AI made answers cheap, so deciding which ones deserve your trust is the skill that counts. Here is what the research shows, five questions to ask, and habits a team can share.
Critical thinking for AI is a set of small habits, not a talent. Match the check to the stakes, ask what would make an answer wrong, trace it to a source, sort its claims into fact, inference, opinion and unknown, and slow down where you agree fastest. Use the prompt to make AI argue against itself, but open the source yourself, and share the routine with your team.
An AI answer costs almost nothing to produce now, which moves the hard part of the work. Drafting, summarising and explaining take seconds. Deciding whether the result deserves to be believed still takes a person, and the answer gives that person very little to go on, because a wrong one arrives in the same even tone as a right one.
A procurement manager asks for a summary of a supplier proposal, reads it once and forwards it to the team. Nothing in it looked off. The summary was accurate about everything it included, but the renewal condition in the appendix wasn’t in it, and the first anyone heard of that condition was well after the contract was signed.
My argument fits in a sentence. When answers are cheap, the scarce skill is deciding which ones deserve your trust, and that skill is five small habits you can practise rather than a talent some people have: ask what would make the answer wrong, check where it comes from, separate fact from inference, look for what’s missing, and notice when you agree too quickly.
None of this argues for using AI less. It argues for spending your attention on a different part of the job.
The sections below cover what the evidence does and doesn’t say about thinking with AI, a five-question check you can run in a couple of minutes, a way of sorting any answer into fact, inference, opinion and unknown, a prompt that makes the AI argue against itself, and a set of habits a team can share. Where I link to other guides, it’s because they go deeper on one piece, and I won’t repeat them here.
People who trust AI more tend to say they think less critically about what it gives them, and people who trust their own judgment more say they think more. That’s the pattern in a 2025 survey of knowledge workers, and it fits what many people recognise in their own week.
The study is by Lee and colleagues at Microsoft Research and Carnegie Mellon, presented at the CHI conference in 2025.[1] They surveyed 319 knowledge workers who use generative AI at work at least weekly and collected 936 examples of how they used it. It’s worth being exact about what that evidence is, because the headlines around it tend to blur it.
| What the study did | What it can show | What it can’t show |
|---|---|---|
| Asked people whether they felt they had thought critically on a task, and how much effort it took | How workers perceive their own thinking when they use AI | How well anyone actually thought. Nobody’s reasoning was tested |
| Compared those answers with confidence in themselves and confidence in AI | Higher confidence in AI goes with less reported critical thinking; higher self-confidence goes with more | Which comes first. The authors state that the analysis does not establish causation |
| Read 936 examples of real tasks | Effort shifting from gathering and drafting toward verifying, fitting answers into the work, and answering for the result | Whether people do those jobs well |
| Recruited workers online who use AI at least weekly | A view of regular users | Anyone else. The authors note the sample skewed younger and more technical |
Based on the paper’s own description of its method and limitations, read October 23, 2026.
So I wouldn’t say AI makes people less intelligent, and the paper doesn’t say it either. The authors also note that some participants mixed up less effort using the AI with less effort thinking, which is a reminder of how hard this is to measure by asking people.[1]
What I find useful is where the effort moved. People described spending less effort finding and drafting, and more on verifying, fitting answers into their work and being accountable for the result. The thinking didn’t disappear. It changed jobs, from producing to checking, and checking is a skill, which means it can be practised.
Five questions cover the core of what a careful colleague would do with an answer before putting their name to it. I call this the Five-Question Trust Check. It’s a Future Factors working tool built from the habits above, not a validated model.
| # | Question | What it catches | How to ask it |
|---|---|---|---|
| 1 | What would make this wrong? | Answers that only hold under an assumption nobody stated | Ask yourself, or the AI: what would have to be true for this to be wrong, and what did you assume? |
| 2 | Where does this come from? | Claims with no source, or a source that doesn’t say what’s claimed | For each claim that matters, ask which document, page or figure it rests on, then open that source yourself |
| 3 | Which parts are fact, inference or opinion? | Conclusions dressed as facts | Mark each claim with one of four labels. The next sections show how |
| 4 | What’s missing? | Omissions: the clause, the segment, the counter-example | Ask what a careful expert reviewer would expect this to cover and it doesn’t |
| 5 | Did I agree too quickly? | Your own confirmation bias | If the answer matched what you already thought, run question 1 again on purpose |
A Future Factors tool. Run it on one answer at a time, in order, and stop when you’ve found something to fix.
You won’t run all five on everything, and you shouldn’t. Checking a list of workshop title ideas for a team away-day deserves less effort than checking a figure in a board paper. The depth should follow what it costs to be wrong.
The more it costs to be wrong, the more of the check you run.
| Stakes | Examples by role | Questions to run |
|---|---|---|
| Low | A content marketer brainstorming angles for a post; a team lead listing agenda ideas | Question 5 only. Notice whether you liked the answer because it was good or because it was fast |
| Medium | A manager summarising a long thread for the team; an analyst drafting commentary others will skim | Questions 1, 2 and 4 |
| High | Anything with a figure, a quote, a person’s data, a legal or policy point, or an outside reader: a finance analyst’s board numbers, an HR business partner’s policy note | All five, plus a second person who didn’t write it |
A Future Factors working guide. Where a situation sits between two rows, use the higher one.
Much of the effort sits in questions 1 and 2, so those get their own section.
The answers that deserve the closest scrutiny rarely look as though they need it. A weak human draft tends to show its weakness, with gaps, hedges and a missing paragraph. An AI draft on a task it handles badly can look finished.
A marketing manager asks for a competitor comparison for a pitch, gets a clean table with confident one-line verdicts, and drops it into the deck. It’s well laid out. Some of the verdicts rest on guesses about products the model had no information on, and the only hint is that they’re slightly too neat.
Evidence that you can’t see the line from the outside comes from a preregistered experiment with 758 knowledge workers. On realistic tasks inside what the AI handled well, people using it outperformed those without. On one managerial task chosen to sit outside its abilities, they were 19% less likely to produce a correct solution.[2] That used a 2023 model, so I’d hold the exact figure loosely.
For your week, the point is that nothing in the output says which kind of task you’re on. That’s why the first two questions don’t depend on how the answer looks. Take the competitor table. What would make it wrong? The verdicts fail if the model lacked current information on those products. Where does it come from? Ask which source each verdict rests on, then open that source. If the answer is “general knowledge,” you’ve learned something useful before the pitch, not after it.
A polished answer tells you about the writing, not about the facts.
These signs are worth a second look, whatever the tool:
Our guide to fact-checking ChatGPT goes through the verification steps in detail. This guide is about the judgment before you start verifying, which is deciding what is worth checking at all.
The third question changes how you read. An AI answer arrives as a block of confident sentences, and those sentences are doing four different jobs. A fact can be checked against a source. An inference is a conclusion the facts support without stating. An opinion is a recommendation that depends on goals. An unknown is something the answer asserts that the material can’t support.
Once you sort, checking becomes proportionate: facts get verified, inferences get examined, opinions get owned, and unknowns get chased or dropped. Here is one worked example. A product manager pastes a customer feedback export into an AI tool and gets back a short summary.
| Claim in the summary | Sort | What to do |
|---|---|---|
| “Many complaints are about onboarding.” | Fact, if it can be counted | Count the rows that mention onboarding. If the summary can’t say how many, treat “many” as unverified |
| “Customers find the new pricing page confusing.” | Inference | The export has comments that mention pricing. “Confusing” is the model’s reading, so read a handful of the actual comments |
| “The onboarding redesign should be the next priority.” | Opinion | A recommendation that depends on goals and costs only you know. Keep it as an input to your decision |
| “Enterprise customers are the least satisfied group.” | Unknown | The export has no customer-segment column, so the summary can’t know this. Find the data or delete the claim |
A worked example of sorting claims. Replace the summary with your own and sort each sentence before you use it.
The fourth row is the one to build the habit around, because it’s also question 4. The summary never said it lacked segment data. It said something that sounded as though it had it.
Sorting also tells you who to ask. Facts go to whoever owns the data, inferences to someone who knows the customers, and opinions to whoever owns the decision. A finance analyst, an HR business partner and a content marketer will sort the same summary differently, because each knows a different part of what’s missing.
Asking AI to argue against its own answer is the quickest way to run questions 1 and 4, as long as the request is specific enough to produce something you can use. This is the prompt:
| Here is an answer you gave me: [paste the answer]. Act as a sceptical reviewer who has read the same source material I have. List up to five ways this answer could be wrong, from most to least likely. For each one, tell me what I would look for in my own documents to confirm or rule it out. Then list up to three things a careful reviewer would expect this answer to cover that it doesn’t. Don’t rewrite the answer. Mark any point where you are guessing. |
Paste the answer where the brackets are. What comes back is a list of things to check, not a verdict.
Know where this stops before you rely on it:
If you’re building a prompt that a whole team will reuse, our guide to testing AI prompts before you trust the output covers that side. For breaking an AI-drafted structure on purpose, see AI consulting frameworks, which does it in a consulting setting.
The prompt is where AI does the real work in this routine, so it’s worth being explicit about who owns what.
| What AI does | What you still own | How it gets checked |
|---|---|---|
| Lists ways the answer could be wrong and what it leaves out | Which of those matter in your situation, and what you do about each | You open the source for every item you act on, and note what you found |
| Sorts the claims into fact, inference, opinion and unknown | Whether the sort is right, especially the items marked unknown | Spot-check a couple of the items it called facts against the original each time |
| Rewrites the answer after your corrections | The decision to rely on it and put your name to it | Reread the final version against the five questions before it leaves your desk |
A Future Factors working split for any task where AI does real work and a person answers for the result.
One of the hardest answers to question is the one that agrees with you. When a summary confirms what you already thought, the check feels unnecessary, which is exactly when it gets skipped.
An HR business partner suspects that turnover is concentrated in one team, asks AI to look at a set of exit-interview notes, and gets a paragraph that agrees. She quotes it in the leadership update. The paragraph never mentioned the other teams at all, and she didn’t notice, because nothing in it contradicted her.
A lab experiment gives a measured version of that feeling. People using AI to solve logical reasoning problems from the Law School Admission Test scored about three points higher than a comparison group, and overestimated their own results by about four.[3] Those with more technical AI knowledge were more confident and less accurate in judging themselves.
It was a lab task with one test format, with 246 people in the first study and 452 in a replication, so I’d read it as a caution about self-assessment, not a measurement of your week. The practical reading is that how sure you feel after using AI isn’t a reliable signal of how right you are.
Slow down where you agree fastest.
So build the pause into the process instead of relying on mood. Before you open the source, write down how confident you are in one word. Then check, and write down what you found. After a few weeks of this you can see where your confidence runs ahead of the facts, which differs by role. An analyst’s blind spot tends to be the number, a manager’s the plan the summary seems to support, a content marketer’s the claim that sounds right for the brand, and an HR business partner’s anything that describes people. Those are places to look first, not findings.
One person can practise this alone, and the habit lasts longer when a team shares it. The pattern to avoid is a single colleague who checks everything while everyone else forwards. That’s the supervision load our guide to botsitting describes, and it lands on whoever is most careful.
| Habit | Who | When |
|---|---|---|
| State the status when you share AI output: “unchecked draft” or “checked against [source]” | Whoever shares it | Whenever AI output moves between people |
| Say what you checked and what you didn’t, in one line | The author | In the message that carries the output |
| Ask “what’s missing?” before anything goes to a decision-maker | The reviewer | In reviews of any AI-assisted report or summary |
| Keep a short catch log: what the AI got wrong and how it was found | A named owner, rotated monthly | Updated weekly, reviewed at a team meeting |
| Rotate the sceptic: one person argues against the output | Rotating | At each review of high-stakes work |
A Future Factors working list. Adjust the cadence to your team’s rhythm.
Teams can start this on their own with the table above. Training from someone who has taught these habits before shortens the learning curve, builds the same habit across the whole team instead of in one keen person, and lets people practise on their own documents rather than generic examples. Workflow design then gives the check a place to happen, such as a named reviewer on the monthly review. Reinforcement is the weekly catch log and the manager who asks about it, and adoption is what you see when the habit survives a busy month. That’s the work our corporate workshops are built around.
The catch log can be small. Three columns are enough:
| What the AI produced | What was wrong or missing | How it was found |
|---|---|---|
| A summary of a long supplier thread | Left out the revised delivery date | Question 4, checked against the last message in the thread |
| A draft policy note for managers | Described one rule as firm that the policy text calls guidance | Question 2, compared with the policy page |
| Commentary on a monthly variance table | Nothing found | Questions 1 and 2. Logged anyway, so the quiet weeks show too |
A completed example. Replace the entries with your own; the third column is what teaches the team where to look.
To know whether it’s working, look past whether people are doing it:
This week, pick one AI output you’re about to rely on. Run the five questions on it, write down what you found, and share the surprise with a colleague.
The research can’t say that yet. The Microsoft Research survey is self-reported and its authors state it doesn’t establish causation, and the lab studies are short. What you can do is keep some reps without it: draft your own view before you ask, then compare. Our guide to cognitive debt covers how to test that in your own work.
For a medium-stakes answer, a couple of minutes is a reasonable starting point, and much of that is opening the source. That’s a suggestion, not a measured figure. Low-stakes answers get one question. High-stakes ones take longer and should have a second person.
You can, and the prompt above does a version of it, but treat the reply as a list of things to check. The same model shares the same blind spots and can confirm its own mistake. For facts about the outside world, the check is the source itself.
Ask an expert reviewer for question 4 and let them name the gaps, or ask the AI what a specialist in that field would look for and then take that list to a colleague who knows. Building that shared knowledge across a team is a good use of structured training, because the gaps people can’t see alone are easier to find together.
This is a personal habit for one answer at a time. A company checklist is about whether an AI system is ready to scale across a team or process. If you’re deciding that, our AI Trust Checklist takes it from there.
I read these sources on October 23, 2026: the Microsoft Research page and full paper for the CHI 2025 survey of 319 knowledge workers (Lee et al.), the abstract of the Harvard Business School working paper on the jagged technological frontier, and the Aalto University record and arXiv preprint for the study on AI and metacognition (Fernandes et al.). I did not read the full text of the Fernandes journal article, so I have used only what its abstract states. I also left out several studies on brain activity and skill decline in AI-assisted work because they are preprints, rest on small or unclear samples, or measure something other than critical thinking. I did not test any tool hands-on for this article. The Five-Question Trust Check, the claim-sorting method and the team habit list are Future Factors working methods, not research-validated models, and the scenes are illustrations, not client stories.