New scale measures how Chinese college students delegate memory, tasks, and thinking to AI

As generative artificial intelligence becomes woven into everyday study habits, researchers have developed and validated a new instrument designed to measure precisely what college students delegate to AI systems when they learn. The AI Cognitive Offloading Scale, or AI-COS, was created by Hongying Chang of Minzu Normal University of Xingyi and published in Frontiers in Psychology. Rather than treating AI use as a single behavior, the scale distinguishes three functions of cognitive delegation: memory offloading, in which students hand storage and retrieval of information to AI; assistive offloading, in which AI participates in executing and organizing academic tasks; and thinking offloading, in which AI contributes to judgment, decision support, and other higher-order cognitive activities.

The scale was built from a pool of 26 items generated through literature review, theoretical analysis, and expert evaluation. Five experts with doctoral degrees in psychology or educational technology rated each item on a three-point relevance scale, and the wording was revised in response to their feedback. The study then proceeded in two stages with independent samples. Study 1 drew on 665 college students, recruited through university student and class groups using convenience sampling and surveyed online, for item analysis and exploratory factor analysis. Study 2 used a fresh sample of 1,123 students for confirmatory factor analysis, reliability testing, criterion-related validity assessment, and measurement invariance testing across gender.

In the exploratory stage, all 26 preliminary items passed empirical screening: item discrimination t values ranged from 14.32 to 27.05, item-total correlations ranged from 0.57 to 0.75, and the overall Cronbach’s alpha was 0.948. Principal axis factoring with direct oblimin rotation was then applied, chosen because the three dimensions were theoretically expected to be correlated. The analysis retained a three-factor solution explaining 54.68 percent of cumulative variance, with a KMO value of 0.965 and a significant Bartlett’s test of sphericity. Six items were removed during purification, three for factor loadings below the 0.40 retention threshold and three for semantic redundancy, leaving a final 20-item scale with loadings between 0.447 and 0.750 and no substantial cross-loadings.

The resulting dimensions map onto distinct cognitive functions. Memory offloading comprises five items describing tendencies to delegate remembering, storing, and retrieving information to AI, such as relying on AI history records instead of attempting recall. Assistive offloading contains ten items covering AI-supported analysis, content generation, summarization, and task completion, including habits such as having AI generate drafts of assignments before making only minor revisions. Thinking offloading includes five items reflecting reliance on AI for judgment, decision support, and cognitive guidance, such as letting AI make final decisions among options or adopting AI suggestions as guiding principles. The authors emphasize that this ordering describes the cognitive functions represented by the items, not a developmental or temporal sequence through which students progress.

Study 2 compared three competing confirmatory factor models in the independent sample of 1,123 students. A one-factor model showed poorer fit, and a first-order three-factor model fit reasonably well, but a second-order model, in which AI cognitive offloading was specified as a higher-order construct underlying the three dimensions, fit best, with CFI of 0.951, TLI of 0.944, RMSEA of 0.047, and SRMR of 0.036. Standardized CFA loadings ranged from 0.598 to 0.786 across items, and the three first-order dimensions loaded strongly on the higher-order factor, at 0.868 for memory offloading, 0.964 for assistive offloading, and 0.866 for thinking offloading. This structure supports reporting both a total score and dimension-specific scores while preserving the distinct interpretation of each function.

Reliability evidence was strong. In the Study 2 sample, the overall scale showed Cronbach’s alpha of 0.944 and split-half reliability of 0.932, with subscale alphas ranging from 0.834 to 0.912. A subsample of 99 students completed the scale again after one month, yielding a test-retest reliability coefficient of 0.937 for the total scale and coefficients between 0.833 and 0.899 for the dimensions. Composite reliability values for the three dimensions were 0.840, 0.906, and 0.869, and average variance extracted values were 0.501, 0.507, and 0.570, supporting convergent validity. The authors note candidly that because no heterotrait-monotrait ratio or latent-factor correlation matrix was retained, strict statistical discriminant validity among the three highly related dimensions has not been fully established and should be examined in future validation work.

Criterion-related validity was assessed against three external measures. The total AI-COS score correlated positively with a measure of student generative AI dependence (r = 0.619), and all three dimensions showed positive correlations with dependence as well (r = 0.489 to 0.593). The total score and all three dimensions were negatively associated with metacognitive monitoring, with correlations ranging from −0.180 to −0.245. Associations with need for cognition were more differentiated: the total score (r = −0.076) and assistive offloading (r = −0.121) showed significant negative correlations, whereas memory offloading and thinking offloading were not significantly related to this measure. The authors stress that these are cross-sectional correlations and do not establish that offloading causes cognitive impairment or benefit; rather, they suggest that the psychological meaning of offloading may depend on which cognitive function is delegated and how learners regulate the process.

Measurement invariance testing across gender supported the scale’s comparability between male and female students. Multigroup confirmatory factor analysis tested configural, metric, and scalar models sequentially, and fit deteriorated negligibly at each step, with changes in CFI, TLI, and RMSEA well within the prespecified criteria. The study therefore reports full scalar invariance across gender, indicating that the instrument measures the same underlying construct in both groups and providing a basis for future gender comparisons. The authors also acknowledge that content validity relied on expert review without formal item-level or scale-level content validity indices, that no qualitative interviews informed item generation, and that the samples were drawn from undergraduate students in a specific contextual setting, limiting generalizability to other regions, institution types, and graduate populations.

The study’s contribution is positioned as complementary rather than first-in-field. Earlier work produced an AI-Based Cognitive Offloading Scale organized by cognitive-demand level, distinguishing lower-order from higher-order offloading, and other research has distinguished dependent from autonomous offloading according to whether users retain cognitive agency. The AI-COS instead classifies offloading by the primary cognitive function delegated during college learning, offering a three-dimensional, learning-context-specific operationalization that can be used alongside these existing approaches. The authors frame the instrument as a foundation for research on AI-mediated cognition, learning regulation, and student development, and for guidance in higher education that focuses not merely on whether students use AI but on which cognitive functions they delegate and how they monitor that delegation.

Several limitations temper the findings. The design was cross-sectional, so dynamic changes in offloading during learning and causal inference about its antecedents and consequences remain beyond reach. The rapid evolution of generative AI may continue to reshape how students use such systems, requiring future item revision. The authors recommend that future studies incorporate qualitative interviews, learning logs, or behavioral data to examine whether different forms of offloading correspond to different actual learning outcomes, and that longitudinal and experimental designs directly test when offloading preserves or enhances cognitive performance and when it may undermine learning, cognitive independence, or metacognitive control. For now, the scale offers researchers a structured way to ask not simply how much students rely on AI, but what exactly they are handing over to it.

Subject of Research: Development and validation of a scale measuring AI cognitive offloading among Chinese college students

Article Title: Development and validation of the AI cognitive offloading scale for Chinese college students

Article References: Chang, H. (2026). Development and validation of the AI cognitive offloading scale for Chinese college students. Frontiers in Psychology, 17, Article 1884676. https://doi.org/10.3389/fpsyg.2026.1884676

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1884676

Keywords: AI cognitive offloading, generative AI, psychometrics, scale development, college students, memory offloading, assistive offloading, thinking offloading, metacognitive monitoring, need for cognition, measurement invariance, educational psychology