How can people use AI without weakening their critical thinking?

What critical thinking means in an AI-assisted task

Critical thinking is not merely being skeptical or finding faults. It is the ability to define a question, separate evidence from assertion, identify assumptions, consider alternatives, judge uncertainty, and take responsibility for a conclusion. You can use a calculator, search engine, spreadsheet, colleague, or AI while thinking critically. The key question is whether you could explain the reasoning and recognize what would change your mind.

AI is strongest when it expands the set of possibilities you inspect. It is weakest when its answer becomes invisible infrastructure for a decision you cannot explain. A useful habit is to keep the human role at the points where error matters: deciding the question, choosing relevant evidence, setting standards, resolving tradeoffs, and accepting responsibility for the result.

Generative AI also has a particular reliability problem. NIST calls it "confabulation" when a system confidently generates false or erroneous content, including misleading logic or citations. This is not a rare formatting mistake that a polished prompt can completely remove. It is a risk that needs checking, especially in open-ended or consequential work. NIST Generative AI Profile

Put the work in the right place

Not every task deserves the same amount of unaided effort. Copying a list into a calendar, formatting notes, or brainstorming headings is reasonable offloading. Learning a concept you must later use, evaluating evidence for a recommendation, deciding how to spend money, or approving work that affects others requires more direct engagement.

The table is a rule of thumb, not a claim that routine work never deserves care. The higher the cost of being wrong and the more likely you need to explain the decision later, the less you should outsource the central reasoning. In medicine, law, finance, safety, employment, or decisions affecting other people, an AI response is not a sufficient basis for action.

Use an attempt, challenge, verify, decide, reflect loop

The following loop keeps the most valuable mental work with the person using the tool. It also makes AI useful even when the first answer it gives is imperfect.

Attempt first

Before asking AI for a solution, write a brief provisional answer. This can be a sketch, calculation, claim, outline, or list of assumptions. Set a time limit if the task is large: five minutes for a simple question, perhaps 20 minutes for a harder problem. This helps identify what you already know and where a more specific explanation would help.

For learning, this is especially important. Retrieval practice means trying to recall or produce an answer from memory, rather than only rereading a correct explanation. In a controlled study, repeated testing improved delayed recall more than repeated study after learning, and the authors found that students were not good at predicting their own performance. Karpicke and Roediger, Science If AI supplies the answer before you attempt retrieval, it can create recognition without demonstrating that you can produce or use the idea.

Useful first-attempt records include: "My current conclusion is X because A and B," "I am assuming C," and "I cannot explain step 3." These notes give AI something concrete to challenge and later let you see whether you changed your mind for a good reason.

Ask for teaching, not completion

The same model can either replace your work or help you inspect it, depending on the request. Ask for questions, hints, contrasting approaches, definitions at the right level, and a critique of your explanation. Do not begin with "give me the answer" when the skill is the thing you are trying to develop.

Try prompts such as:

• "Ask me one question at a time until I can explain this concept. Do not give the final answer unless I ask."
• "Here is my solution. Identify the first unsupported step, then ask a question that helps me test it."
• "Show two plausible approaches, state the assumptions behind each, and tell me what evidence would distinguish them."
• "Act as a skeptical reviewer. Name the strongest counterargument to my conclusion and the evidence I would need to answer it."

These requests do not make the system a reliable teacher by themselves. You still need to check the content. They do make it less likely that you will passively consume a finished response. UNESCO's guidance on generative AI in education calls for a human-centred approach and stresses the need for learners and educators to take a critical perspective toward generated outputs. UNESCO guidance

Generate alternatives and counterarguments

People often use AI to make the first plausible plan sound more polished. A stronger use is to ask it to make that plan compete. Have it produce a serious alternative, a counterexample, a failure mode, or a stakeholder perspective you may have missed. Then judge the alternatives using criteria you state in advance.

For a work decision, ask for a table of options that includes upside, downside, assumptions, evidence needed, and conditions that would make each option fail. Do not accept the categories just because the tool proposed them. Add the constraints that matter in your setting, such as cost, accessibility, legal obligations, customer impact, reversibility, and operational risk.

For a belief or explanation, ask the tool to steelman the strongest reasonable opposing view. A steelman is the best fair version of an opposing argument, rather than its weakest version. Then ask what observation would make your current view less credible. Use the response to examine an assumption you might otherwise overlook. It does not mean that every alternative deserves equal weight or that the tool's output settles a disagreement.

Verify claims that could change your action

Ask AI for sources, but treat the list as a lead, not as verification. Open the cited source. Check its publisher, date, authorship, method, scope, and whether it actually supports the sentence you want to use. Confirm numbers with the primary source or original data where possible. If you cannot find the source, do not repeat the claim.

Do not be reassured merely because two models agree. They may have learned from overlapping material, copied the same earlier error, or generated the same plausible pattern. Agreement can tell you that a claim is common, not that it is true. Independent verification means consulting evidence that does not depend on the model's wording.

Use a stronger check as the consequences rise:

NIST notes that confabulated outputs can include citations and reasoning that appear to justify an incorrect answer. NIST Generative AI Profile That is why source checking is a thinking task, not a ceremonial click on a reference list.

Decide in your own words

Before acting, write a two- or three-sentence decision record: the choice, the main evidence, the important uncertainty, and the reason you accepted one tradeoff over another. If you cannot write it without copying the AI response, you probably do not yet own the reasoning.

This record is useful even for a small decision. It exposes whether the answer depends on an unsupported claim, whether a missing fact matters, and whether you have confused the tool's confidence with your own judgment. For collaborative work, it also gives colleagues something concrete to review instead of forcing them to guess which parts were generated and which were evaluated.

Reflect after use

Finish with a short review:

1. What did I do before asking for help?
2. What did AI add that was genuinely useful?
3. What did I reject, correct, or fail to verify?
4. What will I try to do without AI next time because it is a skill I need to retain?

This is not paperwork for its own sake. It helps you notice whether the tool is teaching, accelerating, or quietly taking over a capability you value. Change the workflow if you repeatedly skip the attempt, cannot explain the final result, or find that the tool is doing the same core reasoning every time.

Work example

Hypothetical example: A product manager must recommend whether to delay a feature for an accessibility fix. Before using AI, she writes the decision criteria: user harm, legal and contractual commitments, engineering effort, release impact, and whether a temporary mitigation is real or merely cosmetic. She writes her preliminary view and names the data she lacks.

She then asks AI for the strongest case for shipping, the strongest case for delaying, assumptions that each case depends on, and questions a reviewer would ask. The response surfaces a possible customer-support burden and a documentation risk. She does not assume either exists because the tool mentioned it. She checks current support data, consults the accessibility lead and legal or policy owner, and revises the recommendation using the evidence they provide.

Her final note says why she chose the option, what risk remains, who reviewed it, and what would cause reconsideration. The useful contribution from AI was better coverage of alternatives. The critical thinking was setting the criteria, testing the claims, assigning weight to tradeoffs, and being accountable for the decision.

Study example

Hypothetical example: A student is learning why a historical policy produced an unintended effect. First, the student closes all AI tools and writes a short explanation from the assigned reading, including two pieces of evidence and one uncertainty. The student then asks AI to pose three questions that would expose a weak causal link in the explanation, without rewriting the answer.

After answering those questions, the student opens the course readings and primary sources to check the dates, language, and interpretation. Only then does the student ask AI to compare the revised explanation with a competing interpretation, while citing what would count as evidence for each. The student writes the final response in their own words and discloses or follows the course policy for AI use.

The learning value came from recall, comparison, and revision. The AI interaction added feedback and alternative framing. It would have been weaker if the student had started with a polished essay and then merely recognized it as plausible. Retrieval practice is valuable precisely because producing an answer shows what is and is not understood. Karpicke and Roediger, Science

Habits that look helpful but are not

Some common AI habits feel efficient while reducing your ability to detect mistakes:

• Starting with a full answer. You lose the chance to find your own gap or form an independent view.
• Accepting citations without opening them. A citation can be irrelevant, outdated, nonexistent, or unable to support the claim attached to it.
• Asking several models and taking a vote. Consensus among systems is not independent evidence.
• Using a summary as a substitute for the underlying material. Summaries are good maps; they are poor substitutes when you must assess method, context, wording, or uncertainty.
• Letting the model choose the goal. It can optimize a stated objective, but it cannot decide which values, harms, or constraints matter to you.
• Handing it private or restricted material by default. Privacy, confidentiality, intellectual-property, and employer rules may limit what can be entered into a tool, even when the output is useful.

The remedy is not to ban AI. It is to make the independent step visible. Keep a scratchpad, a source list, a decision record, or a first attempt. These small artifacts preserve the ability to audit your own thinking.

A short habit to keep

Before using AI, decide which of these modes you are in:

• Production mode: I already understand the task and need speed. Let AI draft or format, then review the output.
• Learning mode: I need to build a skill or remember the material. Attempt, retrieve, and explain before seeing a completed answer.
• Decision mode: My choice affects money, safety, rights, reputation, or other people. Use AI to widen questions and surface assumptions, then verify evidence and retain human responsibility.

Switching deliberately between modes prevents a productivity shortcut from becoming a learning habit. It also avoids the false choice between rejecting AI and deferring to it. In production mode, offloading can be sensible. In learning and decision mode, the tool should create better questions, not remove the need to answer them.