A controlled study published on 8 September 2026 tested six ways of showing people an AI’s reasoning and found the format participants preferred — a numbered plan-and-solve trace — produced nearly triple the false alarms on correct answers that a plain chain-of-thought trace did. People trusted the plan more, and checked it less.
The hosts work through the study’s numbers, then follow the mechanism into Claude Code’s plan mode, which ships that exact preferred format as the checkpoint before any file gets touched. They also hold onto the study’s own limits and the opposing case, from Nielsen Norman Group and a CMU-led position paper, for richer explanation rather than less. The full article, with sources, is at pipelinemag.ai.