LLMorphism: When humans come to see themselves as language models

Technical Summary of "LLMorphism: When humans come to see themselves as language models"

Problem Statement
The paper identifies a novel cognitive bias termed LLMorphism: the biased belief that human cognition operates like a large language model (LLM). While extensive literature addresses the anthropomorphization of machines (attributing human minds to AI), the author argues that the rise of conversational LLMs creates a "reverse inference." As artificial systems produce fluent, human-like language, observers may increasingly infer that human thought itself is merely a form of LLM-like generation, prediction, and pattern completion. This bias is problematic because similarity in linguistic output does not entail similarity in cognitive architecture. The paper posits that this conflation risks collapsing critical distinctions between speaking and understanding, fluency and knowledge, and generation and judgment, potentially leading to the attribution of "too little mind" to humans.

Methodology and Theoretical Framework
This work is a theoretical conceptual paper rather than an empirical study. It does not present new experimental data, psychometric scales, or field trials. Instead, it employs a conceptual analysis to:

1. Define and Delineate: Introduce "LLMorphism" as a specific construct and distinguish it from adjacent concepts such as anthropomorphism, mechanomorphism, computationalism, dehumanization, objectification, and predictive-processing theories of mind.
2. Propose Mechanisms: Theoretically outline two psychological mechanisms through which this bias may spread:
   • Analogical Transfer: Based on Gentner's structure-mapping theory, where the observable similarity in linguistic output leads observers to align the relational structures of LLMs and humans, projecting LLM features (e.g., next-token prediction) onto human cognition.
   • Metaphorical Availability: Drawing on Conceptual Metaphor Theory (Lakoff & Johnson), the paper argues that LLM-specific vocabulary (e.g., "prompting," "hallucination," "training data") is migrating into cultural discourse, becoming the dominant source domain for conceptualizing human thought, memory, and creativity.
3. Analyze Implications: Theoretically map out potential societal consequences across work, education, healthcare, and moral responsibility, framing these as hypotheses rather than established facts.
4. Identify Boundary Conditions: Discuss factors that may moderate the spread of LLMorphism, such as attention to disanalogies, competing metaphors, and professional experiences that emphasize embodied, non-verbal human interaction.

Key Contributions

• Conceptual Innovation: The paper introduces "LLMorphism" as a distinct form of mechanomorphism. Unlike general mechanomorphism (viewing humans as machines), LLMorphism is specific to the statistical, language-generation architecture of LLMs. It is narrower than general mechanomorphism but potentially more consequential due to the centrality of language in social cognition.
• Reversal of Heuristics: The author reframes the anthropomorphism debate. While the public focus is on whether we attribute too much mind to machines, LLMorphism suggests the inverse problem: whether we are beginning to attribute too little mind, agency, and grounding to humans by reducing them to output generators.
• Differentiation from Existing Constructs: The paper rigorously distinguishes LLMorphism from:
  • Computationalism: Which views cognition as symbol manipulation but does not necessarily imply LLM-like text generation.
  • Dehumanization: Which involves moral exclusion; LLMorphism is a representational bias that may occur without hostility.
  • Predictive Processing: Which is often embodied and action-oriented, whereas LLMorphism implies a disembodied, text-centric view of cognition.
• Identification of Societal Pathways: The paper outlines five specific pathways through which LLMorphism could reshape social judgment:
  1. Replaceability: Viewing humans as output generators justifies automation and labor substitution.
  2. Fluency: Mistaking sophisticated text production for genuine expertise or understanding (e.g., in education and professional advice).
  3. Agency-Thinning: Redefining human action as "pattern completion," thereby eroding concepts of reasons, accountability, and moral responsibility.
  4. Disembodiment: Over-weighting verbal reports in fields like healthcare while ignoring nonverbal, affective, and situational cues.
  5. Epistemic Shift: Replacing the evaluation of grounded, justified claims with an evaluation of coherence and plausibility (epistemia).

Results and Claims
As a theoretical paper, it does not report empirical results. Instead, it presents a set of hypotheses regarding the prevalence and impact of LLMorphism. The author claims that:

• The bias is psychologically available due to the fluency of modern LLMs.
• The bias is logically invalid because linguistic output similarity does not imply cognitive architecture similarity.
• The bias is distinct from, though related to, other forms of mechanomorphism and dehumanization.
• The spread of this bias is not inevitable; it is subject to boundary conditions, including the salience of disanalogies (e.g., embodiment, emotion) and the availability of alternative metaphors (e.g., religious or humanistic worldviews).

Significance
The paper concludes that the public and academic debate on AI has focused primarily on the "anthropomorphization" of machines (giving them too much mind). The significance of this work lies in highlighting the "other half" of the problem: the potential de-grounding of human cognition. By suggesting that humans are beginning to view themselves through the lens of the tools they created, the paper warns of a shift where human agency, moral responsibility, and embodied experience are undervalued in favor of a model of cognition based on statistical prediction and text generation. The author calls for future research to develop psychometric scales to measure LLMorphism and to empirically test its effects on social judgments in domains such as labor, education, and healthcare.