Universities are increasingly asking whether generative artificial intelligence weakens students’ ability to think critically. A new study suggests that the more useful question may be how students organise their work with AI rather than simply how often they open the tool.
Among 430 college students in China, researchers found sharply different relationships with critical thinking depending on whether AI was used to take over parts of a task, provide guidance, or participate in an iterative collaboration. The strongest positive adjusted association appeared when students remained actively involved in shaping, evaluating and integrating AI output.
The findings do not show that one style of AI use causes stronger or weaker critical thinking. The study was cross-sectional, meaning students’ existing thinking habits could also influence how they choose to use AI.
Three ways of working with the same technology
The study, published in Frontiers in Psychology on 28 September 2026, surveyed students enrolled in vocational, undergraduate and postgraduate programmes across 30 provincial-level administrative regions in China. Data were collected online from May to July 2026.
The researchers began with 447 responses and excluded 17 participants who selected the same response option across all 25 focal questionnaire items. That left an analytic sample of 430, a retention rate of 96.2%.
The sample was 52.8% female and 47.2% male. About 43.3% were bachelor’s students, 39.8% attended vocational colleges and 17.0% were postgraduates.
AI use was already routine for much of the group. More than half, 54.4%, had used generative AI for over a year, while 40.5% reported using it almost daily and another 37.2% used it three to five times a week.
Instead of collapsing those behaviours into a single measure of AI adoption, Shuai Yin, Yitao Gan and Qing Zhao separated use into three orientations. The distinction centred on who retained responsibility for doing and evaluating the academic work.
The first was a Tool orientation. This represented comparatively delegative use, including submitting task requirements to AI, directly using generated output and adopting answers with relatively little independent revision.
A Guidance orientation described students asking AI for explanations, examples, clues or diagnostic help while retaining responsibility for understanding the problem. Collaboration went further, capturing repeated exchanges in which students supplied task information, adjusted prompts, evaluated responses and developed the output over successive interactions.
Collaboration showed the strongest positive association
Critical thinking disposition was measured using 11 items covering a student’s inclination to question claims, consider alternatives, use evidence and reflect on judgement. The measures showed strong internal consistency, with Cronbach’s alpha ranging from 0.839 to 0.943 across the six reported scores.
At the simple correlation level, Tool orientation had virtually no relationship with critical thinking disposition. Its correlation was -0.044 and was not statistically significant.
Guidance and Collaboration looked very different. Guidance correlated with critical thinking disposition at 0.384, while Collaboration produced a correlation of 0.525. Both were statistically significant at p below 0.001.
The researchers then used hierarchical ordinary least squares regression to separate the orientations from demographic characteristics and general AI experience. The first model accounted for gender, education level and field of study, while the second added duration of AI use, weekly frequency and self-reported proficiency.
Adding the three usage orientations increased explained variance by 10.37 percentage points. The complete model explained 38.95% of the variance in critical thinking disposition, with an adjusted R-squared of 36.4%.
Collaboration had the largest positive standardised coefficient at 0.321, with a 95% confidence interval from 0.159 to 0.483. Guidance was also positively associated with critical thinking disposition, but more modestly, with a standardised coefficient of 0.126 and a 95% confidence interval from 0.023 to 0.229.
The researchers did not formally test whether the Collaboration coefficient was statistically larger than the Guidance coefficient. Its larger point estimate should therefore not be treated as proof that collaboration is definitively superior.
Delegation became negative only after adjustment
The Tool result was more complicated. Although its simple relationship with critical thinking was close to zero, its adjusted coefficient became negative once the other orientations and covariates were considered.
In the complete model, Tool orientation had a standardised coefficient of -0.189, with a 95% confidence interval from -0.274 to -0.104 and p below 0.001. The authors describe the shift as consistent with statistical suppression.
This matters because the three orientations were not mutually exclusive. A student could delegate one activity to AI, seek guidance on another and work collaboratively with it on a third.
Tool use itself was positively correlated with both Guidance and Collaboration. Once those overlapping patterns were statistically separated, comparatively greater reliance on AI-centred task completion was associated with lower critical thinking disposition.
That is a conditional association, not evidence that asking AI to complete a task directly damages critical thinking. The study did not observe students over time or randomly assign them to different AI-use strategies.
Frequency alone tells an incomplete story
The results challenge a common way of discussing AI in education. Counting logins, hours or prompts can describe exposure to the technology, but it may obscure what students are actually doing cognitively while using it.
Two students could both use a chatbot every day while allocating responsibility very differently. One may paste an assignment requirement into the system and accept a finished answer, while another may request an explanation, challenge the response, refine a prompt and combine the output with independent evidence.
The study’s results are consistent with the idea that these patterns contain information that broad measures of experience miss. AI-use background increased explained variance by 4.9 percentage points after demographics, while the orientation block added a further 10.4 percentage points.
For universities, this suggests that policies focused only on whether students may use generative AI could be too blunt. Course design may need to pay greater attention to which parts of intellectual work students are expected to retain, how AI output should be evaluated and where independent judgement must remain visible.
That does not require treating delegation as inherently illegitimate. Cognitive offloading can free attention for more demanding work, just as calculators, reference materials and search engines do. The educational question is what is being offloaded and whether students still evaluate the result.
The study cannot establish cause and effect
The strongest limitation is timing. All focal variables were measured at one point, so the analysis cannot determine whether collaborative AI use contributes to critical thinking disposition or whether students who already think more critically are more inclined to collaborate with AI.
The convenience sample also cannot represent all Chinese college students. Motivation, prior academic achievement, course design, institutional AI policies, digital literacy and previous critical-thinking training were among the potentially relevant factors that were not measured.
The researchers also caution that self-reported disposition is not the same as demonstrated critical-thinking performance on a task. Future longitudinal and experimental research would be needed to test direction, causality and whether different AI-use patterns produce measurable changes in unaided reasoning.
Even with those boundaries, the study offers a useful refinement to the debate. Generative AI use is not a single behaviour, and the same technology can occupy very different roles in a student’s academic work.
The emerging question for education may therefore be less about how much AI students use and more about whether they remain responsible for questioning, evaluating and integrating what the machine gives them.
Source Information
Study Title: Associations between generative AI usage orientations and critical thinking disposition among college students in China
Authors: Shuai Yin, Yitao Gan and Qing Zhao
Journal: Frontiers in Psychology
Year: 2026
DOI: 10.3389/fpsyg.2026.1957029
