Practices: use AI without forgetting how to think
This is a live Practicum page for Chapter 8. Verified: 2026-08-18. Practices change as models change, so check the date. These are examples for different roles, not permanent rules.
The Chapter 8 principle is simple: give the rough work to the machine, but keep the part that helps you grow. The table shows what people in different roles can hand to AI, what useful friction they keep, and what they repeat without AI.
| Role | Gives to AI | Keeps as useful friction | Repeats without AI |
|---|---|---|---|
| Analyst | SQL drafts, charts, summaries | writes the hypotheses and checks edge cases personally | once a week, completes one analysis from start to finish without a model |
| Lawyer | summaries and version comparisons | keeps risk assessment and strategy | writes the central argument, then checks it with an AI critic |
| Software developer | boilerplate, tests, code explanations | keeps architecture and security review | diagnoses a difficult bug before asking AI |
| Marketer | headline options and resizing | keeps the brand voice and final selection | develops one campaign a month from a blank page |
| Manager | summaries and email drafts | keeps decisions and conversations with people | writes the important decision before asking AI |
One rule for every role
Ask the model for options and weak points, not a finished answer. The training, not answers page has useful prompts. Once a week, do one important task without AI. Think of it as practice that keeps the skill available.
What the 2026 research actually found
The chapter's argument now has direct evidence behind it, and the useful part is not the scary half.
The MIT study that gave "cognitive debt" its name ran 54 undergraduates from MIT, Harvard, and Tufts through essay writing. Heavy, unguided reliance on a model produced measurable costs: weaker memory of what they had just written, reduced neural engagement, and output that converged toward sameness. A separate mixed-methods study of 666 people across age groups found heavy AI use associated with lower critical thinking, with cognitive offloading as the mechanism in between.
The finding worth building a habit on comes from a later session in the same MIT study. When participants who had been writing without any model were then given one, they showed elevated neural activity and used more sophisticated prompting than the group that had leaned on the tool from the start. The tool did not decide the outcome. The order did.
That is the whole practice on this page in one sentence: the person who treats the model as a tutor and argues with it stays sharp; the person who treats it as an oracle does not. It is also why the right column of the table matters more than the left one.
Caveats worth keeping. These are small samples of students on writing tasks, over short periods. They do not measure a career. Treat them as a reason to keep the friction, not as proof of brain damage.
Update rule: review once a quarter. Mark outdated material instead of deleting it. Related pages: mode audit · checklist. Version: 2026-08-18.