Every workplace, every culture, every friend group operates on a dense web of unspoken rules that nobody explicitly teaches, yet somehow everyone knows. A new study offers a surprisingly simple explanation for how people may pull off this trick, borrowing its answer from the way toddlers learn grammar.
Rather than carefully calculating the best option or mindlessly copying whoever they last interacted with, people appear to gather social information in a loose, almost random fashion until a mental threshold seems to click into place. Once enough evidence piles up in favor of one behavior, a kind of internal switch seems to flip and the person becomes much more likely to stick with that choice. Before that threshold, behavior looks messy and exploratory. After it, behavior becomes stable.
The threshold in question, called the tolerance principle, was originally developed to explain how young children absorb rules like the English past tense, where exceptions like "ran" and "swam" coexist with the regular "-ed" pattern. Researchers published in the Proceedings of the National Academy of Sciences found that the same basic equation, tested across a range of memory assumptions and checked against alternatives, outperformed the leading theories of social learning when matched against real human behavior. It beat both the "people are copycats" model and the "people are rational decision-makers" model, raising pointed questions about how scientists have been thinking about social learning for decades.
How the Social Learning Experiments Measured Human Behavior
Research led by Douglas Guilbeault, Spencer Caplan, and Charles Yang analyzed data from a well-known series of experiments called the "name game." Participants in social networks were randomly paired with one another, round after round, and asked to independently type a name for an unfamiliar human face. If both people typed the same name, they earned a financial reward. If they typed different names, they were penalized. Any name could work as long as both people agreed on one.
At the start, nobody knows what anyone else will say. Many names float around and coordination is a mess. But over roughly 20 to 30 rounds, a single name begins to dominate and spreads through the entire network. Original experiments covering networks of 24, 48, and 96 participants totaling 264 people documented this pattern but didn't identify the individual-level mental process behind it.
Guilbeault's team reconstructed each participant's interaction history round by round and tested four models against what participants actually did next. One assumed people copy the most recent successful name. A second assumed people always pick the most common name in memory. A third assumed people pick names proportionally to frequency. The fourth was the tolerance principle model, which sets a specific cutoff: once one option fills at least 8 of 12 recent memory slots, the person stabilizes on that choice. Below that line, behavior stays exploratory.
Why the Rational Decision-Maker Model Falls Flat for Social Learning
After participants' memories were full, the tolerance principle model correctly predicted their next choice 87.9% of the time. The imitation model managed 81.1%. The optimization model, the approach that dominates current academic thinking about social learning, hit only 75.2%.
One of the study's most pointed findings is how poorly the optimization model performed early in the game, predicting participants' behavior only about 31% of the time before any convention had taken hold. Under that model, a rational agent always picks the statistically most likely winner. Instead, participants appeared to be sampling loosely from their experiences.
Even when a leading name occupied 48% of a participant's memory, participants chose it only about 61% of the time. Under the optimization model tested here, a perfectly optimizing decision-maker would choose it 100% of the time. People, it turns out, are not optimizers, at least not until they've gathered enough evidence to cross a threshold.
A Committed Minority Can Flip Social Norms, But Only Past a Tipping Point
The researchers also tested their model against experiments on how established conventions get overturned. After networks of players had settled on one name, experimenters secretly introduced planted participants who all pushed the same alternative. Previous research found that when this committed minority reached roughly 25% of the network, the old convention flipped. Below that threshold, the convention held firm.
The tolerance principle model reproduced these tipping point patterns more accurately than any other model, landing within the range observed in the human experiments across different network sizes. The imitation model predicted conventions would crumble even when dissenters made up just 10% of the population, which didn't match reality.
Beyond re-analyzing existing data, the team ran a preregistered experiment called "the mind-reading game," designed to test the competing models under precisely controlled conditions. Participants had to figure out behavioral patterns while navigating controlled levels of interference. The tolerance principle again outperformed all rivals, adding stronger evidence that the threshold may reflect something real in how people process social information, not just a quirk of the name game.
A three-year-old learning to say "walked" instead of "goed" and an adult decoding the dress code at a new job may be running the same basic mental software. If that holds up, it's a genuinely humble view of how the human mind handles one of its most taken-for-granted skills.