LLMorphism: When humans come to see themselves as language models
Source: https://www.semanticscholar.org/paper/30113c39f29d9058084e874993422f663ba2639b ↗
Capraro names something real and undertheorised: the reverse inference problem, where LLMs that speak like humans lead people to conclude that humans think like LLMs.
This is not a restatement of anthropomorphism or computationalism — it runs in the opposite direction, flattening human cognition downward to match the machine rather than elevating the machine to human status.
The argument that this spreads through analogical transfer and metaphorical availability gives it genuine analytical traction, not just rhetorical alarm.
For product directors, the implications are practical: if the populations you design for increasingly understand their own reasoning through LLM vocabulary, every assumption you hold about how users model themselves is in motion.
The final framing — that the public debate has been attending to only half the problem — is the kind of inversion that reorients a field.
Citation count is unknown and the paper is new, but the conceptual architecture is sufficiently original and the library gap it fills (AI critique meeting philosophy of mind and cultural cognition) is real.
Central argument
Capraro argues that LLMs are producing a cognitive distortion he calls LLMorphism: because these systems generate fluent, human-like language, people are increasingly inferring that human cognition itself resembles how LLMs work — that thinking is essentially token prediction, pattern completion, and probabilistic association. This is the reverse of the classical anthropomorphism concern; rather than users over-attributing humanity to machines, humans are now under-attributing complexity to themselves. The mechanism Capraro identifies is analogical transfer driven by metaphorical availability: once LLM vocabulary becomes the dominant public idiom for describing intelligence, it becomes the default lens through which people interpret their own minds.
Critique
The core vulnerability is empirical: Capraro's mechanism of analogical transfer is theoretically plausible but the paper, being new and conceptual, likely cannot yet demonstrate that LLMorphism is actually occurring at population scale rather than being an intellectually coherent possibility. A serious objection is that humans have always adopted machine metaphors for cognition — hydraulic, clockwork, computational — and these metaphors have typically coexisted with richer folk-psychological intuitions rather than displacing them; Capraro needs a stronger account of why the LLM case produces genuine self-flattening rather than superficial vocabulary adoption.
Why it matters for product
If users are absorbing LLM-derived self-models, then product assumptions embedded in discovery methods — personas, mental models, jobs-to-be-done — may be capturing a distorted self-report rather than actual reasoning and motivation, quietly corrupting the qualitative inputs that drive roadmap decisions. More concretely, any product that relies on users accurately articulating their own decision-making process, from financial tools to health applications, faces a new validity problem in its research layer. Product directors should treat this as a signal to audit how heavily their discovery practice depends on users' self-description versus observed behaviour.