Most people use AI in one direction.
They ask. AI answers. They edit. They send.
That workflow is fast. It also skips the part where your own thinking gets stronger.
When the answer arrives before your position, you lose the chance to notice what you believe, what you know, and what you have not considered.
I do not want AI to think instead of me.
I want it to make my thinking harder to fool.
The shift
Recent research is pointing to a more useful role for AI than answer machine.
A Microsoft Research survey of 319 knowledge workers found that higher confidence in the AI was associated with less critical thinking. The work also showed that critical thinking does not disappear completely. It moves. People spend less time gathering information and more time verifying, integrating, and supervising the response.
A 2026 study of 120 young adults found that AI improved accuracy on a logical fallacies task and reduced reported mental effort. But the researchers saw a sharp difference between strategic engagement and complete delegation. The same tool can act as a scaffold or a substitute.
That distinction is the skill.
Do not ask AI for the answer first. Ask it to improve the quality of the question, the options, and the reasoning around your decision.
The thinking loop
1. Start with your own rough view. Write the problem, your current position, the evidence you have, and what feels uncertain. This gives AI something to examine instead of something to replace.
2. Ask for questions. Tell AI to ask what a skeptical colleague, customer, or board member would ask before it offers a recommendation.
3. Ask for alternatives. Generate different approaches, not more versions of your first idea. Ask what each option makes easier, harder, and more likely.
4. Rethink your position. Read the objections. Change your view where the evidence requires it. Keep the parts that survive.
5. Run a reasoning audit. Ask AI to separate facts, assumptions, inferences, and recommendations. Ask what evidence would change the conclusion.
6. Decide yourself. The final output should include your rationale, not just polished language. If you cannot explain the decision without the chat open, the work is not finished.
Microsoft Research calls this direction “better thinking through AI.” The phrase matters. The goal is not to keep every bit of effort. The goal is to spend human effort on judgment, meaning, and consequences.
Use AI to widen the frame. Then bring the decision back to yourself.
1. Microsoft Research: Better Thinking Through AI
A March 2026 guide with ten practical strategies. The best starting points are asking questions, surfacing overlooked issues, offering alternatives, and using AI to support learning.
Read it here
2. AI, Help Me Think, but for Myself
In this Microsoft Research study, an AI that extended a person’s own rationale integrated better with the decision process than a recommendation-first AI. People stayed closer to their reasoning.
Read the study
3. A 2026 study on cognitive offloading
AI improved accuracy and reduced mental effort in a logical reasoning task. The authors’ key warning is that strategic engagement and complete delegation produce different outcomes.
Read the abstract
Use this before asking AI to solve an important problem.
Do not solve this immediately. Help me think. My problem is: [describe the decision, project, or question] My current view is: [state what I currently believe and why] My known evidence is: [list the facts, sources, and constraints] First, ask me five difficult questions that would expose missing context, weak assumptions, or an unclear success measure. Wait for my answers. Then give me three genuinely different approaches. For each one, explain the upside, downside, likely consequence, and what it assumes. Then argue the strongest case against my current view. Finally, separate the result into facts, inferences, open questions, and a recommendation. Tell me what evidence would change the recommendation. Do not make the decision for me. Keep my judgment visible.
“Wait for my answers” creates a pause before the model starts producing certainty.
“Keep my judgment visible” prevents a polished recommendation from hiding who actually made the call.
Where do you reach for AI too early?
The first draft? The decision? The plan? The reply you have not thought through yet?
Reply with the moment where you want AI to slow your thinking down before it speeds your work up.
Thinking with AI gets easier when the assistant has a stable picture of how you work and what good means to you.
My guide on building an AI second brain shows how to capture your working context, structure it, and retrieve it when the task begins.
For the smaller building blocks, the free Prompting Hub includes 19 laws and 32 cheat codes. No email required.
More next week.
Kapish
Enterprise AI Architect · T-Minus AI