Editor's Synopsis
A new Stanford HAI-supported study published in Nature Human Behavior challenges the assumption that people use AI chatbots as shortcuts or for trivial tasks. Analyzing a quarter-million anonymized conversations between users and Anthropic's Claude, researchers led by assistant professor Diyi Yang found that the vast majority of interactions involve meaningful, consequential work — over 56% were deemed "consequential" or higher in stakes. Rather than passively offloading tasks, users actively engage with AI as a dynamic collaborator, with conversation length nearly doubling as stakes increase.
A key finding is the role of "friction" in human-AI collaboration — a creative tension present in roughly half of conversations that drives clarification, refinement, and learning. In two-thirds of interactions, users prompt AI to explain and teach rather than simply generate content. While friction can turn negative and cause collaborations to break down, nearly 80% of users experiencing negative friction attempted to recover the conversation. AI's teaching role proved most beneficial in high-consequence domains like law, software engineering, health, and home improvement.
The researchers raise concerns about a potential widening inequality gap, where AI's benefits concentrate among higher-stakes knowledge workers while others risk intellectual capitulation and de-skilling. They recommend increasing AI literacy so more people can use AI responsibly, recognize when output requires scrutiny, and leverage friction productively. The team is now developing tools to help users identify when AI output needs greater scrutiny and designing AI systems that support longer-term cognitive engagement, while calling for more academia-industry partnerships to advance transparent research at scale.
