Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations | Stanford SALT Lab - BRIAN HEGER

You can also access this resource and many others in Issue 363 of my Talent Edge Weekly newsletter.

Most of what we know about how employees actually use AI at work comes from surveys and self-reports. However, a new Stanford study offers a rigorous behavioral look instead, and it is the first of three studies from a new Anthropic pilot giving outside research teams (Stanford’s SALT Lab, Oxford’s Human Information Processing Lab, and METR) privacy preserving access to real Claude usage data. Each team received access to roughly 250,000 real Claude conversations. Here are just two of the many findings from Stanford’s report (the first to be released) along with some of the practical implications. 1) 56% of the tasks people brought to AI were classified as consequential or high stakes, meaning work that is hard to reverse and carries real professional or financial weight, concentrated in advisory and professional domains. Implication: this challenges the assumption that employees mostly hand AI low stakes work, and suggests governance should focus on the majority of work that actually carries higher stakes. 2) friction, meaning moments where the human-AI collaboration broke down or required correction, showed up in nearly half of all conversations, but researchers classified much of it as productive, since working through it tended to improve the output or deepen understanding of the issue. Implication: organizations can frame this kind of friction as a normal, even necessary, part of working with AI rather than a sign the tool has failed.