LLMs Do Not Reason Like Humans, But They Can Deliver Reasoning-Like Performance
Despite the marketing hype, large language models (LLMs) do not reason, at least not in the ways humans do. Underlying that assertion is the fact that we humans do not apply the term “reason” consistently. In a philosophical sense, reason is most often associated with activities such as inference with justification, belief revision, action guidance, causal modeling, and embodiment and social argumentation. While LLM research shows real gains on multistep tasks and tool-using workflows, recent research also documents brittleness, shallow causal depth, and weak robustness in open-ended settings.
FROM SYMBOLIC LOGIC TO STATISTICAL LANGUAGE
Before today’s arguments about whether LLMs “reason,” AI spent decades trying to make reasoning explicit. Expert systems encoded rules from specialists and applied them to narrow domains with impressive discipline. Case-based reasoning took a related path, solving new problems by comparing them to prior cases and adapting what had worked before. Both approaches treated intelligence as something that could be organized, represented, and executed through formal structures. That ambition gave AI some of its clearest early wins, especially in bounded environments, where the world could be described in advance and the relevant variables kept under control.
The limits arrived just as clearly. Expert systems depended on brittle rules, expensive knowledge engineering, and domains stable enough to tolerate exhaustive modeling.
Vasant Dhar, in his 2024 article in Communications of the ACM, “The Paradigm Shifts in Artificial Intelligence” (cacm.acm.org/research/the-paradigm-shifts-in-artificial-intelligence), explained this well. Case-based reasoning handled similarity and precedent better than rule systems alone, but it still relied on curated representations and human choices about what counted as the best analogy. Neither approach scaled easily to the ambiguity, incompleteness, and contextual drift of ordinary language and everyday work. Logical reasoning did not fail because logic was wrong. It failed because the world refused to stay still long enough to be captured as a clean set of symbolic relationships. The more open the environment, the more symbolic AI ran into combinatorial overload, maintenance fatigue, and the stubborn fact that much of human judgment depends on tacit knowledge that resists formalization.
The rise of LLMs was as much a reaction against the disappointments of symbolic AI’s search for explicit, general-purpose reasoning as it was a triumph of compute and data.
Where expert systems required engineers to specify the structure of thought in advance, LLMs learn patterns from vast corpora and infer structure from use. They do not achieve reasoning by implementing logic at scale. They bypass much of it by producing outputs that often look like reasoning without relying on hand-built symbolic representations.
In that sense, the failures of the symbolic era did not disappear. They were absorbed into a new architecture that traded formal clarity for probabilistic fluency and, in doing so, reopened the question that still defines the field—whether systems that perform reasoning-like acts have actually captured anything like reason itself.
WHAT PHILOSOPHY AND COGNITIVE SCIENCE MEAN BY REASONING
Claims that LLMs reason often result in people hearing, “AI solved a hard problem.” They infer deliberation, understanding, intention, and judgment from the assertion. Human reasoning, however, has always carried more weight than procedural success. In philosophy, reasoning is often defined as an inferential process that transforms one set of attitudes into another, but practical reason, “the general human capacity for resolving, through reflection, the question of what one is to do” (Wallace, R. J., Benjamin Kiesewetter, “Practical Reason,” Stanford Encyclopedia of Philosophy; plato.stanford.edu/entries/practical-reason), goes further. It concerns what an agent ought to do, how action is justified, and how reflection connects to intention. In other words, reasoning is not just output. It is bound up with commitment, normativity, and action.
Cognitive science broadens the picture even more. Contemporary work on human reasoning, no longer treats logic as the sole or even primary standard. Research now emphasizes probabilistic judgment, belief revision, dual-process dynamics, causal modeling, and the construction of mental models that simulate possible states of the world (Over, David E., Jonathan St. B. T. Evans, Human Reasoning, 2024; cambridge.org/core/elements/abs/human-reasoning/6227B8E27932FB6949DD8FB008E0E9A9; and Johnson-Laird, Philip N., “Mental Models and Human Reasoning,” Proceedings of the National Academy of Sciences, V. 107, No. 43 (2010): pp. 18243–18250; pnas.org/doi/10.1073/pnas.1012933107).
Other work argues that reasoning is deeply social, developed in part to generate and evaluate arguments in communication, while embodied cognition claims that concepts and inferences are shaped by the body’s interaction with the world (Mercier, Hugo and Dan Sperber, “Why Do Humans Reason? Arguments for an Argumentative Theory,” Behavioral and Brain Sciences, V. 34, No. 2 (2011): pp. 57–74, pubmed.ncbi.nlm.nih.gov/21447233; Shapiro, Lawrence, Shannon Spaulding, “Embodied Cognition,” Stanford Encyclopedia of Philosophy, plato.stanford.edu/entries/embodied cognition). Those perspectives do not agree on every mechanism, but they do agree on one point: Human reasoning is richer than the execution of a multistep procedure.