Is understanding an anthropocentric concept?

Discussions of whether large language models (LLMs) understand what they reason about typically proceed as though “understanding” names a single, well-defined target that a system either possesses or lacks. This is a curious starting point, since understanding resists definition even in its familiar human case: the term is used constantly and defined rarely. Transposed onto LLMs, the difficulty is obscured rather than resolved by a change of register. Debates about machine understanding are conducted largely in the idiom of benchmark performance: a system is said to understand a concept to the degree that it produces the outputs a competent human would produce. This substitution carries an underexamined cost, because the behaviours counted as evidence, the concepts checked for, and the tasks taken to require understanding at all are, without exception, chosen and calibrated by humans for human cognitive architecture. The resulting picture of machine understanding is anthropocentric not as an incidental bias that better benchmark design could remove, but as a structural feature of the enterprise.

This paper develops that claim through a specific empirical case rather than through conceptual analysis alone, drawing on a companion project that evaluates conceptual understanding in LLMs on ConceptARC [Moskvichev et al., 2023], a benchmark built on Chollet’s Abstraction and Reasoning Corpus [Chollet, 2019] and organised around a fixed taxonomy of sixteen human-labelled concepts. We separate a model’s conceptual understanding, scored by a judge model reading only its reasoning trace, from whether its final answer was correct. Two features of the data motivate the argument: understanding and correctness dissociate substantially in both directions, and the concepts on which models score worst are concentrated in a single family, those requiring reasoning about depth, layering, and occlusion, precisely the family for which human cognition has the most direct embodied purchase and a token-based language model has the least.

From this pattern we draw three connected claims: that “concept understanding,” as standardly measured, is best interpreted as alignment with a human-authored taxonomy rather than as a general-purpose measure of internal structure; that the concepts on which this alignment breaks down track a deeper asymmetry in embodied access to the physical world, so that part of what evaluators call a failure of machine understanding is better described as a mismatch between a task’s cognitive demands and the kind of system performing it; and that the human case does not escape this problem, since the same labelling task given to human participants shows a structurally similar pattern of concept-specific confusion. None of this argues that current LLMs understand in the full philosophical sense; it argues that the received way of posing the question already contains a substantive, anthropocentric answer to what understanding is.

A useful starting point is the distinction, familiar from linguistics and reintroduced into comparative cognitive science by Firestone [2020], between competence and performance. A system may fail to produce an expected output not because it lacks the relevant underlying capacity, its competence, but because of incidental constraints on demonstrating that capacity in a given task format, its performance. Hu and Frank [2024] make a related point for language models, showing that evaluation methods with greater incidental task demands systematically depress measured performance, most for smaller, less capable models. A single scalar accuracy score therefore conflates whether a system has the relevant competence with whether the task allows that competence to be expressed. Standard benchmarks, ConceptARC among them, report only the scalar.

There is a second, deeper respect in which the standard evaluation of machine understanding is anthropocentric, concerning not the conditions under which competence is tested but the content of the competence tested for. ConceptARC, like most benchmarks aimed at general reasoning, defines its target concepts by reference to human perceptual categories: spatial relations salient to an upright, binocular organism, topological distinctions salient to an organism that manipulates bounded objects with hands, occlusion relations salient to an organism with stereoscopic depth perception. None is arbitrary from a human point of view, but all are chosen because they are carvings a human finds natural and easy to name. Once the target concepts are fixed this way, a low score can mean either that the system failed to construct adequate internal structure for the task, or that it constructed structure that solves the task through a route that does not project cleanly onto the human-named concept being checked for, so the score is better read as a measure of taxonomic alignment than of understanding as such.

Moskvichev et al. [2023] built ConceptARC’s taxonomy explicitly to test whether systems generalise the way humans do, and reported that human participants substantially outperform AI systems, a comparison informative only on the assumption that the taxonomy is a neutral coordinate system rather than one native to a single party. Mitchell [2024] argues that much of the vocabulary used to describe machine cognition is imported wholesale from folk psychology, obscuring rather than clarifying what is actually happening inside an artificial system, a caution Shevlin and Halina [2019] extend to psychological terms in AI generally. A taxonomy built from that vocabulary defines the object of measurement in human terms from the outset. Hewitt et al. [2025] offer a constructive response, arguing that interpretability is better framed as a communication problem between distinct conceptual vocabularies, requiring new terms for machine-native concepts rather than forcing machine behaviour into human categories.

One resource for a non-anthropocentric account comes from work treating intelligence, and by extension understanding, as inherently task-relative. Legg and Hutter [2007] define intelligence formally as an agent’s capacity to achieve goals across a wide range of environments, a definition silent on which environments and goals. Holm and Banerjee [2024] develop a related, explicitly pluralist picture: a heliocentric rather than geocentric view of intelligence, in which human cognition is one point among many rather than the fixed centre against which all others are measured.

Applied to understanding, this suggests three connected claims: that understanding claims about any system should be indexed to a task and an access channel rather than asserted as unqualified verdicts; that low performance on a human-authored taxonomy should be treated by default as evidence of misalignment, with genuine absence of competence a further hypothesis requiring channel-independent evidence such as representational probing; and that a concept being intuitive and easily tested for humans is evidence that it is a privileged unit of human cognition, not evidence that it is foundational for cognition as such.

This paper has not offered a first-order theory of what understanding is. It has argued, from a concrete empirical case, that the dominant practice for measuring understanding, in machines and, we suggest, implicitly in humans too, smuggles in a substantive and anthropocentric answer before the inquiry has properly begun: that understanding just is alignment with a taxonomy of concepts native to embodied human cognition, tested through tasks and channels designed around human perceptual access. Decentering the human case does not mean abandoning it; it means holding that vocabulary at arm’s length, treating it as one coordinate system among several rather than as the definition of the phenomenon under study. A metaphilosophical inquiry into understanding of understanding has to begin by asking not only what understanding is, but whose understanding the question was asking about all along.

References

Arseny Moskvichev, Victor Vikram Odouard, and Melanie Mitchell. The conceptarc benchmark: Evaluating understanding and generalization in the arc domain, 2023. URL https://arxiv.org/abs/2305. 07141.

François Chollet. On the measure of intelligence, 2019. URL https://arxiv.org/abs/1911. 01547.

Chaz Firestone. Performance vs. competence in human–machine comparisons. Proceedings of the National Academy of Sciences, 117(43):26562–26571, 2020. doi: 10.1073/pnas.1905334117.

Jennifer Hu and Michael C. Frank. Auxiliary task demands mask the capabilities of smaller language models. 4 2024. URL https://arxiv.org/abs/2404.02418v2.

Melanie Mitchell. The metaphors of artificial intelligence. Science, 386(6723):eadt6140, 11 2024. doi: 10.1126/science.adt6140.

Henry Shevlin and Marta Halina. Apply rich psychological terms in ai with care. Nature Machine Intelligence, 1(4):165–167, 2019. doi: 10.1038/s42256-019-0039-y.

John Hewitt, Robert Geirhos, and Been Kim. We can’t understand ai using our existing vocabulary. 2 2025. URL https://arxiv.org/abs/2502.07586v1.

Shane Legg and Marcus Hutter. Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4):391–444, 2007.

Halfdan Holm and Soumya Banerjee. Intelligence in animals, humans and machines: A heliocentric view of intelligence? AI & Society, 40(2):1169–1171, 2024. doi: 10.1007/s00146-024-01931-1. URL https: //doi.org/10.1007/s00146-024-01931-1.