The document is nearly finished.
A worker has asked an AI system for a first draft, then a tighter version, then a list of objections. The final text is competent. The deadline is close.
There is one remaining question.
How much thinking happened because the tool was there, and how much thinking disappeared because the tool was there?
That question is difficult to answer experimentally in real workplaces.
A 2025 CHI paper by Hao-Ping Lee and colleagues took a different route. Rather than trying to infer critical thinking from the final artifact alone, the researchers asked knowledge workers to describe real tasks in which they used generative AI and to report when critical thinking was enacted and how effort changed.
The paper has become an important reference in arguments about whether AI weakens human thinking.
It deserves to be read more carefully than that slogan allows.
The paper
The full title is The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.
It was published in the Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems.
The authors are Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks and Nicholas Wilson.
The study surveyed 319 knowledge workers recruited through Prolific. Participants provided 936 first-hand examples of tasks for which they used generative AI at work.
The researchers asked two broad questions:
- When and how do workers perceive themselves to enact critical thinking when using generative AI?
- When and why do they perceive AI as increasing or decreasing the effort involved in critical-thinking activities?
Why the study matters
A large part of the public debate about AI and cognition jumps directly from output quality to internal capacity.
If an AI-assisted document is less diverse, perhaps the user thought less.
If a task becomes faster, perhaps the user learned less.
If people delegate drafting, perhaps independent problem-solving is weakening.
All of those are plausible hypotheses.
None follows automatically from the artifact.
Lee and colleagues explicitly recognized this. They focused on enaction of critical thinking, meaning actions or manifestations that participants associated with critical thinking, rather than claiming to directly observe a pure mental faculty.
That conceptual modesty is one of the paper's strengths.
What participants described as critical thinking
Across the qualitative material, participants described critical thinking as work needed to ensure quality and intentionality.
The study organized these practices across three broad phases of AI-assisted work:
- goal and query formation;
- inspecting the response;
- integrating the response into the task.
People described verifying AI outputs against external sources or their own expertise, refining prompts, checking whether responses met criteria, assessing tone and relevance, and deciding how much of an AI output to use.
This matters because critical thinking did not simply disappear from the workflow.
Its location changed.
From information gathering to verification
For knowledge and comprehension tasks, participants often described generative AI as reducing the effort required to find and assemble information.
That did not mean effort went to zero.
Some of it moved into verification.
The system could produce an answer quickly, but the worker still had to decide whether the answer was accurate enough to trust.
This creates a new asymmetry.
Generation is fast.
Verification may still require domain knowledge, source access and time.
A worker can therefore receive more information while becoming more dependent on their ability to evaluate it.
From problem-solving to response integration
In application tasks, participants often described a shift from doing the full problem themselves to adapting or integrating AI output.
This can be productive.
A good external draft can release effort for higher-level judgment.
But integration is not cognitively free.
Someone must understand the task well enough to know whether the output fits the actual context.
A fluent response can make this requirement less visible because the material arrives in a form that already looks complete.
The paper therefore places response integration near the center of human responsibility in AI-assisted work.
From task execution to task stewardship
For analysis, synthesis and evaluation, Lee and colleagues describe a broader shift from execution toward task stewardship.
The human increasingly specifies intentions, steers the system, evaluates intermediate outputs, checks quality and remains accountable for the final work.
This is a useful framing because it avoids the misleading image of a machine simply replacing a task while leaving the human role unchanged.
The role can change in kind.
A person who once produced each sentence may now supervise a chain of generated options.
That can lower production effort while increasing the importance of judgment.
The confidence finding is important
The quantitative analyses found that confidence mattered in opposite directions.
Higher confidence in generative AI was associated with less perceived critical-thinking enactment and with lower perceived effort for several cognitive activities.
Higher self-confidence on the task was associated with more critical-thinking enactment and, in some cases, more perceived effort.
The authors interpret this pattern cautiously.
Workers who trust the AI more may rely on it more heavily, while workers with stronger task confidence may be more able or willing to inspect, steer and evaluate what the system produces.
The result is intuitively important.
It is also correlational.
Association is not causation
The paper does not show that confidence in AI causes critical thinking to decline.
Several alternatives remain possible.
People may be more confident in AI on routine tasks where little critical thinking was required in the first place.
Workers who know a domain well may notice more errors and therefore report more evaluative work.
People who habitually reflect may select different tasks for AI assistance.
Self-confidence may track expertise imperfectly.
The study used regression models to examine associations while accounting for measured task and user factors. That improves interpretation.
It does not turn a survey into a randomized causal experiment.
The dependent variables are self-reports
This is the paper's most important evidentiary boundary.
Participants reported whether they enacted critical thinking and how effort changed.
They were not given a standardized critical-thinking performance test before and after AI use.
The researchers did not measure long-term skill retention.
They did not observe whether unaided problem-solving ability deteriorated across months.
They did not randomly assign workers to different AI-use regimes and then test transfer.
The results therefore concern perceived enactment and perceived effort in reported work examples.
That is valuable evidence.
It is not evidence of measured cognitive atrophy.
The study itself acknowledges this problem
Lee and colleagues note that participants sometimes conflated lower effort in using AI with lower effort in critical thinking.
They also note that self-confidence may not be well calibrated to objective expertise.
They recommend future studies using other measures, including think-aloud methods and task-based assessments.
This is a good example of evidence-grade honesty inside the source itself.
The paper's limitations are not external objections pasted onto the study. They are part of the authors' own interpretation.
The sample is broad, but not universal
The study's 319 participants were professionally diverse knowledge workers who used generative AI at work at least weekly.
That gives the paper useful ecological breadth compared with narrow laboratory tasks.
But the sample does not represent all workers, all countries, all occupations or all levels of expertise.
Nor do 936 reported examples equal 936 independent participants.
The same person contributed multiple task examples, which the statistical models addressed using participant-level random intercepts.
Generalization should remain tied to the actual design.
Critical thinking was operationalized through a particular lens
Participants were introduced to critical thinking through examples aligned with Bloom's taxonomy and AI-assisted work.
This helped make an abstract term concrete.
It may also have shaped what participants noticed and reported.
That is not necessarily a flaw. Every measure operationalizes a construct somehow.
But it means the study should not be read as if it discovered the one universal form of critical thinking.
The paper observed a task-specific enactment framework.
The effort results are not simply "AI makes thinking easier"
Across several cognitive activities, many participants reported less effort with AI.
Evaluation showed a more mixed pattern than some lower-level activities.
The key interpretation is not that less effort is always bad.
Cognitive tools are supposed to reduce unnecessary effort.
A calculator that reduces arithmetic effort can improve a larger engineering task.
The concern is whether the effort being removed is productive cognitive work that the person still needs to retain, or whether it is routine burden that can be delegated safely.
The CHI study does not answer that longitudinal question.
It helps locate it.
Efficiency and capability are different outcomes
A worker can become more efficient without becoming more capable unaided.
They can also become more capable because the tool creates better feedback, examples or scaffolding.
Both pathways are possible.
The paper reports examples in which participants used AI to scaffold complex tasks and create feedback loops they did not otherwise have.
It also reports concerns about reliance and reduced independent engagement.
That tension is more faithful to the evidence than either "AI makes us smarter" or "AI makes us stupid."
The paper changes where responsibility sits
If production becomes cheaper, judgment becomes more valuable.
The worker who did not write every sentence is still responsible for whether the sentence belongs in the report.
The analyst who did not gather every fact is still responsible for whether the facts are reliable.
The manager who asked a model for options is still responsible for the decision.
The study's language of stewardship captures this institutional shift well.
AI can move human effort upward in the workflow.
It can also make it easier to skip the upward work.
Why this matters in Northern Europe
European education and workplace policy increasingly treats AI literacy as a practical competence involving critical evaluation, not merely tool familiarity.
The 2026 European Commission and OECD AI Literacy Framework for primary and secondary education explicitly includes engaging with, creating with, managing and shaping AI.
The CHI paper adds a workplace counterpart to that policy direction.
If AI use shifts cognitive labor toward verification and stewardship, then European institutions need to train those capacities intentionally rather than assuming they emerge automatically from access to the tools.
This is an Institute synthesis from the paper and the policy context, not a finding tested by Lee et al.
The 936 examples are qualitative data, not 936 independent people
The paper reports 936 real-world task examples from 319 knowledge workers.
That abundance of examples is a strength because it provides rich descriptions of how people say they use generative AI across work tasks. But the unit structure matters.
Three examples from one participant are not equivalent to three independently sampled participants. They share the same person's habits, occupation, confidence and interpretation of “critical thinking.” Analyses need to account for clustering rather than treating every example as independent evidence about the population.
The distinction also matters rhetorically. “936 examples” sounds larger than “319 workers,” but the study's human sample is 319 people. Responsible summaries should report both and keep their roles distinct.
Self-reported critical thinking is not directly observed reasoning performance
Participants were asked to describe whether and where they engaged in critical thinking and how much effort it required.
That is valuable evidence about experience and perceived work practice.
It is not the same as watching participants solve standardized reasoning problems with and without AI, scoring the quality of their reasoning, and then testing whether any difference persists after the tool is removed.
Self-report introduces familiar uncertainties.
People may differ in what they count as critical thinking. Some may recognize verification as cognitive work; others may regard it as routine checking. People may also underreport effort when a task feels fluent, or overreport it when they believe careful work is socially desirable.
This does not make the data meaningless. It defines the claim boundary.
The study is especially useful for describing perceived shifts in cognitive work and for generating hypotheses about where future performance-based research should look.
Confidence is directionally interesting and causally ambiguous
One of the paper's most discussed results concerns confidence.
Higher confidence in generative AI was associated with less reported critical-thinking effort and enactment. Higher self-confidence was associated with more.
Several causal stories are compatible with that pattern.
Trust in the system could reduce checking. People who habitually check less might become more confident in the system because they encounter fewer detected errors. Domain experts might distrust AI more precisely because they are better able to notice its limitations. Task type could influence both confidence and effort. Organizational norms could shape all of them.
The observed association cannot separate these explanations.
A stronger causal design would manipulate aspects of system reliability, confidence cues or verification requirements and measure resulting behavior. Longitudinal studies could examine whether trust changes with experience and whether calibration improves or deteriorates.
Until then, confidence should be treated as a useful predictor and design signal rather than a proven mechanism.
Different forms of AI assistance may redistribute cognitive work differently
“Using generative AI” is not one intervention.
A worker who asks for a first-draft outline faces a different cognitive task from someone who asks the system to decide between strategic options, summarize legal material or generate code.
Interface design matters too. A system that provides sources, uncertainty cues and editable reasoning traces creates a different environment from one that produces a single polished answer with no provenance.
Future studies therefore need to distinguish assistance modes.
Does AI replace search, drafting, memory retrieval, computation, evaluation or final decision-making? At what stage does the human enter? Is the person required to produce an initial judgment before seeing the AI output? Are errors visible enough to support learning?
The critical-thinking consequences are likely to depend on where cognition is redistributed, not simply on whether AI is present.
Productivity metrics can hide developmental cost
Workplaces often evaluate tools through immediate output: speed, volume, cost or perceived quality.
Those are legitimate outcomes, but they do not reveal what happens to human capability over time.
A tool can improve today's document while reducing tomorrow's independent practice. It can also free attention from routine production and create more opportunity for higher-level judgment. Both trajectories are plausible.
This distinction is especially important for junior workers and students, whose current tasks are partly training environments. An expert may safely offload a step because they already possess the underlying model needed to detect errors. A novice may offload the same step before that model has formed.
The CHI paper does not resolve this developmental question.
It makes the question harder to ignore.
Organizations evaluating AI should therefore separate task performance from capability formation. Short-term efficiency studies should not be treated as evidence that expertise is being preserved, and capability concerns should not be used to deny demonstrable efficiency gains.
The paper is a map of changed work, not a verdict on human decline
The most useful reading of the study is neither celebratory nor alarmist.
It documents a redistribution of self-reported cognitive effort around a powerful tool.
Some traditional work becomes easier. New verification and stewardship work appears. Confidence influences whether that work is undertaken. The result is a changing cognitive ecology.
That is enough to justify new research, new training and better interface design.
It is not enough to say that generative AI is making knowledge workers incapable of critical thought.
What the evidence does not show
The CHI 2025 study does not show that generative AI causes permanent decline in critical-thinking ability.
It does not show that workers who report less effort perform worse.
It does not measure long-term transfer to unaided work.
It does not establish that high confidence in AI causes low critical thinking.
It does not show that all uses of AI reduce engagement.
It does not show that task stewardship is always cognitively superior to direct execution.
And it does not settle which parts of knowledge work humans should continue to practice unaided.
What this changes for Inner Technology
The study sharpens a central Inner Technology question.
The issue is not whether a tool performs part of a task.
The issue is which human capacity moves upstream when the tool does.
If gathering moves to verification, verification must be trainable.
If drafting moves to integration, integration must be trainable.
If execution moves to stewardship, stewardship must be trainable.
If confidence in the tool changes how much scrutiny is applied, calibration becomes part of the workflow.
This does not mean preserving effort for its own sake.
It means asking which effort carries the human capacity we still need.
The finished document
The worker looks at the nearly finished draft.
The AI has done a great deal.
The important question is no longer, "Did I write this myself?"
It is, "What am I still responsible for knowing, checking and deciding?"
The CHI 2025 paper does not answer that for every profession.
It gives us a much better map of where to look.
Research references
- Lee HP, Sarkar A, Tankelevitch L, Drosos I, Rintel S, Banks R, Wilson N. 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 CHI 2025. Article 1121, 1-22. https://doi.org/10.1145/3706598.3713778
- Microsoft Research. Publication record for Lee et al. 2025. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
- Buçinca Z, Malaya MB, Gajos KZ. 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. 2021;5(CSCW1). https://doi.org/10.1145/3449287
- Fischer H, Huff M. Metacognition as a target of science communication. Nature Reviews Psychology. 2025;4:554-555. https://doi.org/10.1038/s44159-025-00478-2
- OECD, European Commission. Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education. OECD Publishing. 2026. https://doi.org/10.1787/65cd27d4-en
- European Commission. New AI Literacy Framework helps schools prepare learners for the age of artificial intelligence. 18 June 2026. https://education.ec.europa.eu/whats-new/news/new-ai-literacy-framework-helps-schools-prepare-learners-for-the-age-of-artificial-intelligence