Keywordsusing ai without losing critical thinkingBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searchesai and critical thinking at workhow to keep thinking when using chatgptai deskillingchatgpt critical thinking at workavoid ai deskilling professionalsusing ai without losing critical thinking
Keep the hard step human: question, check, decision
Critical thinking at work is not a mood. It is three labours: framing a question that is worth answering, checking claims against the world, and taking a decision you will own. ChatGPT is good at producing a fluent substitute for all three. That is the deskilling risk. You open a chat instead of writing the question. You accept the first frame because it is tidy. You skip the check because the paragraph already has a number. You let the last sentence choose. Using AI without losing critical thinking means the model may draft, list objections, or rephrase. It does not get the question, the check, or the decision. Those stay with a named human, every time.
Watch a junior analyst who used to build the query and now pastes the request into chat. Within a month they still produce commentary, and they can no longer say which filter dropped the internal transfers. The commentary still looks senior. The skill did not compound. The same pattern hits lawyers who stop opening the clause, operators who stop walking the floor, and managers who stop writing the decision they want before the meeting. Fluency hides the loss until a nasty week, when the model is confident and the file is not. The nasty week is too late to rebuild the habit. Build it now by forcing the hard step to happen before generate and after generate.
This essay is a practice, not a sermon against tools. You can use ChatGPT every day and still get sharper if the question is written first, the check is a file you open, and the decision is a sentence you would sign. You can also avoid ChatGPT and still think poorly. The tool is an amplifier. If you bring a blank mind, you get a fluent blank. If you bring a framed problem, you get a draft you can argue with. Arguing is the point. Agreement with a model is not thinking. It is comfort. Comfort is how judgement thins.
If you cannot state the question, the check, and the decision without opening the chat, you are already outsourcing the hard step. Write those three lines first.
Where thinking still has to happen, even with a draft
The question is hard because it includes what you will ignore. ChatGPT will answer the prompt you typed, including a bad one. If the real question is whether the north warehouse should close, you needed a different frame. Write the decision the work must support before you generate. The check is hard because it is physical: a cell, a clause, a ticket, a call. A second chat that agrees with the first is not a check. The decision is hard because it names a no, a wait, or a spend. Models prefer a balanced path that offends nobody. Balanced is not a decision. You still have to pick, and you still have to be the name on the pick.
Use the model on purpose as a challenger after you have a view, not as an oracle before you have one. Ask for the case against your recommendation, the facts you would need, and the way a sceptical finance partner will attack the slide. Then you go and look. That sequence trains thinking. The reverse sequence, generate a recommendation then admire it, trains dependence. Dependence feels like speed until the first time the recommendation was a pattern completion from some other industry's blog, and you are the person in the room. Keep a record of decisions you took against the model's preferred close. That record is evidence you still work here.
| Hard step | If the model takes it | If you keep it |
|---|---|---|
| Question | A tidy frame you did not choose | The decision the work must support |
| Check | A second chat that agrees | A file, a cell, a clause, a call |
| Decision | A balanced path with no owner | A signed no, wait, or spend |
| Practice | Commentary you cannot rebuild | A log of filters, sources, and picks |
A loop that uses the model without handing it the judgement
Write the question and your current view in five lines. Generate only after that. Ask for objections, missing facts, and a critique of your frame, not for the answer. Then run the check in the real system. Then write the decision yourself, including wait. If you want a draft of the artefact, generate that last, from the decision you already took. Do not reverse the order because the draft is more fun. Fun is how the question disappears. For juniors, require the five lines in the pack they show you. If they bring only a chat, send it back. That is teaching, not hostility to tools. The skill that still compounds is the habit of framing, checking, and owning.
Schedule thinking that does not involve a model. One analysis a week rebuilt from the extract without a chat. One meeting decision written before anyone opens a copilot. The point is not purity. The point is that the muscle still fires when the tool is down, the file is messy, or the fluent answer is wrong. Deskilling is gradual. So is the cure. Five lines before generate is a small enough habit to survive a busy week. A weekend course on critical thinking will not, if Monday's workflow still starts with a blank prompt and a hope that the model will know what you meant.
Role: You are a sceptical colleague, not a decision maker.
Task: Challenge the question and the view I paste. Do not replace them.
Context: I will state the decision this work must support and the facts I already have.
Constraints:
- Do not give a recommended option as if I had chosen it.
- List objections, missing facts, and attacks a finance partner might make.
- Mark [missing] where I have not given evidence. Do not invent the evidence.
Output: Critique of my frame, a list of checks I should run, and questions I should answer myself.
Quality checks: Where this reply might tempt me to skip the file and take your wording as the decision.Run the checks it listed in the real system, not in another chat. Write the decision in your own sentence, including no or wait if that is the honest pick. Only then draft the artefact, if you still need a draft. Save the five lines that started the loop. If a junior brings you a chat without those lines, send it back. That refusal is how you use AI without losing critical thinking on a team, not only in your own window.
First-frame capture, skipped checks, and comfortable consensus
First-frame capture is when the model's tidy problem statement becomes the problem. You forget the warehouse question and chase the marketing story because it was well written. Interrupt that by keeping your five lines visible while you read. Skipped checks are when a number in the draft is close enough. Close enough is how commentary drifts. Comfortable consensus is when a group uses the same chat and treats the output as the room's view. It is not the room's view. It is one completion. Make people speak before the screen is shown. If the meeting cannot decide without the generated option list, you have already handed over the hard step. Take it back. A list of options is not thinking.
Do not confuse role-play with judgement. Asking the model to be a hostile board member can surface questions. It cannot tell you what you should do. The board member is fictional. The capital is real. Also watch performance metrics that count artefacts produced with AI. You will get more pages and thinner questions. Count decisions that survived contact with the file, and people who can still rebuild the analysis. That is the anti-deskilling measure. Everything else is theatre about innovation. Theatre is how organisations lose the habit and keep the vocabulary.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Blank prompt first | The model chooses the question | Five lines of view before generate |
| Chat as check | Another window agrees | Open the file and rerun |
| Balanced close | A path that decides nothing | Write no, wait, or spend |
| Junior brings only chat | Commentary nobody can rebuild | Require the frame and the filter |
If the model wrote the question and the close, you did not think with AI. You watched a completion. Write the question and the decision yourself, or do not send the page.
Related reading on StudyGrid
Read next: Skills That Still Compound AI as a Thinking Partner Human and AI. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
On the next real piece of work, write five lines before you open any chat: the question, your current view, the check you will run, and that you will decide. Use the model only to attack that view. Run the check in the file. Write the decision. Draft last if you must. Repeat for five working days. That is using AI without losing critical thinking, as a practice rather than as a slogan.