Unearthing the Nature of Knowing

Who to trust • What to reject • How to generalize

In today’s world, the pull of new information is constant, ubiquitous, and limitless. To be knowledgeable citizens, people must sort through the mess and decide which sources are trustworthy, what information to absorb, and what to reject. They are also constantly evaluating how new situations and new information fit with their past experiences and knowledge. Psychological scientists are asking what underlies all this information processing: How do people know what they know? And how do they understand what others know? 

Researchers have found that people can gauge others’ knowledge with only a few pieces of information, and people learn to purposefully avoid inconvenient information starting in childhood. And although humans are good at generalizing their knowledge to new situations, they often make mistakes by oversimplifying.

But if human knowledge is both incredibly impressive and imperfect, artificial intelligence (AI) systems seem comparable: mind-blowing in their capabilities, but also far from flawless. And in some ways, nowhere near as capable as human intelligence.

Who to trust 

People learn quickly from those around them. But to do that, they need to know who to rely on for accurate information.

“We’re very good at making snap judgments using heuristic shortcuts to arrive exactly at the kind of social information we need,” said Olivier Morin, a researcher at Institut Jean Nicod, in an interview with the Observer. “But we don’t know exactly how it works.” 

Morin hypothesized that knowledge is nested: People who know only a little about a subject know common pieces of information, whereas people who know a lot about a subject know both common and rare pieces of information. This allows people to trust information from someone who knows rare tidbits as they should be well informed on the whole breadth of the subject.

To test whether people can infer others’ knowledge this way, Morin and his coauthors asked 848 U.S. adults to estimate others’ knowledge of a subject on the basis of whether the other person had answered one question about that subject correctly or incorrectly. The results, published in a 2025 Psychological Science paper, revealed that people were able to accurately predict others’ knowledge, suggesting knowledge is nested so people can gauge others’ knowledge according to whether they know rare pieces of information. To make this prediction, they also intuitively determined how widespread or rare knowledge is.

“Clearly they are making some sort of intuitive judgment,” said Morin. “It cannot be reasoning because it’s not something you can reason your way to the right answer, [it’s] not like a math problem. You have to take a shortcut.”

Morin suspected people may use superficial cues such as the rarity of certain words. Even with very little information—whether the other person answered one question correctly or not—people were able to infer others’ knowledge. During real conversation, people would have many more clues to base their inferences on.

But knowing rare information on one subject does not necessarily translate to knowledge about other subjects. Someone who knew a lot about astronomy did not necessarily know a lot about American history.

In the current landscape of constant information bombardment, the study offers comfort that people are naturally able to tell who to trust. Morin said the general public listens to experts, even when they may claim otherwise, and that while “fake news” looms large in the public consciousness, it actually makes up only a small fraction of what people come across. Furthermore, people can generally tell the truth from the lies.

“Many people do try to influence us with lies, but how many of these lies make it into your feed? And out of those, how many of them did you actually believe? Reassuringly, that’s way less,” he said.

What to reject 

Sometimes people reject information because it’s false, but other times people choose to avoid learning information that they’d rather not know (such as related to their health or financial status). Yet children start out extremely curious, so Radhika Santhanagopalan, postdoctoral researcher at the University of Chicago, wondered when information avoidance is learned.

“I was curious what accounted for this transition between us turning from these seemingly indiscriminate seekers of information as children to what seems to be these selective information avoiders as adults,” said Santhanagopalan in an interview. “That paradox is what motivated me.” 

In a 2025 Psychological Science paper, Santhanagopalan and her team examined information avoidance in 320 5- to 10-year-old children. Children were presented with various scenarios and asked if they wanted to receive information such as learning about a prize they did not receive or learning they were the kid not invited to a birthday party. 

“Five- and 6-year-olds were generally kind of information grabbers. They wanted to seek out information, but around 7 and 8 is when kids started to exhibit these avoidance patterns,” she said. 

Older kids may be more aware that certain information could make them feel bad, or younger children may not have the inhibitory control to overcome their natural curiosity. When told to make decisions that would help them feel happy, even younger children started to show these avoidance patterns, supporting the idea that anticipating emotional states is what’s driving avoidance. 

In one experiment, participants were asked whether they wanted to learn how their choice would affect a partner’s payoff before deciding between two payoffs, one of which would give them more stickers. Older children avoided learning how their actions would affect someone else, so they had moral “wiggle room” to choose the self-interested action. 

“This speaks to this tension that older kids and also adults have, which is … you want to do what’s in your best interest, but you also care a lot about appearing fair,” said Santhanagopalan. “This moral wiggle room allows us to exploit ambiguity in order to reach both goals where we can get the thing that we want while also having this illusion of fairness or veil of ignorance.” 

Although children generally exhibited avoidance with age, the exception was information about academic competence. Older kids wanted to know whether they were the kid who scored really poorly on a test, despite the negative feeling that might come with that knowledge.

“This was the reverse effect basically than what we found in the other studies,” said Santhanagopalan. She speculated that these children may have internalized the narratives discussed in elementary school about growth mindset and how intelligence is not fixed. 

This might be an area where adults can learn from younger people. Like children seeking to know about their grades, it could help if adults are more deliberate about their actions. 

“If there’s a medical checkup that we should be getting, we should be reflective about why it is that we keep postponing that,” said Santhanagopalan. 

“Oftentimes we are motivated to focus on our short-term comfort over our long-term benefit,” she said. “Sometimes we don’t even realize that we’re actively avoiding something.”

Modeling information-seeking behaviors around children, peers, and colleagues, especially in situations that give rise to discomfort, can help people resist this unconscious inclination to avoid information.

“If we’re in environments in which people are constantly being challenged and being asked to question things, that will help facilitate an environment in which there’s less willful ignorance,” said Santhanagopalan. “Short-term discomfort is something that we have to push through in order for long-term gains and that’s true for any broader societal issue, whether it’s climate change or politics or health.” 

How to generalize

When people are in new and challenging situations, they often generalize. These generalizations help people recognize a known object in a new scene and quickly adapt and make sense of novel situations. In a 2025 Current Directions in Psychological Science paper, Mirko Thalmann and Eric Schulz from the Institute for Human-Centered AI at the Helmholtz Center for Computational Health discussed three cognitive mechanisms that allow people to generalize: applying simple rules, judging new objects by their similarity to previously encountered objects, and applying abstract rules. 

But, while people have been portrayed as excellent at generalization, especially when compared to artificial intelligence systems, people’s preference for simple rules often leads to mistakes.

“We presented a more detailed view and showed that humans often systematically fail tests of generalization,” the authors wrote. “They rely too heavily on rules.”

To overcome the bias for solving a task with a simple rule, people may need thousands of training examples and feedback to eventually understand the complexity and generalize accordingly. By contrast, when a simple rule is required, humans perform well because they can use their lifetime of experience to quickly solve it.

“The true art of human generalization lies not in a universal ability to generalize, but in the well-formed craft of using adapted, efficient representations in the face of limited processing capacity,” the authors wrote.

In the paper, the authors compared human generalization to artificial intelligence systems. AI systems use the same generalization mechanisms as humans, but differ in how they extract and apply patterns, making them sometimes better and sometimes worse than humans. AI systems are often modality-specific; a system that excels at object recognition might struggle to process language. AI systems also don’t observe and interact with objects in different contexts like humans do, which limits their ability to form internal models of the world. They may perform better at a given task within a particular modality, but only the task they are trained upon.

“If you change the materials from materials that they’re more familiar with to materials they’re less familiar with, then their performance plummets precipitously,” said Robert Goldstone, an APS Fellow and professor at Indiana University who studies computational modeling of human cognition, in an interview with the Observer.

Goldstone’s fear is not the science fiction worry of computers taking over and controlling humans, but rather that humans will trust machines too much and lose their ability to think independently.

“The more immediate risk for us is overestimating what these systems are capable of doing and underestimating what the human brings to flexible generalizable cognition,” he said. 

AI systems—particularly systems based on the transformer algorithm, such as ChatGPT, Gemini, and Claude—are also limited because they tend to be word-based, whereas people are not only using language but also physically experiencing the real world and interacting with other people.

“One of the things that makes humans so smart is having to interact with other smart humans,” said Goldstone. In addition, “People are manipulating their world, they’re categorizing their world, they’re trying to make sense of their world, they’re trying to explain their world, and that’s not what these systems are trained to do.” 

As psychological scientists probe how human intelligence works—how people absorb, reject, choose, and apply information—artificial intelligence will continue to improve in tandem. But with that comes the risk of overreliance on AI systems. 

“Psychologists are going to be very important for figuring out what is the nature of intelligence and acting as thoughtful critics of the current generation of AI systems,” said Goldstone. “We can be impressed by what they’re doing, but we should not be satisfied, and we certainly shouldn’t be thinking that they can hold a candle to what people are able to do.”

Back to top

Feedback on this article? Email [email protected] or login to comment.

References

Dubourg, E., Dheilly, T., Mercier, H., & Morin, O. (2025). Using the nested structure of knowledge to infer what others know. Psychological Science, 36(6), 443–450. 

Santhanagopalan, R., Risen, J. L., & Kinzler, K. D. (2025). Becoming an ostrich: The development of information avoidance. Psychological Science, 36(7), 528–544.

Thalmann, M., & Schulz, E. (2025). How can we characterize human generalization and distinguish it from generalization in machines? Current Directions in Psychological Science, 34(5), 293-300.