Papers
Topics
Authors
Recent
Search
2000 character limit reached

LLMorphism: When humans come to see themselves as language models

Published 6 May 2026 in cs.CY | (2605.05419v1)

Abstract: LLMorphism is the biased belief that human cognition works like a LLM. I argue that the rise of conversational LLMs may make this bias increasingly psychologically available. When artificial systems produce human-like language, people may draw a reverse inference: if LLMs can speak like humans, perhaps humans think like LLMs. This inference is biased because similarity at the level of linguistic output does not imply similarity in cognitive architecture. Yet, LLMorphism may spread through two mechanisms: analogical transfer, whereby features of LLMs are projected onto humans, and metaphorical availability, whereby LLM vocabulary becomes a culturally salient vocabulary for describing thought. I distinguish LLMorphism from mechanomorphism, anthropomorphism, computationalism, dehumanization, objectification, and predictive-processing theories of mind. I outline its implications for work, education, responsibility, healthcare, communication, creativity, and human dignity, while also discussing boundary conditions and forms of resistance. I conclude that 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.

Authors (1)

Summary

  • The paper introduces LLMorphism as a novel bias where human cognition is compared to large language models.
  • It distinguishes LLMorphism from similar constructs using analogical transfer and metaphorical framing, highlighting societal and epistemic implications.
  • The study warns that equating language fluency with mechanistic output risks eroding genuine understanding of human agency and expertise.

LLMorphism: The Reversal of Anthropomorphic Attribution in the Era of LLMs

Introduction

The paper "LLMorphism: When humans come to see themselves as LLMs" (2605.05419) introduces and rigorously articulates the concept of "LLMorphism": the biased belief that human cognition operates analogously to LLMs. The author positions LLMorphism as a newly salient bias resulting from the rapid integration of conversational LLMs into various domains of human interaction. By critically distinguishing LLMorphism from other related constructs, the author addresses a consequential inversion within AI discourse—suggesting that the proliferation of human-like language output by LLMs may engender not only anthropomorphic projections onto machines, but a conceptual transfer wherein humans begin to be seen (and see themselves) as mechanistically LLM-like.

Anthropomorphization and Its Inversion

A foundational observation is that humans have long relied on a heuristic: fluent language use is indicative of a thinking mind. LLMs, by virtue of producing context-sensitive and coherent linguistic output, are positioned to elicit anthropomorphic attributions. Prior research has identified relatively shallow dialogue systems such as ELIZA as sufficient to trigger perceived human-likeness; current LLMs dramatically amplify this effect due to their scale and fluency.

However, the paper's central contribution is to unpack a theoretically and culturally significant reversal: as LLMs achieve greater linguistic sophistication, observers may analogically invert the anthropomorphic heuristic—contending that, if LLMs produce human-like language, perhaps human cognition itself is LLM-like. The author identifies two mutually reinforcing mechanisms: analogical transfer (structural mapping from LLMs to human cognition) and metaphorical availability (LLM-derived terminology infiltrating lay and academic discourse about thought).

Defining LLMorphism and Distinguishing Adjacent Constructs

The paper differentiates LLMorphism from related but distinct explanatory frames:

  • Mechanomorphism: Attribution of machine-like qualities to humans, but not necessarily predicated on linguistic output or LLM architectures.
  • Computationalism: The general view of the mind as information processing, foregrounding rule-based symbolic manipulation rather than probabilistic, lexical generation.
  • Dehumanization and Objectification: Modes of diminishing or denying humanness or subjectivity, typically with explicit moral connotations, whereas LLMorphism’s principal effect is representational.
  • Predictive Processing: Bayesian or generative models of cognition, which do not imply reducibility to LLM-like generative architectures or output-focused models.

LLMorphism, as conceptualized, is more specific: it is the (often tacit) stance that the observable similarity between LLM output and human language production justifies model-theoretic interpretations of human thought.

Risks and Societal Implications

The author systematically outlines hypothesized pathways through which LLMorphism, if sufficiently culturally entrenched, could inflect social, epistemic, and institutional domains:

  • Replaceability Mechanism: Viewing humans primarily as output-generating systems increases perceived replaceability and potentially legitimizes automation-driven labor displacement. This aligns with trends in algorithmic management where human value is operationalized via measurable outputs.
  • Fluency Mechanism: Institutions might misidentify textual fluency with functional expertise, undermining the recognition and transmission of tacit knowledge, situated understanding, and disciplinary rigor.
  • Agency-Thinning Mechanism: Describing human behavior in terms of input-output mappings and pattern completion may erode attributions of moral and legal responsibility, accountability, and the interpersonal foundations of trust and repair.
  • Disembodiment Mechanism: Overweighting verbal output risks diminishing attention to nonverbal, embodied, or contextual cues, with critical implications for healthcare, mental health assessment, and pedagogy, where lived experience and affect are irreducible to language.
  • Epistemic Mechanism: LLMorphism commodifies plausibility and coherence as substitutes for justification and truth, contributing to what the author terms "epistemia," a degraded epistemic environment where surface fluency occludes deeper evaluation.

Each pathway suggests that LLMorphism could retrench and normalize shallow conceptions of cognition, personhood, and expertise, with significant downstream impacts.

Moderating Factors and Potential for Resistance

Despite these risks, the paper identifies substantive moderators:

  • Salience of Disanalogies: Cognitive and educational interventions that foreground the differences in embodiment, affect, and agency between humans and LLMs can counteract the spread of LLMorphism.
  • Cultural and Professional Alternative Framings: Metaphorical pluralism, humanistic philosophies, religious frameworks, and professional experiences (e.g., in nursing, psychotherapy, early childhood education) that emphasize irreducible aspects of human experience function as bulwarks against LLM-like construals.
  • Exposure and Technical Literacy: The paper speculates that increased exposure to or technical understanding of LLMs may either amplify or attenuate LLMorphism, depending on how users internalize the underlying architecture and limitations of these models.

Future Directions

The establishment of LLMorphism as a distinct construct enables several lines of empirical and theoretical inquiry:

  • Psychometric Construct Validation: Development of multi-dimensional scales to measure LLMorphic beliefs in the population and their demarcation from adjacent constructs.
  • Individual Difference Research: Investigation of the factors that predict susceptibility to LLMorphism, such as education, occupation, and societal context.
  • Causal and Field Studies: Experimental designs to test whether exposure to LLMs and LLM-framed discourse causally influences conceptions of human cognition, expertise, and agency.

Conclusion

The introduction of LLMorphism constitutes a significant reframing of the impact of conversational LLMs on cultural understandings of human cognition and personhood. By systematically articulating its definition, mechanisms, differentiation from related concepts, and potential for societal consequences, the paper establishes a critical agenda for sociotechnical research. The recognition of LLMorphism highlights the necessity of epistemic and ethical vigilance—not solely regarding oversubscription of mind to machines, but in resisting reductive reinterpretations of the human mind itself.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

There was an error generating the whiteboard.

Explain it Like I'm 14

What is this paper about?

This paper introduces a new idea called “LLMorphism.” An LLM is a LLM—think of a super-powered autocomplete that has read lots of text and predicts what words are likely to come next. LLMorphism is the belief that human thinking works the same way an LLM works.

The author warns that because chatbots now sound very human, some people might flip their thinking: instead of just asking “Are machines becoming more human?”, they might start asking “Maybe humans think like these machines.” The paper argues this is a mistake, because sounding similar doesn’t mean working the same on the inside.

What questions does it ask?

The paper sets out to answer a few simple questions:

  • What exactly is “LLMorphism,” and how could it spread?
  • How is LLMorphism different from other ideas like anthropomorphism (treating machines like people) or seeing people as “just machines” in general?
  • What could happen in everyday life if people start believing humans think like LLMs?
  • Are there ways this belief might be resisted or kept in check?
  • What should future research look at to measure and understand this belief?

How did the author approach it?

This is not an experiment or a data study; it’s a concept paper. The author:

  • Reviews past research about how people think about minds—both human and machine.
  • Uses clear examples and theories from psychology and philosophy to explain how a new bias could form.
  • Outlines testable predictions and suggests future studies to measure how common this belief is and what effects it might have.

If you think of science like building a map, this paper draws a new part of the map and shows where to look next, rather than measuring every hill and tree itself.

What are the key ideas and how do they work?

To make the ideas easy, imagine two songs that sound alike. Even if the tunes are similar, they might have been made with totally different instruments, skills, and goals. Likewise, human speech and chatbot text can look similar but come from very different “engines.”

Here are the main ideas the paper explains:

  • What LLMs do
    • LLMs are trained on tons of text and are very good at guessing the next word. That makes them great at producing fluent, well-formed sentences.
    • But they don’t have bodies, feelings, or lived experience. They don’t have skin in the game—no real-world consequences for being wrong. Their strength is form and fluency, not grounded understanding.
  • What LLMorphism is
    • LLMorphism is the biased belief that because LLMs talk like us, we must think like them. It treats similarity in how the output looks (fluent language) as proof that the inner process is the same—which it isn’t.
  • How this belief could spread
    • Analogical transfer: When two things seem similar on the surface, people often assume they’re similar deeper down. If both humans and LLMs can write good essays, someone might wrongly conclude they “think” the same way.
    • Metaphorical availability: As LLM words go mainstream—like “prompt,” “hallucinate,” “training data,” “next-token prediction”—people start using them to describe human thoughts and behavior. The language shapes how we see ourselves.
  • How it differs from other beliefs
    • Not anthropomorphism: That’s when we treat machines like people. LLMorphism is the reverse—treating people like machines, specifically like LLMs.
    • Not just “humans are computers”: That older idea (computationalism) says thinking is information processing. LLMorphism is narrower: it says human thinking is like an LLM’s text prediction, which is a specific kind of processing.
    • Not necessarily dehumanization or objectification: Those involve treating people as less worthy or as tools. LLMorphism can overlap but is mainly about a wrong picture of how minds work.
    • Not the same as “predictive processing” theories in neuroscience: Those theories say our brains make predictions, but still emphasize bodies, action, and real-world grounding—not just text generation.

What does the paper find, and why is it important?

Because this is a concept paper, the “findings” are arguments and hypotheses, not experimental results. The author’s main points are:

  • Language output isn’t the same as thought
    • LLMs are fluent, but fluency isn’t proof of understanding. Humans use language rooted in body, emotion, memory, social life, and accountability. LLMs use patterns in text.
  • LLMorphism can blur vital distinctions
    • If we mix up speaking with understanding, or effortless fluency with expertise, we risk judging people the wrong way and trusting the wrong things.
  • Possible real-world effects (things to watch for)
    • Work and replaceability: If people are seen as just “output generators,” replacing them with machines feels easier and more justified.
    • Education and expertise: Schools and employers might mistake slick writing for deep understanding. We could start valuing smooth answers over real knowledge.
    • Responsibility and agency: If we describe human choices as “inputs leading to outputs,” we may weaken ideas like blame, apology, trust, and moral responsibility.
    • Healthcare and embodiment: Doctors might pay too much attention to what patients say and not enough to how they look, move, or feel—especially risky in mental health.
    • Truth and knowledge: People might judge ideas by how plausible they sound rather than whether they’re well-evidenced and true.

These points matter because they touch core parts of society—jobs, schools, law, and health—and they protect what’s special about being human: understanding, caring, choosing, and being accountable.

What could limit or resist LLMorphism?

The paper also suggests forces that could push back:

  • Noticing the differences: Showing clear, concrete ways humans and LLMs differ can weaken the bad analogy.
  • Competing metaphors: If we describe minds with richer images—like “a living body in the world” or “a storyteller in a community”—the LLM metaphor has less grip.
  • Lived experience and certain jobs: Work that centers on care and real presence (nursing, therapy, early childhood education) highlights body language, emotion, and relationships—things LLMs don’t have. The humanities and qualitative social sciences can help here too.

What does the paper suggest for future research?

The author proposes three main next steps:

  • Measure it: Create a survey or scale to see how strongly people hold LLMorphic beliefs, and whether those beliefs cluster into parts (for example, beliefs about language, learning, creativity, or truth-seeking).
  • Find who’s most susceptible: Does more exposure to LLMs increase or decrease this belief? How do education, job type, or other factors matter?
  • Test the consequences: Run experiments to see if stronger LLMorphic beliefs actually lead people to overvalue fluency, underrate expertise, or downplay responsibility.

Bottom line: Why does this matter?

The public debate often focuses on one side: “Are we giving machines too much mind?” This paper says there’s another half we’re missing: “Are we taking too much mind away from humans?” If we start treating people like word machines, we risk confusing smooth talk with real understanding, swapping caring for metrics, and weakening the bonds of responsibility and trust that hold communities together.

In short, LLMs can be useful tools—but we shouldn’t let their style of talking rewrite what it means to think, know, and be human.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Unresolved knowledge gaps, limitations, and open questions

Below is a single, concrete list of gaps and open questions the paper leaves unresolved that future researchers could act on:

  • Prevalence: No empirical estimates of how widespread LLMorphic beliefs are across populations, cultures, age groups, or professions.
  • Construct operationalization: Lack of a validated, reliable psychometric instrument to measure LLMorphism (including item pools, factor structure, and scoring).
  • Dimensionality: Unclear whether LLMorphism is unitary or multidimensional (e.g., language-generation beliefs, creativity/learning beliefs, introspection/confabulation beliefs, truth-seeking beliefs), and how these subcomponents covary.
  • Discriminant validity: No evidence that a putative LLMorphism scale can be distinguished from adjacent constructs (anthropomorphism, mechanomorphism, dehumanization, computationalism, techno-optimism/pessimism, need for cognition, authoritarianism).
  • Measurement invariance: Unknown whether LLMorphism measures function equivalently across languages, cultures, and levels of AI literacy; translation/adaptation challenges for terms like “next-token prediction.”
  • Temporal stability: No data on the test–retest reliability of LLMorphism or its developmental trajectory (e.g., adolescence vs adulthood).
  • Implicit vs explicit endorsement: Absence of methods to capture implicit LLMorphic tendencies (e.g., IAT-style tasks, reaction-time vignettes) versus explicit self-reports.
  • Behavioral validation: No link established between LLMorphism scores and behavioral outcomes (e.g., hiring choices, grading, clinical decisions) beyond attitudinal measures.
  • Causal inference: No experiments manipulating exposure to LLMs or LLM-framed metaphors to test whether LLMorphism increases and for how long effects persist.
  • Mechanism parsing: The relative contribution of analogical transfer vs metaphorical availability remains untested; no mediational models linking exposure → mechanism → belief change.
  • Thresholds and triggers: Unknown conditions under which linguistic similarity is sufficient to trigger LLMorphic inferences (e.g., model fluency level, persona cues, domain context).
  • Dose–response: No evidence on whether LLMorphism scales with intensity, recency, or type of LLM interaction (professional vs casual; text-only vs multimodal/agentic systems).
  • Counterfactual metaphors: No tests of competing metaphors (e.g., embodied, narrative, ecological) as interventions to attenuate LLMorphism or their durability and side effects.
  • AI literacy interventions: Limited guidance on which training components (architecture education, limitations, embodiment gaps) reduce LLMorphism without inducing reactance or over-skepticism.
  • Boundary conditions: Proposed moderators (e.g., caregiving roles, humanities training, religious/humanistic worldviews) are hypothesized but not empirically verified.
  • Cross-cultural variation: No cross-national/cultural studies assessing whether societies with different dominant metaphors or tech discourses exhibit different LLMorphism levels.
  • Longitudinal dynamics: Unclear how LLMorphism evolves as models improve (e.g., tool-use, embodiment via robots); whether novelty effects wane or beliefs entrench.
  • Bidirectional coupling: The reciprocal influence between anthropomorphizing machines and LLMorphizing humans is posited but not modeled or tested as a dynamic feedback loop.
  • Domain specificity: No evidence on whether LLMorphism varies across domains (law, medicine, education, creative arts, management) or interacts with domain norms of evidence.
  • Replaceability mechanism: Lacks field evidence that LLMorphism independently predicts automation decisions beyond economic/managerial drivers; need for designs that isolate its unique effect.
  • Fluency–expertise confusion: No experiments showing that LLMorphism increases reliance on verbal fluency as a proxy for understanding or devalues tacit/situated expertise in evaluative tasks.
  • Agency-thinning: No measures linking LLMorphism to reduced attributions of intent, negligence, accountability, or altered punishment/forgiveness judgments in legal/moral vignettes.
  • Disembodiment in healthcare: No clinical simulations or field studies testing whether LLMorphism reduces attention to nonverbal/embodied cues, with differential impacts on vulnerable groups (e.g., non-native speakers, cognitively impaired patients).
  • Epistemic consequences: The link between LLMorphism and “epistemia” (plausibility-over-evidence evaluation) is hypothesized but lacks experimental and corpus-based confirmation.
  • Media/discourse diffusion: No corpus analyses tracking the spread of LLM-derived vocabulary in public discourse, its correlates, and the agents (media, industry) amplifying it.
  • Stakeholder mapping: No identification of institutions or practices that most propagate LLMorphic frames (e.g., UI design, policy documents) or where mitigation would be most impactful.
  • Policy levers: Absence of evaluated interventions (disclosures, interface cues, professional guidelines) to curb LLMorphism in high-stakes settings (courts, hospitals, education).
  • Equity implications: Unexplored whether LLMorphism exacerbates biases against groups whose embodied or discourse cues diverge from “fluent” norms (accented speakers, neurodivergent individuals).
  • Developmental effects: No research on how early exposure to LLMs shapes children’s concepts of mind, agency, learning, and creativity.
  • Alternative explanations: Predicted societal effects might stem from pre-existing dehumanization, managerial ideologies, or economic incentives; no designs disentangle LLMorphism’s unique variance.
  • Formal models: No causal diagrams or computational models specifying pathways and feedback loops from LLM exposure to beliefs to social outcomes, to guide hypothesis testing.
  • Generalization beyond language: Unclear whether similar biases arise from other generative AI (images, audio, agents) and whether a broader “GenAI-morphism” construct is warranted.
  • Ethical/normative criteria: The paper labels LLMorphism as “biased” but lacks explicit criteria for when LLM metaphors are informative heuristics versus distortive reductions.
  • Open science plans: No proposed shared datasets, preregistered protocols, or benchmark tasks to enable cumulative, comparable research on LLMorphism.
View Paper Prompt  View All Prompts 

Practical Applications

Immediate Applications

Below are specific, deployable practices and tools that organizations and individuals can adopt now to mitigate LLMorphism’s risks and leverage its insights.

  • Industry (management/HR): Redesign performance evaluation to value judgment, context, and tacit work
    • What: Update KPIs and review templates to include “reason-giving,” context awareness, and client/stakeholder outcomes, not just text artifacts (docs, tickets, emails).
    • Tools/workflows: “Reason and evidence” fields in report templates; decision logs that capture alternatives considered; peer review of judgment on high-stakes outputs.
    • Assumptions/dependencies: Managerial buy-in; slight time overhead; willingness to revise incentive schemes.
  • Industry (product/UX for AI tools): Calibrate user mental models to reduce LLMorphism
    • What: Add UX cues that discourage treating humans as output generators or LLMs as agents (e.g., depersonalized model voice, disclaimers like “plausible text, not verified knowledge,” links to sources).
    • Tools/workflows: Model UIs with built-in “evidence required” prompts; toggles to reveal training-data limits; “fluency ≠ expertise” banners in high-stakes contexts.
    • Assumptions/dependencies: Product teams’ willingness to trade slight friction for safety; legal review of copy.
  • Education (K–12 and higher ed): Assessment focused on process, grounding, and accountability
    • What: Grade for reasoning process, data use, and oral defense (vivas, whiteboard derivations) rather than final text alone.
    • Tools/workflows: Process portfolios/lab notebooks; “How do you know?” prompts; reflective memos on uncertainty; AI-use declarations.
    • Assumptions/dependencies: Faculty development; rubric updates; time allocation for orals or checkpoints.
  • Education (AI literacy): Teach fluency vs understanding, generation vs justification
    • What: Short modules that contrast fluent answers with grounded knowledge and explain how LLMs differ from human cognition.
    • Tools/workflows: Micro-courses integrated into first-year seminars, CS intros, and writing centers; example-based exercises showing “plausible-but-wrong.”
    • Assumptions/dependencies: Curriculum owners’ support; materials adapted to discipline.
  • Healthcare (clinical practice): Make embodied and nonverbal cues a first-class data source
    • What: Update intake and progress notes to capture nonverbal observations and context, not just textual complaints.
    • Tools/workflows: EHR templates with structured “Nonverbal/behavioral observations” fields; standardized patient checklists; reminders for interpreter use.
    • Assumptions/dependencies: EHR customization; clinician training; attention to time burden and privacy.
  • Healthcare (mental health): Balance verbal report with behavioral indicators
    • What: Triage and session templates that prompt for psychomotor signs, affect, and alliance quality alongside patient narratives.
    • Tools/workflows: Session checklists; supervision prompts to discuss nonverbal data; policies to avoid LLM-only documentation assistance in diagnostics.
    • Assumptions/dependencies: Clinical governance approval; training; documentation standards.
  • Law and public sector: Preserve responsibility through reason-giving and attestation
    • What: Require human-signed justifications for decisions; prohibit filing LLM-produced text without attorney/official attestation of due diligence.
    • Tools/workflows: Decision memos with case-linked reasons; mandatory “human accountability” sections in briefs and policy drafts.
    • Assumptions/dependencies: Court/agency rules; enforcement mechanisms; professional liability norms.
  • Media and communications: Separate plausibility from verification
    • What: Editorial standards that prevent equating fluent copy with confirmed facts; visible verification steps.
    • Tools/workflows: “Verification checklist” blocks in CMS; source-link requirements; labels when AI assistance is used.
    • Assumptions/dependencies: Editorial policy changes; training; audience education.
  • Corporate governance/procurement: LLMorphism-aware AI-use policies
    • What: Policies that forbid replacing human judgment with AI-generated text for high-stakes decisions and define where human review is mandatory.
    • Tools/workflows: “Human-in-the-loop” gates; risk-tiered approval matrices; procurement checklists assessing reliance on fluency proxies.
    • Assumptions/dependencies: Board/legal support; vendor compliance; auditability.
  • Workplace training (cross-sector): Metaphor hygiene and prompt hygiene
    • What: Train staff to avoid LLM-centric metaphors for human cognition in policy and evaluation (e.g., “hallucination” for people), and to avoid prompts that invite overtrust.
    • Tools/workflows: Microlearning modules; style guides; automated document linting that flags LLMorphic phrasing about people.
    • Assumptions/dependencies: Communication teams’ adoption; tool integration in writing platforms.
  • Research (immediate start): Rapid pilot instruments for LLMorphism
    • What: Create short, provisional survey items to gauge LLMorphic beliefs in organizations or classrooms and correlate with decisions (e.g., willingness to automate).
    • Tools/workflows: 5–10 item pilot scale; pre/post measures around AI exposure; simple experiments showing shifts after LLM interaction.
    • Assumptions/dependencies: IRB where needed; small-sample tolerance; iterative refinement.
  • Daily life (personal and parenting): “Grounding check” before acting on fluent advice
    • What: Simple heuristics to ask for evidence, alternatives, and consequences before accepting plausible-sounding guidance.
    • Tools/workflows: Pocket checklists; family discussions distinguishing saying vs knowing; journaling prompts about reasons vs outputs.
    • Assumptions/dependencies: Willingness to adopt heuristics; basic AI/media literacy.

Long-Term Applications

These opportunities require additional research, validation, scaling, or policy development to reach robust deployment.

  • Cross-cultural psychometric scale of LLMorphism
    • What: Develop and validate a multidimensional scale (language-generation beliefs, creativity-as-recombination, introspection-as-confabulation, truth-seeking orientation).
    • Tools/workflows: Item banks, factor analyses, norms by sector/culture; open datasets.
    • Dependencies: Multi-site studies; funding; translations and measurement invariance.
  • Organizational “LLMorphism Risk Index”
    • What: An audit framework that scores units on overreliance on fluency, thin agency attributions, and replaceability framing.
    • Tools/workflows: Surveys, text analytics of policy/communications, decision-sampling audits; dashboarding.
    • Dependencies: Access to internal documents; change-management support; data governance.
  • NLP detectors of LLMorphic framing in institutional documents
    • What: Models that flag language equating human cognition with LLM-like generation or substituting plausibility for justification.
    • Tools/workflows: Policy-document scanners; IDE plugins for legal/clinical drafting.
    • Dependencies: Annotated corpora; precision/recall trade-offs; organizational acceptance.
  • Education (systemic): Epistemic virtues curriculum
    • What: K–12 and university curricula emphasizing justification, accountability, uncertainty management, and embodied perspectives across disciplines.
    • Tools/workflows: Cross-disciplinary modules; capstones requiring reason-giving and stakeholder engagement; assessment standards accrediting process quality.
    • Dependencies: Standards bodies; teacher development; longitudinal evaluation.
  • Healthcare (technology): Multimodal clinical support that integrates nonverbal data ethically
    • What: Decision support that prompts clinicians with structured observations from voice, movement, and interaction patterns, combined with patient narratives.
    • Tools/workflows: EHR-integrated multimodal inputs; consent and privacy safeguards; bias and fairness audits.
    • Dependencies: Regulatory approval; robust evidence of utility; infrastructure for secure data capture.
  • Labor and AI governance: Standards and regulation to prevent “fluency-based automation”
    • What: Policies requiring demonstration that tasks automated by LLMs do not depend primarily on tacit, contextual, or accountability-laden human skills.
    • Tools/workflows: Task analyses; human-over-the-loop requirements; certification schemes for high-stakes deployments.
    • Dependencies: Legislative processes; sector regulators; enforcement capacity.
  • Legal doctrine and professional norms: Responsibility-preserving AI practice
    • What: Revisions to professional conduct rules to require documented human reasoning for decisions in law, medicine, finance, and public administration.
    • Tools/workflows: Attestation protocols; audit trails linking decisions to reasons and evidence; sanctions for “explanation displacement.”
    • Dependencies: Bar/medical boards; insurers; jurisprudence development.
  • Industry (expertise assurance): Fluency-vs-grounding audit in expert workflows
    • What: Third-party audits of expert functions (e.g., compliance, research, safety) to ensure conclusions are based on validated models/data, not just polished prose.
    • Tools/workflows: “Grounding score” metrics; replication or adversarial review steps; documented uncertainty quantification.
    • Dependencies: Market demand; cost tolerance; access to data and methods.
  • Public communication and media: Verification-forward platforms
    • What: News and knowledge platforms that foreground verification pathways (methods, sources, accountability chains) over stylistic polish.
    • Tools/workflows: Source maps; machine-assisted verification aids; audience education widgets.
    • Dependencies: Business models; audience incentives; platform partnerships.
  • Workforce development: New roles and training for “grounding stewards”
    • What: Roles dedicated to ensuring decisions retain embodied/contextual grounding and proper attribution of agency and responsibility.
    • Tools/workflows: Checkpoint reviews in high-stakes pipelines; red-teaming against “plausibility creep.”
    • Dependencies: Budget and headcount; role clarity; authority to intervene.
  • Cross-sector experiments on exposure and susceptibility
    • What: Large-scale studies testing whether LLM exposure increases or decreases LLMorphism, moderated by AI literacy, profession, and worldview.
    • Tools/workflows: Pre-registered RCTs; longitudinal panels; behavioral outcomes (e.g., automation choices, empathy measures).
    • Dependencies: Funding; participant recruitment; data-sharing agreements.
  • Cultural counter-metaphors initiative
    • What: Develop and disseminate alternative metaphors for human cognition (embodiment, practice, care) to compete with LLM-centric vocabulary.
    • Tools/workflows: Public campaigns; interdisciplinary humanities/science collaborations; toolkits for educators and managers.
    • Dependencies: Philanthropic and public support; cultural institutions; evaluation frameworks.

Assumptions and dependencies common across applications:

  • Prevalence of LLMorphism may vary; initial diagnostics are needed to justify interventions.
  • Adoption requires leadership endorsement, training capacity, and minor productivity trade-offs.
  • Privacy, consent, and bias concerns are critical for any multimodal or auditing tooling.
  • Regulatory and accreditation changes take time; interim voluntary standards can bridge the gap.
  • Sectoral contexts differ: caregiving, humanities, and early childhood education may naturally resist LLMorphism and can be leveraged as exemplars and training partners.
View Paper Prompt  View All Prompts 

Glossary

  • Agency-thinning mechanism: A hypothesized process where reframing human action as input-output generation weakens perceived agency and responsibility. "A third pathway is an agency-thinning mechanism."
  • Algorithmic management: The governance of workers through metrics and automated systems that track and optimize productivity. "and with algorithmic management, where workers are governed through quantifiable productivity traces (Kellogg et al., 2020)."
  • Analogical transfer: The projection of relational structure from a known domain to a new domain to make inferences about it. "Analogical transfer involves mapping relational structure from one domain to another (Gentner, 1983)."
  • Animalistic dehumanization: A subtype of dehumanization that portrays people as animal-like, denying uniquely human qualities. "Haslam's model distinguishes animalistic dehumanization, which represents others as animal-like, from mechanistic dehumanization, which represents others as objects, automata, or machines (Haslam, 2006; Haslam & Loughnan, 2014)."
  • Anthropomorphization: The act of attributing human-like properties or mental states to non-human entities. "The overapplication of this heuristic lies at the basis of the anthropomorphization of LLMs."
  • Anthropomorphism: The tendency to ascribe human mental states and capacities to non-human agents. "In general, anthropomorphism refers to the tendency to attribute human-like mental states, intentions, and capacities to non-human entities (Epley, Waytz, & Cacioppo, 2007)."
  • Bayesian theories of cognition: Frameworks positing that cognition involves probabilistic inference and updating beliefs with evidence. "Predictive processing holds that the brain continuously generates predictions about sensory input and updates internal models in light of prediction error (Clark, 2013; Friston, 2010; Hohwy, 2013). But predictive processing does not imply that humans are LLM-like, nor that human understanding is merely text generation. Indeed, many predictive-processing accounts are deeply embodied and action-oriented (Allen & Friston, 2018; Clark, 2015; Pezzulo et al., 2024)."
  • Computationalism: The view that mind and cognition are forms of information processing over representations and rules. "LLMorphism is different from computationalism, the broader thesis that cognition is a form of information processing."
  • Conceptual metaphor theory: The theory that people understand abstract concepts by mapping them onto familiar, concrete domains. "Conceptual metaphor theory holds that people understand abstract domains partly by mapping them onto more concrete or culturally salient source domains (Lakoff & Johnson, 2008)."
  • Dehumanization: The denial of humanness to individuals or groups, often resulting in moral exclusion or mistreatment. "Dehumanization involves the denial of humanness to individuals or groups."
  • Disembodiment mechanism: A hypothesized process where linguistic output is overemphasized at the expense of embodied and contextual cues. "A fourth pathway is a disembodiment mechanism."
  • Epistemia: A proposed condition in which linguistic plausibility replaces rigorous epistemic evaluation. "This may reinforce epistemia: a condition in which linguistic plausibility substitutes for epistemic evaluation (Loru et al., 2025; Quattrociocchi et al., 2025)."
  • Essentialist beliefs: Beliefs that human categories have deep, immutable essences that define them. "individuals who strongly endorse essentialist beliefs about human uniqueness (Haslam, Bastian & Bissett, 2005)"
  • Formal linguistic competence: The ability to produce grammatically correct and locally coherent language. "Consequently, LLMs are remarkably strong at formal linguistic competence: they can generate grammatical, coherent, and locally context- appropriate text."
  • Functional linguistic competence: The practical use of language in real-world contexts, relying on reasoning, knowledge, and social understanding. "However, they remain more limited with respect to functional linguistic competence, that is, the use and understanding of language in the world, which depends on extralinguistic capacities such as reasoning, world knowledge, situation modeling, and social cognition."
  • Intentional stance: A strategy of interpreting behavior by attributing beliefs, desires, and intentions to an entity. "people are especially prone to adopt the intentional stance and infer communicative agency from apparently meaningful behavior (Dennett, 1989)."
  • LLMorphism: The biased belief that human cognition functions like a LLM. "LLMorphism is the biased belief that human cognition works like a LLM."
  • Mechanistic dehumanization: A subtype of dehumanization that portrays people as objects or machines lacking agency and warmth. "Haslam's model distinguishes animalistic dehumanization, which represents others as animal-like, from mechanistic dehumanization, which represents others as objects, automata, or machines (Haslam, 2006; Haslam & Loughnan, 2014)."
  • Mechanomorphism: Attributing machine-like characteristics and processes to humans. "mechanomorphism treats the human as machine-like."
  • Metaphorical availability: The cultural accessibility of a metaphorical vocabulary that shapes how people conceptualize domains like thought. "metaphorical availability, whereby LLM vocabulary becomes a culturally salient vocabulary for describing thought."
  • Next-word prediction: Predicting the subsequent token in a sequence based on learned statistical patterns. "For this reason, functional performance often requires forms of augmentation beyond next-word prediction, such as specialized fine-tuning or coupling with external modules (Mahowald et al., 2024)."
  • Objectification: Treating a person as an object or instrument rather than as an autonomous subject. "Objectification involves treating a person as an object, tool, or instrument rather than as a full subject (Nussbaum, 1995)."
  • Patient-centred medicine: A medical approach that views illness as an embodied, contextual experience, not just reported symptoms. "patient-centred medicine treats illness not merely as verbally reportable information, but as an embodied and socially situated experience (Engel, 1977; Kleinman, 1988; Carel, 2016)."
  • Pattern completion: Generating outputs by filling in or completing learned patterns based on context. "what begins as a comparison between two forms of linguistic output may become a broader reinterpretation of human cognition as LLM-like generation, prediction, recombination, and pattern completion."
  • Predictive processing: A theory positing the brain continually predicts sensory input and updates models via prediction errors. "Predictive processing holds that the brain continuously generates predictions about sensory input and updates internal models in light of prediction error (Clark, 2013; Friston, 2010; Hohwy, 2013)."
  • Psychomotor retardation: A slowing of physical and emotional reactions, often observed in clinical depression. "behavioral and nonverbal signs such as psychomotor retardation, agitation, facial expression, vocal dynamics, and posture can provide clinically relevant information beyond verbal report (Dibeklioğlu et al., 2015)."
  • Relational pedagogy: An educational approach emphasizing relationships, attachment, and affect regulation as core to learning. "Early childhood education is organized around relational pedagogy, attachment, affect regulation, and development (Cliffe & Solvanson, 2023)."
  • Replaceability mechanism: A hypothesized pathway where viewing humans as output generators increases perceived substitutability by machines. "One possible pathway is a replaceability mechanism."
  • Reverse inference: Inferring underlying processes from observed outputs, often in a way that can be psychologically tempting but logically invalid. "a reverse inference becomes psychologically available"
  • Situated perspective: Understanding and using language from a contextually embedded, embodied point of view. "by embodied agents who use language from a situated perspective, with perception, action, affective stakes, communicative intentions, and responsibility for what they say"
  • Situation modeling: Constructing mental representations of real-world situations to interpret and use language effectively. "extralinguistic capacities such as reasoning, world knowledge, situation modeling, and social cognition."
  • Social cognition: The mental processes involved in understanding others’ intentions, emotions, and social contexts. "extralinguistic capacities such as reasoning, world knowledge, situation modeling, and social cognition."
  • Structure-mapping theory: A theory of analogy stating that transfer depends on aligning relational structures between domains. "Analogical transfer, according to structure-mapping theory, requires that the source (LLM) and target (human cognition) be aligned on relevant relational predicates (Gentner & Markman, 1997)."
  • Structural alignment: The process of matching relational structures across two domains to support analogy. "Similarity and analogy depend on structural alignment: when two systems appear similar, observers may align their relational structures and project inferences from one domain to the other (Gentner & Markman, 1997)."
  • Tacit knowledge: Implicit, non-codified know-how that informs expert judgment and practice. "Expert judgment depends on tacit knowledge, situated interpretation, uncertainty management, disciplinary norms, and accountability within a practice (Polanyi, 1966; Collins & Evans, 2019)."
  • Therapeutic alliance: The collaborative relationship and bond between therapist and client that supports effective therapy. "Psychotherapy depends not only on verbal exchange, but also on embodied presence, nonverbal communication, and therapeutic alliance (Del Giacco et al., 2020; García et al., 2022)."
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 23 tweets with 715 likes about this paper.

Sign Up for Free