AI and humans show similar effort patterns when solving reasoning tasks

What is it about?

In this study, we compared how long people took to answer a wide range of reasoning questions (arithmetic, logic, relational problems, intuitive judgments, and inductive reasoning problems) with how many "thinking steps" a reasoning AI model produced when solving the same problems. Across all seven tasks, the AI used more steps on items that humans found harder, and fewer steps on easier ones. The number of steps the model used closely resembled people's reaction times, even though the model was never trained on any of the human data. This pattern held not only within each task but also across tasks: the kinds of problems that humans generally take longer to solve, such as intuitive reasoning and challenging inductive reasoning items, were also the ones for which the model required more reasoning tokens. This suggests that these new reasoning-focused AI systems may capture something about the structure of human thinking: they can mimic not just what answers people give, but also how much effort problems require, offering a new tool for studying how humans reason.

Featured Image

Photo by Steve Johnson on Unsplash

Photo by Steve Johnson on Unsplash

Why is it important?

For decades, many researchers have argued that human reasoning must rely on symbolic, rule-based systems, and that neural networks—because they learn statistical patterns—could never capture the structure of human thought. This study challenges that view. Reasoning AI models mirror the effort humans expend on different kinds of reasoning, and the fact that a purely statistical system reproduces these core features of human thought suggests that some aspects of reasoning may emerge from statistical learning without symbolic mechanisms. This opens up new ways to study human cognition using models that were not designed to look human in the first place. Nobody built these systems to behave like people: the goal is simply to get them to solve problems reliably. The fact that a statistical model trained only on text and rewarded for correct answers ends up mirroring the effort patterns of human reasoning is genuinely quite striking.