Problem
People don't always behave the way they say they do and for a long time, we couldn't study that gap at scale. Natural language processing (NLP) has changed what's possible. We can now observe how people talk, move and interact in ways that weren't possible before. These tools let us ask sharper questions about how people communicate, disagree, and live.
Science
We approached this question from three directions. The first study ran a large field experiment analysing more than 27,000 comments to test whether a small word change could improve the quality of input people give each other at work. The second ran randomised control experiments involving more than 1,800 conversations to test whether disagreement works differently when spoken rather than written. NLP was used to measure how open people are to each other's views. The third combined smartphone sensor data, covering GPS, microphone and activity recognition, with self-reports from more than 1,200 students to test whether people with different political views live differently day to day.
Outcome
The findings from this stream of research are clear. Asking for 'advice' rather than 'feedback' produces more concrete, useful input. The word 'advice' points people toward the future rather than the past, and that shift is measurable at scale through text analysis. Spoken disagreement produced greater mutual understanding, lower conflict, warmer impressions and more attitude alignment than disagreement in written exchanges. People use less open language in writing, precisely where it would help most. Liberals and conservatives do behave differently across socialising, movement, work and leisure, but the differences were small, most visible in leisure, and considerably smaller than students themselves assumed.
Change
Taken together, these findings show what's possible when we treat conversations, text and daily routines as data. Small features of language and medium shape how people learn, cooperate and understand each other. That matters for organisations designing feedback processes, for platforms choosing between voice and text and for communities trying to understand their own fault lines.
Read more: