How to use AI without getting dumber

Updated 30 September 2026.

To use AI without getting dumber, keep the parts of the job that make you think. Take a first pass before you prompt, guess the answer before you read it, ask the AI to explain, and check one claim every time. Each of those habits comes from a study, and each takes a few minutes.

Last verified 30 September 2026. Each habit links the study it comes from.

1. Take a first pass before you prompt

Write three bullet points, a rough outline or a failing first attempt at the code, then bring in the AI. In the MIT essay study, people who wrote on their own for three sessions and then got ChatGPT showed higher brain connectivity than the ChatGPT group had shown in its own sessions, and the paper reports higher recall for them.arXiv Decades of memory research point the same way: material you produce yourself is remembered better than material you only read, by close to half a standard deviation across 86 studies.meta-analysis

2. Guess the answer before you read it

Before you open the reply, write down what you expect it to say, even in one line. Then compare. Recalling from memory beats rereading for anything you want to keep a week later, even though rereading feels more effective.paper The guess also gives you something to disagree with, which matters: in the Wharton experiments, people who consulted a chatbot accepted its wrong answers most of the time.Ars Technica

3. Ask why, and ask for the explanation with the answer

In Anthropic's coding trial, three habits kept people's understanding intact while they still used AI: asking only conceptual questions, asking for an explanation alongside generated code, and generating code and then questioning it. Those groups scored 65% to 86% on the quiz. Handing the task over, sliding into full reliance, or asking the AI to fix each error scored 24% to 39%.arXiv The groups were small, between two and seven people each, so treat the exact numbers loosely.

4. Check one claim every time, and slow down when rushed

Pick the claim that matters most in each answer and verify it somewhere else. In the Microsoft and Carnegie Mellon survey, people did less critical thinking on tasks where they trusted the AI more, and time pressure was one of the barriers they named.paper In the Wharton experiments a 30 second timer cut the rate at which people overruled wrong answers, while feedback on each answer raised it.Wharton

5. In the evening, say what you did from memory

Write two or three lines about what you finished, decided or learned, without looking at the chat. If you cannot explain a change the AI made, that is the one to go back and read. In the MIT study the ChatGPT group had the hardest time quoting essays they had written minutes before.arXiv

6. Keep a few reps without the tool

For any skill you want to keep, do part of it by hand on purpose. The maths students who practised with plain GPT-4 did worse on the exam than students who never had it.PNAS The doctors in the colonoscopy study found adenomas less often without the AI after months of working with it, though that study was observational.Lancet

What does a ten minute day look like?

Morning, two minutes: write the one thing you want done today, in your own words, before opening any AI tool.
Before each big prompt, one minute: a rough outline or your own first guess.
After each important answer, one minute: find the claim that matters most and check it.
Evening, five minutes: score your day out of ten, write what you did from memory, then compare it with what actually happened.

Where does Mover OS fit?

Mover OS builds several of these habits into workflows that run inside Claude Code, Codex or Gemini CLI. The evening review asks for your score and your biggest win and biggest avoidance before it shows its analysis, then names the gap between your guess and the data. The daily note has a log and a journal you type yourself. After a long session where the agent worked on its own, it offers a three question quiz on what changed.

It does none of this unless you run it, it cannot stop you handing a whole task to your agent, and it has never been studied. The habits above work with any AI tool and cost nothing. Mover OS is not for anyone who wants a phone app or a chat window: it needs one of those three terminal agents.

Common questions

Should I stop using AI to protect my thinking?

The trials do not point there. Students with a tutor designed to protect learning improved during practice and largely avoided the exam drop, and developers who questioned the AI kept their understanding. The usage pattern mattered more than whether AI was used at all.

Does ChatGPT have a mode that makes me think more?

Anthropic's own write up of its coding trial names ChatGPT Study Mode and Claude Code's Learning and Explanatory modes as options designed to build understanding. Their effect was not measured in that trial.

Is it bad to use AI for boring tasks?

In the Microsoft survey, people reported putting less critical thinking into routine, low stakes tasks with AI, which is often a reasonable trade. The habits here matter most for work you need to understand or defend later.

Sources

Mover OS: $49 once. Runs inside Claude Code, Codex or Gemini CLI. The evening review asks for your own score first. Does not run while the agent is closed.

Get Mover, $49 once