Teaching another person is often treated as a sophisticated act of perspective-taking. A good teacher should work out what the learner already knows, predict what information will be most useful and then choose an example that changes the learner’s understanding. New research suggests that people do not always take that mentally demanding route.
Across three behavioural experiments, researchers found evidence for two distinct teaching strategies. Some adults behaved in ways consistent with reasoning about a learner’s knowledge, while others relied on simpler rules of thumb. Those heuristic strategies could persist even after the task changed and they stopped working well. But when participants were explicitly helped to infer what the learner knew, their teaching improved.
The study, published in Nature Human Behaviour, combines experiments with computational modelling to examine a basic question with implications far beyond classrooms: when people need to help someone else learn, how do they decide whether careful mental reasoning is worth the effort?
Teaching can require a model of someone else’s mind
Imagine trying to teach a colleague how a complicated system works. One approach is to build a mental model of what the colleague already understands, identify the gap and choose an explanation specifically designed to close it. That is a form of mentalizing, because the teacher is reasoning about another person’s internal state.
There is also a cheaper strategy. Instead of modelling the learner, a teacher can follow a heuristic such as choosing the example that seems most informative in general. That shortcut may work surprisingly well when the environment is predictable, but it can fail when the learner’s prior knowledge changes what counts as a useful lesson.
Sevan Harootonian, Thomas Griffiths, Yael Niv and colleagues used a graph-teaching task to distinguish these possibilities. Participants had to select information for a learner rather than simply solve a problem for themselves. Computational models then allowed the researchers to compare behaviour with the predictions of a more demanding Bayesian pedagogy strategy and simpler heuristic approaches.
The first experiment revealed striking individual differences
Experiment 1 involved 100 adults recruited online. The researchers found a pronounced split in how participants approached the teaching task. Some selected examples in a way consistent with an optimal Bayesian pedagogy model, which reasons about how evidence should update a learner’s beliefs. Others behaved more like heuristic teachers, relying on simpler selection rules that did not require the same degree of mentalizing.
This matters because the two strategies can produce similar choices in some situations. If a simple rule happens to recommend the same example as a learner-sensitive model, observing the final teaching choice does not reveal how the teacher arrived there. The study therefore used task conditions and modelling designed to make the strategies distinguishable.
The result challenges the assumption that successful teaching necessarily reflects sophisticated perspective-taking. A person can sometimes produce an apparently sensible lesson using a computationally cheaper shortcut. The important question is what happens when the environment changes and that shortcut no longer points to the best example.
People kept using shortcuts after the shortcuts stopped working
The second experiment tested precisely that problem. In a preregistered study of 253 participants, the researchers changed the task so that previously useful heuristic strategies were no longer effective.
Participants nevertheless showed a significant tendency to persist with those heuristics. The reported effect was P < 0.001, with a rank-biserial correlation of r = 0.287 and a 95% confidence interval from 0.149 to 0.419.
The finding is important because it separates a useful shortcut from flexible expertise. Heuristics are not inherently irrational. They reduce cognitive effort and can perform well when the structure of a problem is stable. But a shortcut becomes costly when a person continues using it after the conditions that made it useful have changed.
For teaching, that creates a familiar practical risk. An explanation that works for one learner may be repeated for another even when the second learner has different background knowledge. A teacher can therefore look consistent and experienced while failing to adapt to the person in front of them.
Scaffolding the inference changed how people taught
The third preregistered experiment was much larger, involving 759 participants. Instead of merely changing the environment, the researchers introduced an auxiliary task that scaffolded participants’ inference about what the learner knew.
That intervention reduced the tendency to rely on ineffective heuristics and improved teaching performance. The scaffolding effect was statistically strong, with P < 0.001 and partial eta squared of 0.107, with a 95% confidence interval from 0.068 to 0.148.
The result points to a more nuanced account than simply dividing people into good and bad teachers. Strategy use appears responsive to the cognitive demands of the task. When inferring the learner’s knowledge is difficult, a heuristic can save effort. When that inference is made easier, people can shift toward a more learner-sensitive strategy.
The broader issue is how people allocate mental effort
The researchers interpret the findings as evidence for arbitration between planning and heuristics. Human cognition has limited time and processing capacity, so people constantly face an implicit resource-allocation problem: when is deeper reasoning worth the cost?
Teaching makes that trade-off unusually visible because the relevant problem exists in someone else’s head. A teacher can either invest effort in modelling that learner or use a rule that has worked before. The first strategy is flexible but cognitively demanding. The second is efficient but risks becoming brittle.
This idea fits a wider pattern in cognitive science in which people do not always use the most sophisticated strategy available to them. Instead, they adapt mental effort to the expected value of thinking harder. The new work adds a social dimension by showing that this trade-off shapes how people transmit information to others.
Why this matters for education and workplace learning
The experiments were not conducted in schools, universities or workplaces, so the results should not be translated directly into a teaching policy. Even so, they raise a useful question for real learning environments: do teachers and mentors have enough support to diagnose what a learner already understands?
Many educational systems focus heavily on the quality of instructional material. The study suggests that another part of the problem may be the inference that comes before instruction. The same explanation can be excellent for one learner and redundant or confusing for another. Tools that make prior knowledge more visible could therefore change not only what teachers know, but which cognitive strategy they use when deciding what to teach next.
The same logic applies to managers, coaches and experienced employees training colleagues. Repeating a familiar demonstration may be efficient, particularly under time pressure. But if the learner’s knowledge differs from the assumptions built into that demonstration, the shortcut can become a mismatch. Asking a diagnostic question or giving a short preliminary task may make the learner’s state easier to infer and support more targeted instruction.
The study does not show that heuristics are bad
A central strength of the research is that it does not frame simple strategies as mere mistakes. Heuristics can be sensible because mentalizing has a cost. If a shortcut works reliably, investing additional effort in modelling another person’s beliefs may add little value.
The problem appears when people fail to switch strategies after the environment changes. Experiment 2 demonstrated that persistence directly. Experiment 3 then showed that making learner inference easier could alter the balance.
This makes the findings relevant to the design of educational technology as well. Adaptive systems often try to estimate a learner’s current state before choosing the next problem or explanation. The human results suggest that useful technology might also support the teacher’s inference process rather than simply automating the final recommendation.
Important limitations
The experiments used adults recruited through Prolific and a structured graph-teaching task. That gives researchers strong control over the information available to teachers and learners, but it is much simpler than real instruction. Classroom teaching involves motivation, language, social relationships, subject expertise, feedback and repeated interaction over time.
The computational models are also explanations of behavioural patterns rather than direct measurements of thought. Behaviour that fits a Bayesian pedagogy model is consistent with learner-sensitive reasoning, but it does not prove that participants consciously performed the exact computations formalised by the model. Likewise, heuristic classifications simplify what may be a richer mixture of strategies.
All three experiments involved adults, so the results do not establish how children teach or how teaching strategies develop with age and expertise. Professional teachers may also have domain-specific routines that differ from the strategies used by participants in a novel laboratory task.
Finally, improved performance after scaffolding does not mean that every form of additional learner information will help. Poor diagnostic information could create false confidence or add unnecessary complexity. Future research will need to test which kinds of scaffolding improve real teaching decisions and when the extra cognitive effort is worth the cost.
A different way to think about teaching skill
The study shifts attention from whether people are capable of reasoning about learners to whether they choose to do so in a particular situation. Across the three experiments, people showed distinct strategies, persisted with shortcuts when those shortcuts became ineffective, and changed their behaviour when inference about the learner was made easier.
That suggests teaching skill may depend partly on strategic flexibility. Knowing a subject is important, and understanding how people learn is important. But effective teaching may also require recognising when a familiar rule is no longer enough and when another person’s knowledge deserves a closer look.
Source Information
Study: Harootonian, S. K., Griffiths, T. L., Niv, Y. et al. “Mentalizing and heuristics as distinct cognitive strategies in human teaching.”
Journal: Nature Human Behaviour.
Published: 26 August 2026.
DOI: 10.1038/s41562-026-02540-2
Study design: Three behavioural experiments with computational modelling, including an initial experiment of 100 adults and two preregistered follow-up experiments involving 253 and 759 adults.
Participants: Adults recruited through Prolific.
