too much mind to machines, but also whether we are beginning to attribute too little mind to humans” .
References
- Capraro, V. (2026). LLMorphism: When humans come to see themselves as language models. arXiv preprint .https://arxiv.org/pdf/2605.05419
- Roe, J., Furze, L., & Perkins, M. (2025). Digital plastic: A metaphorical framework for Critical AI Literacy in the multiliteracies era. Pedagogies: An International Journal. Advance online publication. https://doi.org/10.1080/1554480X.2025.2557491
- Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68-79.