LLMorphism: Bias in Modeling Cognition

• LLMorphism is defined as the cognitive bias that models human cognition on the basis of LLM token prediction, recombination, and statistical generalization.
• It contrasts embodied human cognition—which includes affect, social context, and developmental influences—with the pattern-based operations of LLMs.
• LLMorphism poses risks in domains like work, education, and healthcare by conflating surface linguistic fluency with genuine understanding and ethical responsibility.

LLMorphism denotes a specific cognitive bias: the belief that human cognition operates analogously to the architecture and behavior of LLMs, specifically as systems engaged primarily in token prediction, pattern completion, and statistical generalization. This bias has become increasingly prevalent with the rise of conversational LLMs, prompting some to infer that since artificial systems generate human-like language, human thought must resemble LLM processes. LLMorphism distinguishes itself from anthropomorphism (attributing humanlike qualities to machines), mechanomorphism (likening humans to machines in general), and classical computationalism (regarding the mind as a rule-based symbol manipulator), targeting instead a representational shift that ascribes specifically LLM-like properties to human cognition (Capraro, 6 May 2026).

1. Definitional Scope and Contrasts With Related Metaphors

At its core, LLMorphism consists of projecting the salient properties of LLMs—namely next-token prediction, recombination of training data, and statistical linguistic generalization—onto the human mind. Unlike mechanomorphism, which associates humans with generic machines (clocks, engines, robots, or computers), LLMorphism is tied to the distinctive mechanisms of contemporary generative models. It diverges from anthropomorphism by reversing direction: rather than humanizing machines, it models humans after machines. It is distinct from dehumanization or objectification, as it need not involve hostile or instrumental attitudes but primarily entails a representational realignment. Similarly, while classical computationalism posits the mind as a symbol-processing system, LLMorphism specifically frames everyday creativity and discourse as token-based prediction rather than rule-based manipulation.

2. Psychological Origins: Analogical Transfer and Metaphorical Availability

LLMorphism emerges through two principal and mutually reinforcing cognitive pathways:

1. Analogical Transfer: Drawing upon structure-mapping theory, individuals align relational structures between LLM output (HL: fluent, context-sensitive text) and human language use (HH: fluent, context-sensitive speech). A projection function $$A: \text{Structure}(\text{HL}) \to \text{Structure}(\text{HH})$$ enables mapping LLM mechanisms—token prediction, recombination, distributional "knowledge"—onto human cognition whenever observable behavior appears similar.
2. Metaphorical Availability: Technical terminology from AI—such as "prompting," "generation," "hallucination," "next-token prediction," "training data"—enters ordinary language, forming a lexicon $$M = \{\text{prediction}, \text{pattern completion}, \text{training}, \text{generation}, ...\}$$. Over time, this lexicon shapes everyday descriptions of memory, creativity, introspection, and explanation, displacing earlier mechanical or agentive metaphors and making LLM-based conceptualizations of mind feel self-evident.

3. Foundations of the Bias: Disanalogy Between Surface Output and Cognitive Architecture

LLMorphism is characterized as a bias because it illegitimately infers equivalence between similarity of linguistic outputs and underlying cognitive processes. While both LLMs and humans generate contextually fluent language, human linguistic ability is fundamentally grounded in embodiment, affect, social norms, developmental context, and world knowledge. LLMs, by contrast, extract patterns from massive text corpora without the benefit of first-person perception or goal-directed agency. Conflating next-word prediction with genuine understanding collapses distinctions between fluency and semantic grounding, and between generative competence and normative or ethical judgment (Capraro, 6 May 2026). This conflation risks eroding recognition of dimensions of cognition that remain inaccessible to current LLM architectures.

4. Mechanisms of Propagation and Resistance

LLMorphism is not a deterministic or universal outcome. Several boundary conditions and resistive factors moderate its spread:

• Cognitive Disanalogies: Awareness of key differences—such as the absence of intention, embodiment, and social context in LLMs—attenuates the effect.
• Competing Metaphors: The persistence of hydraulic, agent-based, narrative, or spiritual metaphors provides cultural resistance; for example, people with essentialist or religious worldviews often reject LLMorphic reductionism.
• Professional and Experiential Contexts: Fields centered on lived experience and embodiment (e.g., caregiving, early-childhood education, psychotherapy) produce repeated encounters with aspects of human life resistant to LLM-style abstractions. Training in humanities and qualitative social sciences, which highlight first-person perspective and narrative, also counteracts the substitution of LLM-derived metaphors.

5. Societal Domains: Illustrative Pathways and Risks

LLMorphism shapes several major societal domains, each mediated by distinct theoretical mechanisms:

Domain	Mechanism	Illustrative Risk
Work & Labor	Replaceability mechanism	Worker substitution normalized by "output generator" framing
Education	Fluency mechanism	Equating expertise/learning with surface competence
Responsibility	Agency-thinning mechanism	Sidelining intention and moral accountability
Healthcare	Disembodiment mechanism	Neglect of nonverbal/embodied cues, diagnostic failure
Knowledge	Epistemic mechanism	Plausibility as substitute for evidence or justification

• In work, the view that employees function as "output generators" plausibly justifies technological replacement.
• In education, the risk is that surface fluency becomes conflated with deep expertise, undermining value placed on tacit knowledge and community-derived judgment.
• For responsibility, casting humans as pattern-completing devices thins attributions of agency and erodes practices of norm-based accountability.
• In healthcare, prioritizing self-reported text over embodied interactions may result in diagnostic and therapeutic oversights, particularly in domains where coherent language does not map directly onto clinical status.
• Concerning knowledge and epistemics, LLMorphism fosters a shift where coherence and plausibility are valued over rigorous proof, increasing susceptibility to misinformation and shallow reasoning.

6. Implications for Creativity, Responsibility, and Human Dignity

A pervasive LLMorphic frame recasts creativity as mere statistical recombination, neglecting the richness of embodied inspiration and cultural situatedness. Responsibility and agency are diminished when human moral action is conceived as a function of input-output mappings rather than deliberation and reflective self-governance. This framing undermines human dignity by reducing persons to systems that generate plausible text, threatening conceptions of autonomy, intrinsic value, and depth (Capraro, 6 May 2026).

7. Open Problems and Research Directions

Current research suggests several empirical and theoretical priorities:

• Psychometric Development: Construction of scales to quantify LLMorphism across its dimensions (including beliefs about token prediction, statistical learning, introspective confabulation, and equating plausibility with truth).
• Individual Differences: Examination of whether technical acuity attenuates LLMorphism or if exposure alone increases its prevalence, along with mapping moderating effects for socioeconomic, educational, gender, and occupational variables.
• Experimental Manipulation: Assessment of causal effects of LLM exposure on subsequent judgments about human replaceability, competence, responsibility, and clinical assessment.
• Interventions: Design of prompts and educational modules emphasizing human-LLM disanalogies or reviving alternative metaphors to counteract reductive framings (Capraro, 6 May 2026).

This suggests that addressing LLMorphism is critical for preserving the complexity of human cognition and for maintaining the social and ethical practices underpinning education, healthcare, work, and law. The prevailing risk is not only anthropomorphizing machines, but also diminishing the ascription of mind and agency to humans.