Separating Human Projection from Machine Cognition Is Key to LLM Understanding

Anthropomorphism has become a defining lens for how people interact with large language models (LLMs) and how researchers test them, often treating these systems as if they share human traits such as morality, personality, social intelligence, and even consciousness. But as models grow more capable, the debate intensifies: do LLMs possess genuine human-like understanding, or are they sophisticated pattern matchers?

Support for “genuine understanding” comes from mechanistic work suggesting that LLMs can build internal representations grounded in perceptual structure, despite operating on text alone. Related analyses also report introspective-like access to internal states and evidence that multi-step planning can occur before generating final responses—for example, when producing structured outputs such as poems.

Skeptics counter with a different diagnostic: brittleness. They argue that small changes—paraphrases that preserve meaning, or irrelevant additions—can trigger sharp accuracy losses. Moreover, LLMs often struggle with logic tasks that lack training-like templates, and chain-of-thought benefits can appear to mix reasoning with memorized fragments from training data.

The article proposes experientialism as a unifying framework to dissolve the dichotomy. Instead of assuming that understanding must mirror either objective reality or unconstrained imagination, experientialism views cognition as arising from interactions between an agent and its environment. Under an extended version of this idea, meaning is constructed by a system’s internal representational machinery while being constrained by the environment it learns from.

This view aligns with Bayesian cognition: an agent’s internal model is separated from the outside world by an interaction boundary (a Markov blanket). Lacking direct access, it must infer hidden causes of sensory-like inputs. Crucially, the internal model’s structure depends both on environmental structure and on the representational capacity of the agent itself.

Applying this to LLMs suggests a comparative but non-identical picture of cognition. For humans, embodied sensorimotor experience supports embodied cognition. For LLMs, the Transformer’s training corpus plays an analogous role, shaping how internal representations are constructed—an insight that reframes the question from “Does it understand?” to “How does it construct meaning under its constraints?”

A key empirical example comes from recent work on temporal cognition. In a similarity-judgment task spanning years from 1525 to 2524, large models reportedly develop a subjective temporal reference point and follow a Weber–Fechner-like scaling pattern. Neural and representational analyses are described as showing logarithmic temporal coding, hierarchical abstraction relative to the reference, and non-linear temporal structure embedded in the training data.

Taken together, the proposal reframes LLMs as “the other mind”: not a lesser version of human cognition, but a different cognitive construction process. The danger, the article warns, is treating model outputs as direct surrogates for human behavioral data without verifying cognitive equivalence.

Instead, it advocates task-based cognitive comparisons that test convergence and divergence directly. As LLMs enter high-stakes domains, machine experientialism argues for non-human-centric cognitive science—understanding how internal reality is built, predicting where divergence will emerge, and intervening to mitigate risks.