F
Francesc
Guest
Disclosure: I am the author of Reclaim Your Mind, a book about attention, digital wellbeing and independent thinking.
A useful question to ask before opening an AI tool is: which part of this task do I want to do myself?
That question rarely appears in a prompt box. The interface is ready for a request, and the next response is only a click away. Yet deciding what to delegate is itself an act of judgment. If we skip it, convenience can make the decision for us.
Consider a simple example. You need to explain a difficult concept to a colleague. You could ask a model to produce the explanation immediately. You could also write three rough sentences first, mark what you cannot explain, and ask the model to challenge those gaps. Both workflows use AI. They give your own thinking different roles.
My argument is that we should design more opportunities for that choice, especially in tasks where understanding matters as much as the finished output.
A CHI 2025 study by Lee and colleagues surveyed 319 knowledge workers, collecting 936 examples of AI use. Higher confidence in generative AI was associated with less reported critical thinking, while greater task-specific self-confidence was associated with more. Participants also described thinking moving toward verification, integration and oversight.
This is self-reported survey evidence. It does not establish that using an AI tool causes permanent cognitive damage. It does give us a reason to ask where thinking occurs in a workflow and how users decide that an answer deserves trust.
For organizations addressing these questions more broadly, NIST's Generative AI Profile provides a voluntary resource for considering trustworthiness across the design, use and evaluation of AI systems. An individual's habits are one part of a wider design problem.
The first pause comes before generation. Write down the purpose of the task and your initial view. For a difficult decision, that might be one preferred option, one reason for it and one uncertainty. The aim is to make your starting point visible, so that you can later tell whether the model supplied evidence or merely persuasive phrasing.
The second pause comes before acceptance. Separate the response into claims you can verify, suggestions you may try and decisions that depend on your priorities. Ask what evidence would change the conclusion. If the answer cites a source, open that source and check that it supports the claim. A model's explanation of its own answer is another response to examine.
The third pause comes before the next prompt. Close the response and explain the main idea in your own words. If you cannot, decide whether the task requires understanding before you proceed. Sometimes it does; sometimes the task is a routine transformation whose accuracy you can check directly. The useful distinction is the requirement of the task, not a blanket rule that every action must be slow.
These are proposed habits, not a clinically validated program. Their value should be assessed against the work you actually do.
A warning that appears on every screen can become another thing to dismiss. A stopping point should ask for something specific.
Imagine a team preparing a product proposal. Before generating the document, the team records its intended user, the problem it has evidence for and the assumptions it still needs to test. After generation, reviewers can compare the prose with that short record. A convincing new claim without supporting evidence becomes visible as an unresolved question.
Superorange’s “Stop Asking AI to Write the PRD” on HackerNoon approaches a related problem through traceable requirements: a polished document needs explicit evidence and decisions behind it. My interest is the everyday habit that supports that discipline. Before accepting an answer, identify what makes it acceptable.
The same idea can work outside product development. A student can attempt a problem before requesting a hint. A writer can choose the argument before asking for alternative structures. A manager can identify a decision's constraints before asking for options. Each example preserves a particular contribution from the person using the tool.
An AI conversation can always continue. Another version, another comparison and another refinement remain available. I suggest defining a stopping condition before starting: the question has been answered, the key claims have been checked, or the remaining uncertainty has been recorded for someone able to resolve it.
The appropriate stopping condition depends on the consequences of getting the task wrong. A brainstorming exercise and a decision affecting other people deserve different levels of review. Adding friction everywhere would waste attention that could be used where it matters.
In Reclaim Your Mind, I describe the underlying principle this way: “Every time you introduce intentional friction before using AI, you’re exercising your right to think before you delegate.”
The practical test is modest: after a session with AI, can you explain what you asked it to do, why you accepted its contribution and what still requires your judgment? A useful tool should leave those answers within reach.
A useful question to ask before opening an AI tool is: which part of this task do I want to do myself?
That question rarely appears in a prompt box. The interface is ready for a request, and the next response is only a click away. Yet deciding what to delegate is itself an act of judgment. If we skip it, convenience can make the decision for us.
Consider a simple example. You need to explain a difficult concept to a colleague. You could ask a model to produce the explanation immediately. You could also write three rough sentences first, mark what you cannot explain, and ask the model to challenge those gaps. Both workflows use AI. They give your own thinking different roles.
My argument is that we should design more opportunities for that choice, especially in tasks where understanding matters as much as the finished output.
Be precise about the concern
A CHI 2025 study by Lee and colleagues surveyed 319 knowledge workers, collecting 936 examples of AI use. Higher confidence in generative AI was associated with less reported critical thinking, while greater task-specific self-confidence was associated with more. Participants also described thinking moving toward verification, integration and oversight.
This is self-reported survey evidence. It does not establish that using an AI tool causes permanent cognitive damage. It does give us a reason to ask where thinking occurs in a workflow and how users decide that an answer deserves trust.
For organizations addressing these questions more broadly, NIST's Generative AI Profile provides a voluntary resource for considering trustworthiness across the design, use and evaluation of AI systems. An individual's habits are one part of a wider design problem.
Three pauses worth trying
The first pause comes before generation. Write down the purpose of the task and your initial view. For a difficult decision, that might be one preferred option, one reason for it and one uncertainty. The aim is to make your starting point visible, so that you can later tell whether the model supplied evidence or merely persuasive phrasing.
The second pause comes before acceptance. Separate the response into claims you can verify, suggestions you may try and decisions that depend on your priorities. Ask what evidence would change the conclusion. If the answer cites a source, open that source and check that it supports the claim. A model's explanation of its own answer is another response to examine.
The third pause comes before the next prompt. Close the response and explain the main idea in your own words. If you cannot, decide whether the task requires understanding before you proceed. Sometimes it does; sometimes the task is a routine transformation whose accuracy you can check directly. The useful distinction is the requirement of the task, not a blanket rule that every action must be slow.
These are proposed habits, not a clinically validated program. Their value should be assessed against the work you actually do.
Give the pause a job
A warning that appears on every screen can become another thing to dismiss. A stopping point should ask for something specific.
Imagine a team preparing a product proposal. Before generating the document, the team records its intended user, the problem it has evidence for and the assumptions it still needs to test. After generation, reviewers can compare the prose with that short record. A convincing new claim without supporting evidence becomes visible as an unresolved question.
Superorange’s “Stop Asking AI to Write the PRD” on HackerNoon approaches a related problem through traceable requirements: a polished document needs explicit evidence and decisions behind it. My interest is the everyday habit that supports that discipline. Before accepting an answer, identify what makes it acceptable.
The same idea can work outside product development. A student can attempt a problem before requesting a hint. A writer can choose the argument before asking for alternative structures. A manager can identify a decision's constraints before asking for options. Each example preserves a particular contribution from the person using the tool.
Leave room to finish
An AI conversation can always continue. Another version, another comparison and another refinement remain available. I suggest defining a stopping condition before starting: the question has been answered, the key claims have been checked, or the remaining uncertainty has been recorded for someone able to resolve it.
The appropriate stopping condition depends on the consequences of getting the task wrong. A brainstorming exercise and a decision affecting other people deserve different levels of review. Adding friction everywhere would waste attention that could be used where it matters.
In Reclaim Your Mind, I describe the underlying principle this way: “Every time you introduce intentional friction before using AI, you’re exercising your right to think before you delegate.”
The practical test is modest: after a session with AI, can you explain what you asked it to do, why you accepted its contribution and what still requires your judgment? A useful tool should leave those answers within reach.