New research introduces a novel framework for quantifying common sense, revealing significant variations in common beliefs and the influence of social perceptiveness. Their findings highlight the uniqueness of common-sense perceptions, suggesting future global studies and applications in AI.
Researchers from Penn tackle a significant deficiency in the understanding of knowledge.
Throughout the annals of human existence, the endurance and development of intricate communities have been largely reliant on the accumulation and application of knowledge. Of equal importance is the understanding of shared beliefs regarding what is considered true or untrue, often referred to as common sense. This concept plays a significant role in daily scenarios, such as following traffic regulations: Pedestrians instinctively steer clear of venturing into oncoming traffic, whereas motorists avoid using sidewalks as shortcuts to evade traffic jams.
However, deviations from these seemingly intuitive principles of interpersonal conduct remain prevalent. Despite the ubiquity of common sense, there is no unanimous consensus on what individuals collectively perceive as true or false.
The Study on Common Sense by Watts and Whiting
Now, Penn Integrates Knowledge University Professor Duncan Watts and Mark Whiting of the School of Engineering and Applied Science and Wharton School have developed a unique framework to quantify the concept of common sense. In a paper published in the Proceedings of the National Academy of Sciences, the researchers present a way to quantify common sense at both the individual and collective levels.
“Common sense is something that we all believe we possess, but rarely, if ever, are we forced to articulate which of our beliefs we consider ‘commonsensical’ or who else we think shares them,” Watts says. “What Mark and I set out to do was create a framework for answering these questions in a systematic, empirical way.”
Individual and Collective Common Sense
The researchers first tackled the challenge of defining and quantifying individual perceptions of common sense, which they termed “commonsensicality.” This involved assessing how much agreement exists among people regarding specific claims and how aware individuals are of others’ agreements on these claims.
“Essentially, we sought to measure not just whether people agree on a claim but also their awareness of said shared agreement,” Whiting says. “It’s an approach that moves beyond simply tallying up agreements to understanding the depth and breadth of consensus.”
The second aspect was collective common sense, a concept focusing on shared beliefs across different groups. This measure helped the researchers gauge the extent of common beliefs within groups, and, interestingly, they found that the larger the group the fewer common beliefs are held.
The researchers introduced this measure as the “pq common sense” metric, which has its basis on the idea of mapping out a network of beliefs shared among people—each person and each claim they believe in is connected—with the goal to find clusters or groups within this network where there’s a high level of agreement on certain claims.
“Here, ‘p’ represents a fraction of the population and ‘q’ a fraction of claims,” Whiting says. “The framework then calculates the proportion of claims q that are shared by a certain proportion of people p.”
This is like examining a large group of people and figuring out what percentage of these people agree on a certain percentage of claims, Whiting says. It quantifies the commonality of common sense across a population.
Framework Testing and Insights
To test this framework, the researchers then collected a vast array of 4,407 claims—ranging from philosophical statements to practical truths—and had 2,046 people rate these claims in terms of how commonsensical they found them. Examples of categories of claims corresponded to the top level of Wikipedia’s ontology and included general references: geography and places, mathematics and logic, culture and arts, and philosophy and thinking. They also classified claims based on where they stand on spectrums like fact versus opinion, literal language versus figure of speech, or knowledge versus reasoning.
They then applied their framework to this data, analyzing the network of agreements to find patterns of common belief, and their results showed a significant variation in what individuals consider common sense, with few beliefs universally recognized at the group level.
“Interestingly, demographic factors like age, education, or political leaning did not significantly influence a person’s level of common sense,” Whiting says. “But, social perceptiveness—the ability to understand others’ thoughts—did correlate with higher commonsensicality.”
Their study also highlights the individual uniqueness of common-sense beliefs, showing that agreement on common sense diminishes significantly in larger groups. “Our findings suggest that each person’s idea of common sense may be uniquely their own, making the concept less common than one might expect,” Whiting says.
The researchers note that, given their interest in common sense as a societal concept, expanding their research to a global scale would be a logical next step. This would involve studying common sense across different cultures and societies to understand how it varies and what universal aspects might exist. They are also interested in developing methods to measure and implement common sense in AI systems that could improve AI’s understanding of human contexts and enhance its decision-making capabilities.
“When we think something is common sense, we often feel very strongly about it, but, as we see in this study, we very often disagree with each other about what it says,” Watts says. “So, whether our goal is to better resolve disagreements about matters of common sense or to teach common sense to computers, we had better first have a clearer picture of what it is and isn’t. That’s what we want to accomplish.”
Reference: “A framework for quantifying individual and collective common sense” by Mark E. Whiting and Duncan J. Watts, 16 January 2024, Proceedings of the National Academy of Sciences.
DOI: 10.1073/pnas.2309535121
