A policy analyst asks a generative AI tool to summarize a consultation. The answer looks coherent, yet one recommendation depends on a source the analyst has not opened. Did using the tool reduce critical thinking, or did it move critical thinking from drafting toward checking and integrating? The difference matters. A fluent answer can conceal an unsupported claim; an equally fluent answer can help a worker spend more time on judgment. The 2025 CHI paper by Hao-Ping Lee and colleagues studies how knowledge workers perceive such changes in their own tasks. It cannot by itself decide which path a particular organization has taken.
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers is a source-first research brief. It reports the paper’s question, survey design, principal findings and limitations before considering what its findings might mean for practice. It differs from the Learning Hub’s broader account of designing AI to preserve human capacity: this page stays close to one study and treats the design proposals as implications to test, not as results the survey established.
The question the paper actually asked
The authors asked when and how workers think critically while using generative AI, and when they experience more or less effort in doing so. These are questions about reported activity during actual work examples. They are not direct tests of whether a person’s underlying reasoning skill improved or declined. The title’s phrase “self-reported reductions” is essential to an accurate reading.
The paper responds to a change in ordinary knowledge work. A person may ask a model to draft, search, summarize, revise or advise. A conventional productivity measure can count time saved or output delivered, while leaving open whether the worker checked evidence, understood the result or could explain it later. Lee and colleagues wanted a more precise view of what people say they do cognitively within these interactions.
They used a broad operationalization of critical thinking, informed by Bloom’s taxonomy of cognitive activities. Participants were introduced to examples ranging from checking an email’s tone to verifying code or examining bias in an analysis. That onboarding helped make the survey question intelligible, but it may also have influenced what respondents recognized and reported as critical thinking. The authors acknowledge that critical thinking is not a single, universally agreed observable unit.
A second distinction is between doing a task with less effort and thinking less critically. An assistant can reduce the effort needed to retrieve information while leaving evaluation intact. A worker may instead accept an answer with minimal checking. The survey asks about perceived effort and reported behavior, but cannot always separate those pathways. The paper itself names this ambiguity as a limitation.
Who was studied and how
The researchers recruited 319 knowledge workers through the Prolific online platform. Participants used generative AI for work at least weekly. They described 936 real examples of AI-assisted tasks and indicated how critical thinking entered each one. These examples are nested within workers: 936 examples do not mean 936 independent people. A reader should therefore resist treating the count as a random sample of all jobs, industries or countries.
Participants first described what they were trying to accomplish, which tool they used and how they used it. They selected task types within creation, information and advice categories. They then rated task-specific confidence in their own ability to do the work without AI, confidence in the AI’s ability to do it and confidence in their ability to evaluate its output. The confidence questions used five-point response scales. The survey also asked whether participants thought they had engaged in critical thinking and how much effort cognitive activities seemed to require with AI compared with without it.
The authors analyzed these answers with explanatory regression models that included both task factors and user factors. They also coded free-text descriptions of reported practices and reasons. Combining the two methods can reveal patterns that neither a single percentage nor a small set of quotations would show alone. It does not create an experiment. Participants were not randomly assigned to high or low confidence, nor did the researchers observe the same worker’s performance over years with and without AI.
Recruitment matters to interpretation. The sample consists of people already using AI at least weekly, and the authors report that it skewed younger and more technologically skilled. The survey was conducted in English with fluent English-speaking respondents. Its patterns may differ among workers with little access to AI, older workers, multilingual teams or occupations where checking a result has an unusual cost. The authors also studied a particular period in rapidly changing tool use; a new interface or organizational policy could alter the practices they observed.
The analytic unit also matters. A worker may be cautious in a high-stakes report and casual in a routine invitation. The authors measured confidence about specific tasks because a global statement such as “I trust AI” may miss this variation. In their models, overall trust did not show the same relationship to reported critical-thinking enactment as task-specific confidence in AI. This is a useful reason to ask “trust it for what?” rather than assume one stable attitude governs every use.
What workers said they did
In 555 of the 936 examples, about 60 percent, respondents reported some critical-thinking activity. This is an observed proportion of survey reports, not a measure of how often all AI users think critically. It also depends on how the survey defined and elicited examples. The result cuts against two easy stories: that no one checks an AI answer and that everyone who uses AI reliably does so.
The qualitative accounts describe several forms of activity. Workers checked generated information against external sources or their own domain knowledge. They selected relevant pieces from an answer, changed its language for an audience and decided how it would fit a larger task. The authors group important shifts as movement from information gathering toward verification, from problem solving toward response integration and from task execution toward stewardship of the output and process.
Those shifts are descriptions of reported practice, not a ranking of superior and inferior work. Verification can be demanding, especially if citations are missing or the worker lacks background knowledge. Integrating a response can involve careful comparison with a brief, but can also become cosmetic editing. Stewardship can mean guiding, checking and taking responsibility. It can also become a label for nominal human oversight if the workplace leaves no time to challenge the system. The survey invites these questions without resolving them for every example.
Participants gave reasons for checking. Work quality and avoiding negative consequences mattered; some saw critical engagement as part of developing their professional skill. Others named barriers. The paper reports lack of time, uncertainty about whether verification belonged to their role, insufficient domain knowledge to inspect an answer and difficulty improving an output after noticing a problem. These are not merely individual shortcomings. A workflow can assign one person to generate and another to verify, or can reward speed so strongly that checking becomes unrealistic.
The survey’s free-text accounts are valuable because they show concrete ways a general term is used at work. They are still selected recollections. A respondent may remember an exceptional case, present their actions favorably or misidentify whether a quick scan caught an error. The paper did not audit all source documents, prompts and final artifacts to determine whether each claimed check was successful. “I verified” and “the conclusion was correct” are different findings.
Confidence and reported critical thinking
The most consequential quantitative pattern concerns confidence about a particular task. Greater confidence that AI could do the task was associated with a lower likelihood of reporting critical-thinking enactment. Greater confidence in one’s own ability to do the task without AI and in one’s ability to evaluate the output was associated with more reported critical thinking. These are associations in the fitted survey models. They do not prove that confidence caused the behavior.
A plausible interpretation is that people who feel able to do and judge a task remain engaged when they delegate parts of it. Another is that people who choose to scrutinize an output retrospectively report greater confidence in themselves. Task difficulty, professional experience, organizational norms and the quality of available AI outputs may affect both confidence and behavior. The study’s controls reduce some simple confounding, but a cross-sectional survey cannot remove all such alternatives.
The distinction between self-confidence and calibrated expertise matters. Feeling capable does not guarantee that a worker can detect errors. A novice may be confidently wrong, while a careful expert may express uncertainty. The authors explicitly identify this as a measurement limit: their confidence scores are subjective. An organization using the paper to justify an intervention would need to measure task performance and error detection separately from confidence.
Nor should the finding be turned into a command to distrust AI indiscriminately. Appropriate reliance depends on actual model performance, the task, error costs and available alternatives. Bansal and colleagues showed in human-AI teaming work that the most accurate standalone AI need not maximize team performance in a studied decision workflow. That related experiment reinforces the need to evaluate the person-plus-tool arrangement, but its task and outcome differ from Lee and colleagues’ survey. Neither paper provides a universal trust setting for every generative system.
What “less effort” means here
For a majority of examples across the six cognitive categories used in the study, participants said AI reduced their effort compared with doing the task without it. The proportions varied: roughly seven in ten reported less effort for knowledge, comprehension, application, analysis and synthesis; evaluation had a smaller majority. Those responses describe a comparison imagined or recalled by participants, not a measured brain workload or a test of intellectual decline.
Confidence in AI’s ability to do the task correlated with lower perceived effort for five of the six cognitive activities in the models. Confidence in one’s own task ability correlated with greater perceived effort in applying and evaluating outputs. It is tempting to say that AI confidence “switches off” thinking. The study supports a narrower statement: among these survey responses, confidence and reported effort moved together in specific ways after adjustment for modeled factors.
Four possibilities can produce a lower effort report. The worker may do the same kind of evaluation with useful support. They may delegate part of it to the tool. They may perform less evaluation. They may interpret the ease of completing the task as less critical-thinking effort even when their checking remains substantial. The paper explicitly considers these possibilities. A survey answer alone cannot decide which one happened in a given case.
Reduced effort may be a benefit. Searching a large document manually is not intrinsically valuable if a reliable tool can locate the relevant passage. What matters is whether the worker can assess the passage and use it appropriately. Conversely, if a generated summary removes the only occasion on which a novice would have read conflicting evidence, time saved may come with a learning cost. Lee and colleagues did not follow participants long enough to measure that trajectory.
Tankelevitch and colleagues’ earlier CHI paper frames generative AI use as a metacognitive problem: the user has to decide what to delegate, monitor the result and adjust strategy. That conceptual work helps interpret why the same aid can simplify one operation while creating a new supervisory burden. It is not a replication of the 2025 survey. Its value here is to identify a competing account of effort: work may shift rather than disappear.
What the findings change for research and design
The paper supplies a vocabulary for studying AI-assisted knowledge work at the task level. A future study could record what workers checked, whether they identified planted errors, how they integrated an answer and whether they retained the ability to do the task later. It could compare those outcomes with self-reported effort. Such work would test whether the survey’s associations correspond to observable performance.
For a team deploying AI, the immediate lesson is to specify the human role. If the user is expected to verify a claim, the interface should make the original material accessible and the organization should allow time to read it. If a worker is expected to explain a decision, the workflow should preserve the rationale and unresolved uncertainty. These are design hypotheses derived from the paper’s findings, not interventions the survey tested.
One option is to introduce a moment of active consideration before accepting an AI recommendation. Buçinca and colleagues tested cognitive forcing functions in particular AI-assisted decisions and found they could reduce overreliance, while also imposing costs. That study gives a reason to experiment with such prompts; it does not justify placing friction everywhere. A routine formatting suggestion and a consequential policy inference deserve different thresholds.
Amershi and colleagues’ human-AI interaction guidelines offer prompts about setting expectations, communicating uncertainty, supporting correction and allowing control. They help translate the 2025 survey into questions for a product review: Can users tell what the system has and has not done? Can they inspect evidence? Can they correct an output and recover when it fails? The guidelines are a design resource, not proof that applying them preserves critical thinking.
An evaluation should measure both immediate productivity and the quality of human judgment. Compare users’ final answers with source material, record missed and detected errors, and ask whether a worker can explain a choice without repeating a generated phrase. Repeat the assessment after weeks or months. Include workers with different experience and accessibility needs, and examine whether checking is feasible under their actual workload. None of these tests can be replaced by a single confidence rating.
Why this matters in Northern Europe
Organizations in Sweden and the wider Nordic region may use generative AI in public administration, education and professional services, where a fluent answer can influence someone who did not see the prompt. The CHI sample does not establish a Swedish or Nordic effect. The practical question is how to test these workflows locally, across the languages and institutional duties that the English-only study did not represent.
A local evaluation could compare Swedish-language and English-language tasks, examine how workers verify sources in each, and record whether responsibility for correction is clear. In a public-service setting, a person affected by a decision may need more than a concise summary; they may need traceable reasons and an avenue to challenge an error. Those are governance and design considerations applied to a regional context. They are not findings that Lee and colleagues measured in Nordic participants.
What the evidence does not show
This study does not show that generative AI causes a general decline in critical-thinking ability. It did not randomize workers to tool use, measure a before-and-after skill trajectory or compare their independent performance years later. It does not show that the reported 60 percent of examples with critical thinking reflects the prevalence of effective checking in all knowledge work. The sample was a specific group of frequent users recruited online, answering in English.
Self-reported effort is not a direct measure of mental effort, reasoning quality or task accuracy. Task-specific confidence is not the same as calibrated skill. The regression coefficients identify modeled associations, not mechanisms. Free-text accounts explain how participants understood their experience, not whether every final product was sound. The authors’ concern about possible long-term overreliance is a reason to investigate that outcome, not evidence that it occurred in their cross-sectional data.
The study also does not establish that more effort is always better. Skilled use of a reliable aid can reduce unnecessary work. Equally, demanding verification may be essential when an answer affects others. The relevant outcome is whether people can understand, check and defend a result under the conditions of the task. That outcome requires direct observation and context-specific tests beyond the present survey.
Return to the consultation summary
The analyst in the opening example can now ask a narrower question: which source supports the recommendation, and what did the tool omit? If the answer survives that check, the AI may have saved useful time. If it does not, the worker needs a path to correct it and explain the correction. The 2025 paper makes this work visible as verification, integration and stewardship. It does not promise that those practices happen automatically when a system says “human in the loop.”
The most defensible use of this research is to make an organization’s claims about AI assistance testable. Name the task, the worker’s role, the evidence available for checking and the outcome that matters. Then observe what people do, including when the model is wrong. That approach respects both the promise of faster work and the study’s limits. It keeps critical thinking a practice to examine, rather than a headline to fear or celebrate.
Research references
- Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121. View source
- Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2024). The metacognitive demands and opportunities of generative AI. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, Article 680. View source
- Bansal, G., Nushi, B., Kamar, E., Horvitz, E., & Weld, D. S. (2021). Is the most accurate AI the best teammate? Optimizing AI for teamwork. Proceedings of the AAAI Conference on Artificial Intelligence, 35(13), 11405–11414. View source
- Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. View source
- Amershi, S., et al. (2019). Guidelines for human-AI interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, Paper 3. View source