The Evaluating Self

83The Impact of Generative AI on Critical Thinking

Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks and Nicholas Wilson, CHI '25, 2025

Linked to 17 chapters: supports 9 · contradicts 3 · mechanism 10 · same claim 2 · pairing 2

Critical thinking with generative AI, 2025 — the survey in five moves 319 workers936 examples How it isenacted Confidencedecides Where theeffort goes Stewardship Prolific, weekly AIusers at work. Threeexamples each, codedfor enaction. Goals and queries,inspecting theresponse, integratingit into the work. Confidence in thetool: less thinking.Confidence in self:more thinking. Less effort at all sixlevels, least atevaluation. Trusttracks the drop. Effort moves toverifying, integratingand overseeing thework of the tool. Task type, gender, age and occupation predicted nothing; confidence in the tool did. It measures what workers report — enaction — not critical thinking itself.
Figure 83.0 — The paper's five moves. A survey of real tasks, the practices workers call critical thinking, the one variable that predicts them, the effort they cost, and where that effort has gone.

Lee and his co-authors begin with a list of complaints: Socrates against writing, Trithemius against printing, arithmetic teachers against calculators, and a long line of objections to the Internet. They do not dismiss the list. They take it seriously by way of Bainbridge, whose irony of automation is that mechanising the routine parts of a job removes the routine occasions on which a worker's judgement is exercised, leaving it unpractised when the exceptions arrive. The question is whether generative AI does that to critical thinking, and they ask it of people at work rather than students in a classroom. They surveyed 319 knowledge workers recruited on Prolific, all of whom used a generative tool at work at least weekly, and collected 936 first-hand examples of real tasks. Of each example they asked two things: whether the worker had enacted critical thinking, and whether each of the six cognitive activities in Bloom's taxonomy had cost more or less effort with the tool than without it. Workers reported critical thinking in 555 of the 936 examples, and reported less effort at all six levels. Seven mixed-effects regressions then found that almost nothing about the task or the worker predicted enaction — not task type, not gender, not age, not whether the occupation was exposed to automation. Confidence did. Confidence in the tool went with less critical thinking; confidence in oneself went with more. The qualitative half describes what the remaining thinking consists of: verification, integration, and what the authors call stewardship.

Critical thinking, operationalised

The definition comes from Bloom's 1956 taxonomy of educational objectives, whose six levels are knowledge (recall), comprehension (organising and restating), application (solving problems in new situations), analysis (breaking a problem into parts), synthesis (assembling parts into a whole), and evaluation (judging against criteria). The authors do not claim it is the best available account. They name the Paul–Elder scheme, with its eight elements of thought, ten intellectual standards and eight intellectual virtues, as the more elaborate rival, and choose Bloom because it is old, has survived heavy criticism, and has six dimensions with definitions plain enough to put into a questionnaire. Each level became one five-point item.

The dependent variable is not critical thinking but its enaction — the actions that signal or manifest it — on the argument that thinking as a mental event is hard to observe in oneself and report. That choice is aimed at the measure the prior literature had been using. Sarkar had described the tendency of AI-assisted work to produce mechanised convergence: workers with a generative tool produce a less varied set of outputs for the same task than workers without one. Output diversity is a proxy for critical thinking and, these authors argue, a bad one. A worker who reuses an output unedited may have judged, carefully, that it needed no editing, and that judgement is invisible to any measure of the finished artefact.

They also separate their subject from overreliance, the acceptance of an incorrect recommendation. Buçinca had shown that cognitive forcing functions — making the user wait before seeing the output, or requiring them to revise it — cut overreliance further than explanations do. But a worker who accepts a correct output that is merely adequate has not overrelied, and has still not thought. The failure being chased here is wider than the one that surfaces as an error.

Recruitment ran on Prolific among people who used a generative tool at work at least once a week. Of 333 responses, 14 were dropped for low-effort free text, leaving 319. Participants were first shown examples of critical thinking at several levels of the taxonomy — checking the tone of a generated email, verifying a code snippet, looking for bias in a data insight — so that they would not read the term too narrowly. Each was then asked for three examples of their own, one from each of Brachman's categories of tool use: creation, information, and advice. Of 957 examples, 21 were removed as vague or duplicated, leaving 936; ChatGPT appeared in 97 per cent of responses. Seven mixed-effects models were fitted with participant as a random intercept, one logistic for enaction and six linear for effort, the 98 predictors corrected by the Benjamini–Hochberg procedure to a threshold of p = 0.007.

How knowledge workers enact critical thinking

The free-text answers were open-coded and sorted onto three stages of a working session with the tool: forming the goal and the query, inspecting the response, and integrating it into the work. What the workers describe as critical thinking is almost entirely inspection.

Six of the 319 described forming a goal before opening the tool. Thirty described forming and revising the query, and the examples suggest that the revision is where the thinking happens: a teacher generating an image about hand-washing for a school presentation noticed that her picture had no soap dispensers, changed the prompt to name them, and found that working out what the image had to show gave her a better image.

Inspection carries the largest counts. A hundred and twenty-five workers evaluated the output against objective criteria — whether it answered the prompt, whether the generated code compiled and ran. Seventy-seven applied subjective standards: whether a suggestion for a social care setting was affordable in a real service, whether a set of trading ideas followed logically, whether the output suited the audience. A hundred and fourteen cross-referenced the output against outside sources, from official documentation to a web search to another person; a nurse checked a generated pamphlet for newly diagnosed diabetic patients against her own hospital's guidelines, and a conductor found the repertoire she was offered American-centric and unfit for her country. Twenty-three checked the sources the output itself had cited, the narrower and rarer habit.

Integration splits in two. Thirty-six workers described choosing which part of a long output to use — an auditor writing a résumé had to decide which of the regulations the tool named applied in the country she worked in. Forty-five described changing the form rather than the content: a researcher rewriting his abstract because the tool had produced something too emphatic for a scientific paper, and a worker making an email read as though a colleague had written it.

What workers describe as critical thinking, by stage of the session Goal and query formation Inspect response Integrate response Form the goal Form the query Objective criteria Cross-check sources Subjective standards Check cited sources Use part of the output Adjust the style 6 30 125 114 77 23 36 45 Counts are participants out of 319 whose free text described the practice. Most of the reported thinking is inspection, against criteria and against outside sources.
Figure 83.1 — Where the reported thinking goes. Counts are participants out of 319; inspection dominates both goal formation and integration.

When they enact it, and when they do not

Critical thinking was reported in 555 of 936 examples, a little under sixty per cent. The logistic model found no effect of task type: creation, information and advice were indistinguishable. Gender, age group and whether the occupation was economically exposed to automation all did nothing. The predictors that survived correction were about confidence. Confidence in the tool's ability to do this particular task predicted enaction negatively and strongly (β = −0.69, p < 0.001). Confidence in doing the task oneself predicted it positively (β = 0.26), as did confidence in evaluating the tool's output (β = 0.31). A worker's general tendency to reflect, measured with Kember's reflective thinking inventory, predicted it positively too (β = 0.52).

General trust in generative AI, measured as a disposition, predicted nothing at all (β = −0.01). The authors read this as task-level confidence having already absorbed the variance a general disposition would explain: whether a worker checks is decided task by task, not once and for all.

What predicts whether a worker thinks critically about the output Confidence in the AI Tendency to reflect Confidence in evaluating Confidence in oneself Task type: information Overall trust in the tool -0.69 *** 0.52 *** 0.31 * 0.26 * 0.32 n.s. -0.01 n.s. -0.5 0 +0.5 Coefficient, log odds of enacting critical thinking From Table 4; stars mark significance after correction, threshold p = 0.007.
Figure 83.2 — What predicts the enaction of critical thinking. Confidence in the tool is the largest effect in the model, and it runs the wrong way.

The free text supplies the reasons on both sides. Critical thinking was motivated by the quality of the artefact (74 participants, typically because the output was generic and had to be made specific), by the consequences of getting it wrong (116; a pharmacist double-checked a continuing-development entry because a bad one risked suspension), and by skill development (13, the smallest of the three motivators — a programmer making sure he understood the fix well enough to do it himself next time).

The inhibitors are sorted into three kinds. Awareness: the tool's use was secondary to the goal (14), the task was trivial (55), or the worker trusted and relied on the tool (83). One participant, asked why he had not thought about making a passage sound professional, said "I just never thought about it, as critical thinking didn't feel relevant." Motivation: no time (44 — a sales representative with a daily quota to hit), or verification belonging to another team (11). Ability: not enough domain knowledge to inspect the response (58), and not enough command of the tool to improve it (72), including tools that repeated a wrong suggestion after being asked for a different one.

The authors then point the awareness barrier back at Bainbridge. Skipping the checking on a trivial task is individually harmless, and the paper says so. But skill decays without practice, and if the high-stakes tasks become the only occasions on which judgement is exercised, the worker arrives at them out of practice. Simkute and colleagues had restated Bainbridge's irony for generative tools; this survey supplies the population of low-stakes tasks on which the practice is being given up.

Where the practice is given up Task looksroutine Confidence inthe tool Checkingskipped No practiceat judging Low-stakes work:nothing goes wrong High-stakes work:judgement out of practice 55 participants called the task trivial; 83 named their trust and reliance on the tool. Bainbridge's irony restated: practice is lost on exactly the cases that seem not to need it.
Figure 83.3 — The irony, in the survey's own numbers. The cases that seem not to need judgement are the cases that supply the practice.

Where the effort goes

The six effort scales run from much less effort (−2) to much more (+2). Most workers reported less effort at every level: 72 per cent for knowledge, 79 for comprehension, 69 for application, 72 for analysis, 76 for synthesis, and 55 for evaluation. Evaluation is the exception in both directions. It shows the smallest majority reporting less effort and by far the largest minority reporting more — 120 of its 764 examples, against thirty to forty-three for each of the others.

Effort of each cognitive activity, with the tool against without it less effort Knowledge Comprehension Application Analysis Synthesis Evaluation 72% 79% 69% 72% 76% 55% less effort with the tool about the same more effort From Table 4. N varies by activity: participants could mark one as not relevant. Evaluation is the exception: 120 of 764 examples were reported as more effort.
Figure 83.4 — Reported effort by cognitive activity. Judging the output is the one activity that a substantial minority found harder with the tool than without it.

The same two confidences that predicted enaction predict effort, and they predict it in opposite directions. Confidence in the tool went with less perceived effort at five of the six levels, all but application: knowledge (β = −0.11), comprehension (−0.13), analysis (−0.15), synthesis (−0.12) and evaluation (−0.23). General trust in generative AI, which had done nothing for enaction, does work here: knowledge (−0.12), application (−0.17), analysis (−0.12), evaluation (−0.24). Confidence in oneself ran the other way, and only where the worker has to act on the output: application (0.08) and evaluation (0.10).

The paper states the pair as its result. Higher confidence in the tool goes with less critical thinking, and it is experienced as less effort. Higher confidence in oneself goes with more critical thinking, and it is experienced as more effort. No causal claim is made, and the authors say plainly that self-confidence is subjective and need not be calibrated to real expertise. They are equally plain that a reported drop in effort admits four readings: the same thinking done with support, the thinking offloaded, less thinking, or the general effort of a task confused with the effort of thinking about it. The last of these, they concede, happened most when participants were satisfied with what the tool gave them.

From execution to stewardship

The free text on effort produces the paper's most-quoted finding, which is not a number but a description of three displacements.

For knowledge and comprehension, the effort moves from gathering information to verifying it. A hundred and eleven participants described getting information faster and at greater scale, and 87 described it arriving already organised. Fifty-six described the opposite: a lawyer looking for relevant law found that the tool invented material to agree with the case he was making, so that checking it by hand cost him the time he had saved.

For application, the effort moves from solving the problem to fitting the answer in. Seventy-seven described getting a solution tailored to their own situation, and nine described feeding the tool their own examples to work from. Nineteen described the work of making the output usable, which is sometimes a net loss: a developer found that matching generated code to his existing code and altering it to fit "was more effort than just doing it myself in the end."

For analysis, synthesis and evaluation, the effort moves from doing the task to overseeing it. On the reducing side, 48 described the tool scaffolding a complicated problem, 129 described it automating the making of a document or a piece of code, and 40 described a feedback loop they had previously had only by asking a colleague. On the increasing side, 45 described turning an intention into a query, 48 steering a response that drifts, and 42 judging the result against their own standards.

Three displacements of critical thinking effort gathering information verifying information solving the problem fitting the answer in doing the task overseeing the task Knowledge,comprehension Application Analysis, synthesis,evaluation 111 workers described faster retrieval; 56 described having to verify what came back. 129 described creation automated; 48 described having to steer a response that drifts. Stewardship, not teammate: the material work is delegated, the accountability is not.
Figure 83.5 — The three displacements. Nothing on the right is new work in kind; what has changed is which of it the worker is doing.

Stewardship is the word the authors settle on, in preference to teammate, collaborator and supervisor. The material production has been delegated; the responsibility and the accountability have not. That asymmetry is what makes the human-to-human words wrong, and it is also what makes verification a permanent cost rather than a transitional one.

The design proposals follow the three barriers: proactive interruption where checking is likely to be skipped, rather than assistance only on request; checking positioned as skill development rather than as an auditing task attached to each job; and explanations, suggested points of refinement and guided critique for workers who cannot inspect an answer unaided. Behind all three sits recalibration — giving a worker something to go on in deciding when the tool is to be trusted.

Since 2025

The paper was reported widely and usually wrongly, as a finding that AI makes people less intelligent. It is a cross-sectional survey of self-reports, it establishes no causation, and it measures whether critical thinking was enacted rather than how well. The authors list the limits themselves: an English-only instrument, a Prolific sample skewed young and technically able, subjective confidence that may not track real expertise, and the conflation of general effort with the effort of thinking.

What has held up best is the descriptive half. Verification, integration and stewardship have become the standard account of what is left for the human in AI-assisted work, and the finding that inspection is where the reported thinking concentrates has been reproduced in studies of programming, writing and analysis. What remains contested is the measurement. Kosmyna and colleagues put 54 people through essay writing under EEG in 2025 and found weaker neural coupling in the group using a language model, and that those participants could not quote from essays written minutes earlier. Gerlich, surveying 666 people the same year, found AI use correlated with lower critical thinking scores, and inherited the same cross-sectional weakness. None of the three can say which way the arrow runs.

The confidence result is the part that connects outward. Lee and See had argued that the design target is appropriate trust rather than more trust, and that failing to aim for calibration produces misuse and disuse alike; this survey measures the misuse side at the granularity of a single work task, and finds that it is task confidence, not disposition, that does the work. Risko and Gilbert had already made the point that offloading is neither good nor bad in itself, the question being whether the offloaded capacity will be wanted later; the low-stakes tasks catalogued here are the cases where that question is never asked. And METR's randomised trial of experienced open-source developers, run a few months later, found that the tools slowed them down while they believed they had been sped up — the gap between perceived and actual effort that a survey, by construction, sees only one side of.

The design recommendations have aged less well. Every intervention the paper proposes costs the worker time, and 44 of its own participants said they had none to spare. Buçinca had already found the same trade: the forcing functions that reduced overreliance were the ones participants liked least and, given the choice, turned off.

Claims drawn from this chapter · 9

evidenceHow knowledge workers enact critical thinking

Almost all the thinking workers reported was inspecting what the tool produced

Of 319 knowledge workers asked what critical thinking they had done with a generative tool, 6 described working out the goal before opening it and 30 shaping the query, against 125 who judged the output and 114 who checked it against an outside source.

Becausethe counts come from open coding of free text over three stages of a session, and the imbalance is too large to be a coding artefact.
LimitsInspection is the easiest thinking to notice and describe, so goal formation may be happening unreported. The examples participants were shown were themselves mostly inspection.
PracticeSpecification discipline — write what the output has to contain before asking for it. The teacher who noticed her generated hand-washing image had no soap dispensers got a better image because she worked out what it had to show.
After AIThe framework's first layer is specification; this measures how little of it there is. People kept the part that feels like work — looking at a finished thing — and gave up deciding what it should be.
opinionCritical thinking, operationalised

An unedited output can be the result of a careful judgement

Lee's authors argue that how varied a group's finished work is cannot measure how much they thought: a worker who uses an output unchanged may have decided carefully that it needed no change, and the artefact cannot tell them from one who never looked.

Becauseit is an argument about what a measure can see, offered against Sarkar's mechanised-convergence result, not a finding of its own.
LimitsThe argument cuts both ways: if judgement leaves no trace in the artefact, the survey's own measure is a worker's report of having judged, which is no more inspectable. Convergence results still show a real narrowing of output whatever its cause.
PracticeFailure-mode review — stop reading sameness or editing volume as evidence of thought in your team's work. Ask what was checked and against what, which is a question about the process and not about the document.
After AIReview in most organisations happens on the artefact, because that is what gets submitted. This says the artefact is where the evidence is not — the reviewable unit is the conversation with the tool, not the thing it produced.
evidenceWhen they enact it, and when they do not

Confidence in the tool was the strongest predictor of not checking

Across 936 tasks reported by 319 knowledge workers, confidence that the tool could do that task was the largest predictor of not checking its output (β = −0.69). Confidence in doing the task oneself (0.26) and in judging the output (0.31) ran the other way.

Becausea logistic mixed-effects model over 936 first-hand task examples, with 98 predictors corrected to p = 0.007, so the surviving coefficients are not a fishing result.
LimitsA cross-sectional survey of self-reports establishes no causation, and the authors say self-confidence is subjective and need not track real expertise. The sample was English-only and skewed young and technically able.
PracticeReliance audit — list the tasks where confidence in the tool exceeds confidence in doing the job yourself, and arrange a check from outside for those; your own sense of when to look is the variable running the wrong way.
After AIReported as a finding that AI makes people less intelligent, it says something narrower: what decides whether you check is a feeling about the tool, available to be examined before the task rather than after.
evidenceWhere the effort goes

Judging the output was the one activity that got harder with AI

Workers reported less effort at all six of Bloom's levels, but evaluation was the exception both ways: the smallest majority reporting less (55 per cent against 69 to 79) and the largest minority reporting more, 120 of its 764 examples against thirty to forty-three elsewhere.

Becausesix five-point scales answered about 936 particular tasks, with evaluation separating from the other five levels on the same instrument and the same participants.
LimitsEffort here is perceived effort, reported once, and the authors concede that participants often confused the general effort of a task with the effort of thinking about it — most of all when they were satisfied with what the tool gave them.
PracticeCalibration log — budget evaluation as the expensive part of a task, and write down what you expected to be true before reading the output, so judging it has something to judge against.
After AIThe framework's second rule says the evaluation layer gets harder as capability rises. This is a survey number for it: the one place the tools save no effort is where the output has to be judged, and that is the layer that does not go away.
evidenceWhen they enact it, and when they do not

Nothing about the task or the worker predicted whether they thought critically

In Lee's survey of 319 workers, whether someone checked a tool's output was not predicted by task type, gender, age, whether the occupation was exposed to automation, or general trust in AI (β = −0.01). Only confidence at that one task predicted it.

Becausethe nulls come from the same corrected model that produced the large confidence effect — failures to find an effect in a design that found one.
LimitsA null on 319 self-selected people is weak evidence of absence, and the general-trust scale may simply be a poor instrument. The authors' reading — task confidence absorbing the variance — is an interpretation, not a result.
PracticeDomain audit — stop settling your relationship with the tool once, as a policy or a personality; decide per task, and give the domains where you cannot judge your own competence a fixed rule rather than a feeling.
After AIMost advice about AI is written as a stance to adopt — trust it, distrust it, verify everything. This is the measurement saying stances do no work, and the unit of decision is one task.
evidenceWhen they enact it, and when they do not

Skipping the check was mostly unawareness and inability, rarely lack of time

Asked why they had not thought about an output, 319 workers gave 152 reasons of awareness, 130 of ability, and 55 of motivation. Only 44 said they had no time; 83 said they trusted the tool and 58 lacked the knowledge to inspect it.

Becausethe three kinds come from open coding of first-hand reasons attached to particular tasks, and the ordering survives disagreement about individual codes.
LimitsReasons are reconstructed after the fact, and "it never occurred to me" is what people say when they cannot recall deciding. The largest single code, trusting the tool, could be awareness or a reasonable judgement.
PracticeAutomation ironies audit — sort your own skipped checks into the three kinds before reaching for a fix: not noticing needs a trigger, not being able needs a skill or a second reader, only the third needs time.
After AIAlmost every proposal for keeping humans in the loop assumes the human has the knowledge to look and only needs prompting. On these numbers, as many could not have checked as did not think to.
evidenceFrom execution to stewardship

The effort moved from producing the work to verifying and steering it

Lee's workers described three displacements, not a saving: gathering information became verifying it (111 faster, 56 checking), solving a problem became fitting the answer in (77 tailored, 19 making it usable), and doing a task became overseeing one (129 automating, 48 steering a drifting response).

Becausethe three patterns come from coded free text on 936 tasks with counts on both the reducing and increasing side, reported by workers describing their own week rather than inferred.
LimitsSelf-report cannot weigh the two sides: 111 saying information came faster and 56 saying it had to be checked is not a net figure. One developer found matching generated code to his own cost more than writing it himself.
PracticeFailure-mode review — plan the new work rather than assuming it away: verification, fitting and steering are the tasks that grew, so they get time in the estimate and a checklist of their own.
After AINothing on the increasing side is new work in kind — verifying, fitting and steering were always somebody's job. What has changed is that they are now one person's job, and that person used to do the producing.
opinionFrom execution to stewardship

The production was delegated and the accountability was not

Lee's authors call the worker's new role stewardship rather than teammate, collaborator or supervisor: the making of the thing has been handed over and the responsibility for it has not, which makes verification a standing cost rather than a stage on the way to trust.

Becauseit is the authors' reading of why the human-to-human words are wrong, argued from the displacement pattern in the free text; no measurement bears on it.
LimitsAccountability is a fact about institutions, not tools, and institutions can move it — insurers and vendors already absorb responsibility for machine output elsewhere. If that happens here, the asymmetry dissolves.
PracticeThe stays-human list — write down what the team will not hand over and why, and treat verification as a line in the plan with a name against it, not as something that will shrink as the tool improves.
After AIThe vocabulary problem is the real finding. Every available word for the relation borrows from human working relationships in which responsibility travels with the task, and none of them describes one where it does not.
opinionWhen they enact it, and when they do not

The tasks that seem too small to check are the practice

Skipping judgement on a task too trivial to be worth it harms nothing at the time, but 55 of 319 workers gave triviality as their reason and only 13 checked anything to keep a skill — so the high-stakes tasks arrive unrehearsed.

Becausethe counts are measured but the consequence is an inference, pointed back at Bainbridge by way of Simkute; no skill decay is observed in this study.
LimitsNothing here shows judgement decaying, and the survey is a single snapshot. The reading also assumes the judgement exercised on a trivial task is the capacity wanted on a consequential one, which the paper does not test.
PracticeManual flying hours — keep a low-frequency habit of doing by hand the class of task you have stopped checking, chosen from the trivial end rather than the important end, because the important end will look after itself.
After AIBainbridge's operators lost the practice because the system took it; here workers give it up themselves, one small task at a time, on a judgement correct each time.

All claims · Source map