How to use AI to think better without letting it think for you

Artificial intelligence offers a powerful temptation: stop thinking as soon as a plausible answer appears.

A question goes into a chat. The model generates a convincing response. We copy it, send it, and move on.

That pattern can save time, but it can also turn AI into a kind of cognitive autopilot.

There is a more interesting alternative: use it to force ourselves to think better.

Jeremy Caplan recently made that case in Fast Company in an article titled “5 ways to use AI to sharpen your thinking”. His starting question is simple: how do we benefit from AI without falling into a kind of cognitive surrender where we outsource the mental work itself?

The answer is not necessarily to use less AI.

It is to change the role we assign to it.

Instead of asking it to think for us, we can ask it to:

  • challenge our ideas;
  • identify gaps in our reasoning;
  • organize thoughts that are still incomplete;
  • investigate before a major decision;
  • teach us through a personalized curriculum;
  • or turn learning into a measurable system.

That may look like a small shift, but it changes the relationship with the tool completely.

The problem is not using AI. It is which part of thinking we outsource

In 2025, researchers at Microsoft Research published a CHI study on the impact of generative AI on critical thinking among knowledge workers. They surveyed 319 professionals and analyzed 936 real-world examples of AI use in work tasks.

The most interesting result was not that AI eliminated critical thinking.

It was that critical thinking moved.

When users reported higher confidence in AI, they also reported less critical-thinking effort. When they had greater confidence in their own ability to perform the task, they tended to exercise more critical judgment.

The researchers also found that thinking shifted away from producing everything from scratch and toward verification, integration, and task stewardship.

That suggests a fundamental distinction:

weak pattern
problem → AI → answer → accept

stronger pattern
problem → own hypothesis → AI → critique → verification → decision

AI does not have to replace reasoning.

It can become an additional layer of intellectual friction.

1. Ask it to disagree with you

One of the most useful ways to use a model is almost the opposite of the most common one.

Instead of asking:

Do you think this is a good idea?

try something like:

Here is my hypothesis.

Try to prove me wrong.
Identify my three weakest assumptions.
Find alternative explanations.
Tell me what evidence would change my conclusion.

The difference matters because language models are very good at generating plausible explanations in the direction we already gave them.

If we want more value, we should introduce deliberate opposition.

We can ask the model to adopt multiple roles:

financial analyst
security engineer
skeptical customer
competitor
devil's advocate
technical reviewer
production operator

Each one will notice different risks.

A software architecture can look elegant to the developer and dangerous to the operator who will have to maintain it at three in the morning.

A product can look compelling to the founder and confusing to the person seeing it for the first time.

A business strategy can appear efficient from an operations perspective and fragile from a security perspective.

AI makes it cheap to simulate those viewpoints.

The Reviewer pattern

Software engineering already has a similar pattern:

Developer → Reviewer → QA

The developer proposes.

The reviewer tries to find flaws.

QA tries to break the system with evidence.

We can apply the same structure to ideas:

Thinker → Critic → Verifier

The first role builds the hypothesis.

The second tries to destroy it.

The third looks for evidence.

That loop often produces a better decision than asking for one supposedly “correct” answer.

2. Think out loud before asking for a solution

Another powerful tactic is to use AI as a thought structuring tool, not as the first generator.

Speaking out loud has an interesting advantage: it forces us to externalize ideas that are not organized enough yet to become polished writing.

We can record two or five minutes of whatever is in our head:

what I want to do
what worries me
what alternatives I see
what constraints exist
what I still do not understand
what decision I am trying to make

Then AI enters the process.

Its job is not to invent a position.

Its job is to transform the raw material into something inspectable:

transcript
   ↓
main ideas
   ↓
assumptions
   ↓
open questions
   ↓
contradictions
   ↓
possible next steps

That preserves something valuable: the raw material remains yours.

AI acts as a cognitive editor.

The sequence matters.

If we ask AI for an answer first, our thinking can become anchored to the first structure the model gives us.

If we generate our own material first, AI works on a base that already contains our perspective.

3. Use Deep Research before an important decision

Some questions are too consequential for a quick chat answer.

Choosing an important technology.

Selecting an architecture.

Switching vendors.

Choosing an AI model.

Entering a new market.

Evaluating a job offer.

Designing an investment strategy.

In those situations, the goal is not a fast answer.

It is to reduce uncertainty.

Deep research modes in tools such as ChatGPT, Gemini, Claude, Copilot, or Perplexity can turn a broad question into a multi-source investigation.

But research quality depends heavily on the specification.

A strong research prompt should include:

objective
context
specific questions
constraints
preferred sources
sources to avoid
time window
output format
criteria for comparing alternatives

For example:

I am evaluating three providers for running AI models.

Compare:
- real workload cost;
- latency;
- regional availability;
- rate limits;
- privacy;
- tool support;
- API stability.

Prioritize official documentation and reproducible benchmarks.
Separate facts, inference, and opinion.
Include original sources.

That last point is essential.

Deep Research does not remove the need to verify.

It changes where we spend our verification effort.

Instead of manually investigating every page from scratch, we can focus on checking the claims that would actually change the decision.

4. Make AI teach instead of answer

There is a huge difference between these instructions:

Give me the answer.

and:

Help me reach the answer.

The second turns AI into a tutor.

A good tutor does not immediately answer every question. It asks questions, identifies gaps, and adjusts difficulty.

We can make that behavior explicit:

I want to learn neural networks.

Do not give me answers immediately.
Ask questions to test my understanding.
If I am wrong, give me a hint before explaining.
Increase difficulty progressively.
Every so often, make me retrieve earlier concepts without warning.

That final instruction matters.

Learning does not mean recognizing an explanation when we see it.

It means being able to retrieve and apply knowledge without the answer in front of us.

AI can generate:

  • exercises;
  • retrieval questions;
  • short exams;
  • analogies;
  • progressive problems;
  • practical projects;
  • study plans;
  • spaced reviews.

It can also adapt examples to the learner’s context.

A software developer can learn statistics through latency, observability, and A/B testing examples.

An entrepreneur can learn probability through sales, conversion, and risk scenarios.

Personalization turns the model into something closer to an interactive learning environment than an encyclopedia.

5. Build a learning dashboard

The fifth idea sounds less philosophical, but it can be one of the most effective.

If we want to learn consistently, it helps to make learning visible.

We can ask an agent or software-generation tool to build a small dashboard that tracks:

what I studied
how much time I spent
what exercises I completed
which concepts I missed
what I need to review
which projects I finished
which questions remain open

It does not have to become a complex platform.

A Markdown file, a small local app, or a spreadsheet may be enough.

The point is to create a feedback loop:

learn
  ↓
practice
  ↓
measure
  ↓
find gaps
  ↓
review
  ↓
learn better

AI can help at every transition.

But the system preserves something isolated chats often lose: history.

We can see what we are learning and where we are still failing.

A protocol for not outsourcing your thinking

All of this can be compressed into a simple five-step workflow.

Step 1: produce your own position

Before asking AI, write or dictate what you think.

Even if it is incomplete.

Even if it is probably wrong.

Step 2: ask for opposition

Make the model identify flaws, alternatives, and risks.

Do not ask for validation.

Step 3: separate facts from interpretation

Ask for an explicit structure:

confirmed facts
assumptions
inferences
opinions
uncertainties

Step 4: verify the claims that matter

Not every sentence deserves the same amount of checking.

Focus on the claims that could change the decision.

Step 5: decide yourself

The model output is input evidence.

It is not the decision.

The stack should look like this:

YOUR JUDGMENT
     ↑
verification
     ↑
AI critique
     ↑
your hypothesis
     ↑
real problem

The person remains at the top of the stack.

The better question is not “can AI do this for me?”

A more useful question is:

Which part of this task do I still want to exercise myself?

Sometimes the answer will be: none of it.

We do not need to mentally practice a trivial unit conversion or manually rewrite one hundred rows of data.

But in decisions, learning, strategy, architecture, research, or creativity, delegating too early can remove the most valuable part of the work.

Thinking is not merely producing an answer.

It is forming questions, building hypotheses, finding contradictions, updating beliefs, and deciding under uncertainty.

The best AI does not have to replace that process.

It can become a machine that puts pressure on it.

And that may be one of the most powerful ways to use these systems: not as a prosthetic that helps us avoid thinking, but as an adversary, tutor, and laboratory that makes us think better.

Sources