Good conversation between two people involves “pink noise,” according to new research. This term refers to fluctuations in semantic similarities – alignments in speech patterns and word choice – across time scales during a chat. Kathryn O’Nell and colleagues’ study showed that tête-à-têtes are more enjoyable when there is more pink noise. Moreover, they found that a large language model added pink noise into generated scripts when asked to make them enjoyable. Many theories of speech view conversation as a linear means-to-an-end. This perspective does not consider pink noise, nor its purpose. O’Nell et al . elaborate on what pink noise looks like: “[Speakers] remain aligned for a while, diverge as one introduces something new, then realign as the other picks it up and develops it further.” The team searched for pink noise in two preexisting two-person conversation datasets, analyzing its presence alongside information about how much each pair liked chatting. Pink noise corresponded with enjoyment. It also appeared in conversations from speed-dating encounters and two-host podcasts. Then, the authors prompted a large language model (Claude Sonnet 3.7) to produce both enjoyable and unenjoyable conversational scripts, finding that Claude injected more pink noise into the enjoyable scripts. This outcome, the scientists suggest, may come from the LLM’s ingestion of human-authored text.
Science Advances
Pink noise in dyads’ semantic similarity trajectories predicts conversation enjoyment
6-Nov-2026