Network cognition

GAICIL Insights

Everyone carries a mental map of their social world, and the map is drawn in a peculiar, distorted hand.¹ Network cognition, by Roberta Dessí, Edoardo Gallo and Sanjeev Goyal, measures the distortions. Combining experiments with survey data, the authors study individuals' ability to memorize and recall information about friendship networks.

In the experiment, subjects see a network in which their location is exogenously assigned, then answer questions after it disappears. Three biases emerge. Subjects underestimate the mean degree compared with the actual network. They overestimate the number of rare degrees. They underestimate the number of frequent degrees. Analyzing survey data from two real friendship networks, one from a Silicon Valley firm and one from a university research center, the authors find, somewhat remarkably, the same biases in the field.

Three more findings complete the picture. Network cognition depends on the subject's location. Its accuracy varies with the nature of the network. And it significantly affects economic decisions. Misperceived networks lead to different choices.

In plain terms: people think their world is sparser than it is, and they think the unusual is more common than it is. Those are exactly the illusions a feed can either correct or feed.

The parallel with today's technology is close. Artificial intelligence now draws much of the map for people. Platforms present curated views of networks: who is connected to whom, who matters, what others are doing. If perception is systematically biased, the way AI presents network information can correct or amplify the bias. A recommendation system that spotlights rare connections might reinforce the tendency to overestimate rare degrees. A design choice made for engagement can quietly reshape a population's picture of its own society.

Because network cognition shapes economic decisions, the stakes are concrete. Choices about whom to trust, whom to trade with and how to allocate effort rest on perceived networks. Biased perceptions mean suboptimal choices. AI tools that summarize network structure accurately could improve those decisions by showing individuals their real position and surroundings.

Organizations should pay attention. Even inside a single Silicon Valley firm, employees' mental maps of informal connections are biased. Managers who rely on employees' perceptions to identify influential colleagues or communication channels may be misled. Data-driven tools, including AI analysis of communication patterns, could complement perception with measurement, provided they respect privacy and the informal nature of social ties. The org chart people carry in their heads may be the least accurate document in the company.

The social media debate gets a jolt too. Arguments about online influence often assume users accurately perceive their networks and the prevalence of views within them. The biases here suggest otherwise. Underestimating the mean degree and misjudging the frequency of rare and common types can distort perceived social norms. Interventions that supply accurate information about network structure and behavior may correct misperceptions; interventions that rely on unaided perception may falter.

The method is a template. Pairing controlled experiments with survey data from real networks shows that lab biases appear in the field. The same pairing can test how AI-mediated presentations of networks affect perception: experiments compare designs, and field data check whether effects persist. Before a platform redesigns how it shows people their friends, it can run this kind of test.

Read as a whole, human cognition is part of network infrastructure. Networks operate through decisions by people who perceive them imperfectly, so technologies that shape perception shape behavior and outcomes. AI could improve network cognition by delivering accurate, accessible information about social structure. Realizing that potential requires designs grounded in evidence about how people actually perceive networks, and this paper supplies that evidence.

A final provocation: perhaps the most useful social feature an AI system could offer is a corrective lens, a gentle reminder that people know more people than they think and that rare things are rarer than they seem.