How you use AI decides what it does to you.
The same tools that erode thinking in one usage mode strengthen it in another. This series maps the difference, habit by habit, from the research.
Grounded in randomized trials, observational studies, and preprints, 2011–2026. Each guide names where the evidence is strong and where it is early.
Two ways to use the same tool.
What determines whether AI helps or harms your thinking is not how much you use it. It is whether you are asking it to hand you the finished thing, or to help you build it.
Extraction
Asking AI for finished answers.
“Write this for me.” “What’s the answer?” “Fix it.” The output arrives complete and you move on.
What the evidence shows: immediate performance goes up. Unassisted capability and persistence go down. And most people cannot feel it happening while it happens.
Scaffolding
Asking AI to work on your attempt.
“Here’s my draft — what’s weakest?” “Give me a hint, not the answer.” “Argue against this.” You still do the producing.
What the evidence shows: in the randomized studies, the harm disappears. In the best-designed tutoring systems, learning gains roughly doubled.
Where this is still a hypothesis
The strongest evidence here randomizes the system — students were assigned to a tutor that gave hints or one that gave answers. No study has yet randomized the user’s intent. The finding that hint-seekers show no deficit comes from people who chose that mode themselves, so we cannot separate the mode from the kind of person who picks it. That gap is the frontier CPAI’s own research program works on. Everything in this series is consistent with the intent hypothesis; none of it is proof of it.
Five habits, five guides.
Each one takes a single mechanism, states what the research actually shows, and ends with something you can do differently tomorrow.
Extraction vs. Scaffolding
The same tool, opposite outcomes. What decides whether AI builds your capability or erodes it isn't how much you use it — it's how.
Read the guide →What You Hand Off
Offloading is old, useful, and has a price. The question isn't whether to hand things off — it's which capacities you're willing to let thin out.
Read the guide →The Perception Gap
Experienced developers using AI were 19% slower — while believing they were 20% faster. You cannot habit-correct what you cannot perceive.
Read the guide →Effort Is the Active Ingredient
Effortful, self-produced success is the ingredient both learning and mood regulation depend on. Frictionless help removes it.
Read the guide →Boundaries
Across the companion-AI research, outcomes hinge on one thing: whether AI time adds to human contact or replaces it.
Read the guide →The habits
Six plain practices.
Not a scale, not a score. These are the behaviors the guides keep arriving at from different directions.
Choose deliberately
Decide what you're handing off before you hand it off.
Attempt first
Commit to an answer, however rough, before you open the chat.
Verify what matters
Scale checking to the stakes, not to how confident the output sounds.
Keep people primary
For the things that need a person, go to the person first.
Notice your own reliance
Periodically do the thing unaided, just to see where you are.
Keep a no-AI zone
One domain you insist on doing yourself, chosen by you.
Action for every level of influence.
For yourself
- Try one hints-only day: for everything you're trying to learn, ask for the next step rather than the finished answer, and see what changes.
- Commit to an answer before you ask. Even a rough one. The act of generating a position is what the AI answer then gets compared against.
- Pick one skill you refuse to offload — and do it unaided often enough that you'd notice if it slipped.
For an organization
- Make "show your attempt first" a norm in AI-assisted work. The interesting artifact is the difference between the draft and the revision, not the final output.
- Give every verification step a named human owner. "The team checked it" is how unchecked work ships.
- Distinguish deliverables from skills. Extraction is the right call for output nobody needs to reproduce unaided; it is the wrong call for the capabilities you're trying to build in your people.
For educators
- Design tasks where AI hints are allowed but AI answers are not. This is the single change with the strongest evidence behind it.
- Teach the distinction explicitly. Students who think the rule is "AI is cheating" and students who think the rule is "any AI use is fine" are making the same mistake — treating exposure as the variable when mode is the variable.
- Assess the process, not only the product. A revision diff shows you what a finished document cannot.
For policy
- Fund longitudinal research on usage modes. Every study we have measures immediate effects; nobody has followed users over years.
- Require learning products to disclose whether they are designed to give hints or answers. That design choice is the difference between the two outcomes in this series.
- Resist screen-time framings in AI guidance. Fifteen years of research on the previous panic says quantity is the wrong measure.
The other side of the same evidence
The Six Risks name the failure. This series names the habit.
CPAI’s AI Proficiency framework maps six ways AI use goes wrong. Several of them draw on exactly the studies cited here — read from the risk side rather than the practice side. If you want the diagnosis before the habit, start there.
Where this leads
The habits are one thing. Building the skills is another.
These guides teach you how to use AI without losing capability. The Applied AI Certification builds the capability itself — across all six domains, with structured practice and a cohort. They are not the same thing, and that's rather the point.
What this series is built on.
Every empirical claim across the five guides traces to one of these. Evidence tier is stated on each — a randomized trial and a preprint are not the same kind of thing, and we are not going to let the layout imply they are.
Want CPAI to teach healthy AI use in your school or workplace?
We deliver this material as workshops and cohort programs for schools, libraries, employers, and community organizations.