Summary
The episode works through four questions that students often find difficult when they begin to use AI tools for study and professional work. It examines why a tool that explains demanding ideas well can fail at simple counting, and it uses a short probability calculation to show why three clean runs of a known-answer test cannot establish that a tool is rarely wrong. It considers how much prompts can do about fabrication, including leading questions that carry an assumption into the output, and it works through two verification examples in which a figure matches its source while the wording, denominator or population has changed. It closes with the classification of small counts in aggregate tables and a discussion between the hosts about how much checking is appropriate during private study.