The question the team posed is deceptively simple: when people teach, are they reasoning about another mind, or merely executing a routine? Psychologists have long regarded teaching as a foundational social behavior, one that underpins education, culture and the transmission of knowledge across generations. But the cognitive machinery behind it has remained opaque, because teaching can be accomplished in two very different ways. It can be performed optimally, through mentally effortful reasoning that treats the learner as a mind to be modeled. Or it can be performed frugally, through heuristics that demand little thought and no mentalizing at all. Debates about when humans deploy expensive planning rather than inexpensive habits usually unfold over nonsocial tasks: choosing between rewards, navigating mazes, pressing keys. Teaching offers an unusually clean test case, because the two routes leave measurably different fingerprints in the examples a teacher selects. The researchers set out to establish which route people naturally take, whether they abandon a failing strategy when circumstances change, and what it takes to make them switch.

Experiment 1, with 100 participants, delivered the study’s first surprise: teaching strategies are not a matter of degree but of kind. Participants took the role of teachers, selecting examples to convey a concept to a learner whose knowledge they could not directly observe. The researchers then fitted a family of computational models to each individual’s choices, asking which algorithm best reproduced the observed behavior. For some participants, the best-fitting account was an optimal Bayesian pedagogy model, in which the teacher explicitly reasons about the state of the learner’s knowledge before choosing what to show. For others, the best-fitting account was a set of simple heuristics that require no mentalizing whatsoever — rules of thumb that select plausible-looking examples without ever computing what the learner believes. Crucially, both groups faced the same task, the same instructions and the same information. The difference lay not in what they knew or how well they taught, but in how their minds chose to spend their cognitive budget. Teaching, it turns out, has personality.

To appreciate why that split matters, it helps to unpack what Bayesian pedagogy demands. In this framework, teaching is a recursive act of mutual inference. The teacher maintains an internal model of the learner’s current beliefs — a probability distribution over the rules the learner considers plausible — and then runs a counterfactual simulation for every candidate example: if I show this, and the learner understands that I am deliberately trying to teach, how likely are they to update their beliefs toward the correct rule? The optimal teacher selects the example with the highest expected payoff, the greatest probability of steering the learner’s inference toward the truth. This is heavy cognitive lifting. It requires holding a representation of another person’s mental state in working memory, predicting how that state will change with each new piece of evidence. Cognitive scientists call reasoning about other minds mentalizing, and it ranks among the most demanding computations the social brain performs. The Bayesian teacher is the cognitive equivalent of a chess player who calculates several moves deep while simultaneously modeling the opponent’s style.

The heuristics, by contrast, are mentally frugal. A heuristic-driven teacher might simply pick examples that are themselves excellent specimens of the concept, on the intuitive logic that good examples make good teaching, without checking whether those examples tell this particular learner anything new. Another might repeat instances that worked earlier, or choose items resembling previously successful ones, letting past performance rather than the learner’s current understanding drive the next choice. Such rules can perform respectably in many situations, which is why they persist. But they are fundamentally blind: they never represent the learner’s knowledge, so they cannot detect that the learner has already grasped a point, nor can they recognize a misconception the teacher is unwittingly reinforcing. It is the difference between a physician who orders tests based on a patient’s specific symptoms and one who orders the same standard battery for every patient who walks through the door. The first is expensive but tailored; the second is cheap but indifferent to the very person it serves.

Then came the study’s sharpest test. In a preregistered Experiment 2 with 253 participants, the researchers altered the teaching environment so that the heuristic no longer worked — conditions in which blindly applying the shortcut would steer the learner astray. If people were flexible strategists who adjusted their cognitive spending to circumstances, they should have abandoned the failing heuristic and switched to mentalizing. They did not. Participants persisted in using the now-ineffective shortcut, a statistically robust effect (P < 0.001; rank-biserial correlation r = 0.287, 95% confidence interval 0.149 to 0.419). The rank-biserial statistic, an effect-size measure for two-group comparisons, points to a small-to-medium but highly reliable tendency. In plain terms, even when the cheap strategy stopped paying off, people kept deploying it, apparently because it remained the path of least resistance. The result echoes a familiar theme from research on habits: behaviors that economize on effort become sticky, and a track record of past success is enough to keep a strategy alive long after its expiration date.

The third experiment, preregistered and by far the largest, with 759 participants, showed that this stickiness can be broken — not by urging people to try harder, but by scaffolding the expensive step itself. Participants received an auxiliary task that supported their inference about what the learner knew, effectively lowering the cognitive cost of mentalizing. With that inference partially externalized, the tendency to persist with heuristics was pre-empted: participants shifted toward reasoning about the learner’s knowledge when choosing their teaching examples (P < 0.001; partial ηp² = 0.107, 95% confidence interval 0.068 to 0.148, a medium-sized effect in this design). The barrier to thoughtful teaching was not a lack of ability but a question of cost. When the price of representing the learner’s mind dropped, people paid it, and their teaching changed accordingly. Mentalizing, the study suggests, is not a fixed capacity that some possess and others lack; it is a resource that people purchase when its price falls or its expected payoff rises.

Taken together, the three experiments reveal what the authors describe as “sophisticated arbitration between planning and heuristics during teaching.” The mind appears to run something like a cost–benefit calculation over its own cognitive effort: mentalizing buys accuracy in transmitting knowledge, but it is expensive, so the cognitive system rations it. Heuristics are the economy class of teaching — cramped, limited, but affordable. Individual differences in Experiment 1 show that people price the trade-off differently, with some defaulting to first class and others to economy. Experiment 2 shows that once a cheap strategy is running, it resists shutdown even in the face of clear evidence of failure. Experiment 3 shows that the pricing is not fixed but responsive to context: change the cost structure, and the strategy follows. The findings extend dual-process accounts of cognition — the interplay of fast, automatic and slow, deliberate thinking — into the social domain, where the slow, deliberate option is specifically the construction of a model of another mind.

The implications radiate well beyond the laboratory. In education, a teacher relying on heuristics may deliver polished, reasonable-looking lessons while never noticing that a student’s misconception is quietly being reinforced. This research suggests that tools which surface a learner’s actual state of knowledge — diagnostic feedback, formative assessment, structured insight into what students do and do not understand — could shift even shortcut-prone teachers toward genuinely adaptive instruction. In human–AI interaction, the same asymmetry grows by the year: people now teach machines constantly, from recommendation algorithms to household robots to large language models, and whether they do so with or without mentalizing may determine how quickly and how well those systems learn from the examples people supply. For theories of bounded rationality, the work adds a social dimension to a long-standing principle: intelligence is not about always thinking harder, but about deploying hard thinking precisely where it changes outcomes.

The authors frame their contribution as demonstrating just such arbitration and elucidating “the more general mechanisms involved in adapting mental effort during social interactions” — a framing that positions teaching not as a special talent reserved for gifted educators but as a window onto how any mind manages its cognitive budget in the presence of another person. The natural next questions follow directly from the design. When in development do individuals settle into their pricing schemes for mental effort, and how durable are those schemes across the lifespan? Does the same arbitration govern other social behaviors — cooperation, conversation, deception — where modeling another mind is likewise optional but consequential? And can such scaffolding be scaled from a laboratory task to classrooms, workplaces and the algorithms humans increasingly find themselves teaching? What is already clear is that the gulf between a heuristic teacher and a mentalizing teacher is not a gulf of talent. It is a gulf of effort — and effort, this research shows, can be moved.