Whether generative AI erodes the skills it assists is usually studied by counting how much people use it, and while the tool is still in hand. That design cannot separate cause from selection, or distinguish using AI from depending on it. We report a three-arm experiment (900 randomized; 659 in the primary analysis) that varies the manner of AI use against a no-AI anchor and measures performance 15–20 min later, once the tool has been removed. Participants wrote an argumentative essay under a self-generation-first autonomous workflow (commit one's own position and then interrogate AI), a direct-adoption dependent workflow, or no AI and then wrote a second, parallel essay with no AI available. Two blind raters scored reasoning quality, idea diversity, and originality. On the unaided task, dependent users reasoned worse than participants who had never used AI (adjusted difference = −0.38 on a 1–7 scale, a small effect, d = −0.28, two-sided p = 0.0002), whereas autonomous users did not fall below that baseline. The pattern held under per-protocol, covariate-adjusted, Phase-1-time-adjusted, and full-randomized-sample multiple-imputation checks. A trait disposition toward autonomous offloading was discriminant from reflective thinking and critical-thinking disposition (HTMT ≤ 0.43) and incrementally predicted unaided reasoning. We did not detect a metacognitive pathway, although the design cannot exclude one. Because the two workflows differ in self-generation, elaboration, revision, and effort as well as in autonomy, the estimate belongs to the workflow contrast as a whole rather than to any single component: A generation-effect account is consistent with the pattern, but this design cannot weigh it against the alternatives, and no mechanism is established here. The result is an immediate near-transfer decrement rather than demonstrated lasting de-skilling. What separated the arms was not access to the model but the order in which the interaction required the user to produce something.

Instructors can prescribe which part of a writing task a student may delegate to generative AI, but not how the student works. This four-arm randomised experiment separated them. At Site A, 744 undergraduates were randomised to one of four divisions of labour across four rounds of report writing—AI drafts and the student evaluates and revises; the student drafts and AI evaluates; the student drafts, AI revises, and the student adjudicates each change; no AI—with all AI withdrawn at Week 6 post-test, immediately after the final round (639 completed). A reduced three-arm replication at Site B randomised 126 students (109 completed). Both arrangements in which the student judged AI-produced text outperformed the arrangement in which AI judged student-produced text and the no-AI control (d = 0.35 to 0.36), whereas having AI evaluate was equivalent to no AI. The benefit was unevenly distributed (arm × tendency F(3, 629) = 7.25, p < 0.001), and logs showed students high in dependent tendency re-outsourcing their assigned evaluation (b = 0.072, p < 0.001). Satisfaction carried no information about error detection (r = 0.015), while self-rated quality was weakly associated with writing performance. Prescribing a division of labour governs the assignment, not the cognition.

Continued use of fitness technology is often treated as a prerequisite for health behavior change, yet platform engagement and behavioral regulation may operate through different mechanisms. This four-wave panel study tested whether an engagement pathway in online fitness communities extends across the digital–physical boundary. After a priori quality screening, 506 Chinese adults who had used the community functions of a smart fitness platform, application, or wearable device within the previous three months and had access to a community activity venue were retained at T1, with 417 at T2, 377 at T3, and 394 at the outcome-only T4 wave. A unified full-information maximum-likelihood path system adjusted for prior outcomes and demographics and used 2000 participant-level bootstrap resamples. Perceived interaction at T1 predicted social presence ($\beta$ = 0.297), flow ($\beta$ = 0.355), and belonging ($\beta$ = 0.245) at T2, and all three independently predicted continuance intention at T3 ($\beta$s = 0.203–0.302). Continuance intention is the intention to keep using the community, not observed use. It showed no detectable association with physical activity at T4 once prior activity was controlled ($\beta$ = 0.016, 95% CI [−0.061, 0.082]), and the serial indirect effects were likewise undetectable. This did not follow from conditioning on the strong autoregressive path: the zero-order correlation between continuance intention and later activity was 0.068, while associations along the hypothesized within-platform chain ranged from 0.218 to 0.563. Missing-data, complete-case, log-transformed, and inverse-probability-weighted specifications converged. Exploratory equivalence testing rendered standardized effects of ±0.10 or larger implausible (p = 0.012) but not ±0.05, and a BIC approximation favored the model without the path (BF₀₁ = 17.81). Experiences that sustain digital continuity should not be assumed to produce physical activity spillover.

Evidence that AI-driven adaptive learning outperforms uniform instruction comes almost entirely from student samples in educational settings; whether, why, and for whom it works in real-world workplace training remains largely untested. We conducted a three-wave, matched-pair, cluster-randomized field experiment across six county-level power supply companies of a Chinese provincial electric power utility. Employees ( N = 3,000 invited; analysis sample N = 2,574) received either knowledge-graph-driven adaptive training or traditional uniform training covering the same syllabus. Training batches were matched on batch-level characteristics and allocated by random draw before enrollment. Because the two arms differ in delivery medium as well as in adaptive routing, the contrast estimates the effect of a bundled adaptive delivery system rather than adaptivity in isolation. Training self-efficacy was measured at T1 and T2, whereas transfer intention and knowledge were measured at T1, T2, and at the 6-week follow-up (T3). Cognitive load and perceived personalization fit were measured at T2, and measurement invariance across conditions was confirmed by confirmatory factor analysis. Linear mixed models showed small but robust advantages in training self-efficacy ( d ≈ 0.26), transfer intention ( d ≈ 0.25), and knowledge ( d ≈ 0.29 at T2), with all effects surviving three clustering levels, a company-level wild cluster bootstrap, and a stratum-level permutation test. No reduction in the knowledge advantage was observed at follow-up, and equivalence testing bounded any change to less than 0.10 SD (T3 d ≈ 0.26). Parallel mediation showed statistical indirect effects of reduced extraneous load and higher perceived personalization fit on self-efficacy and knowledge, whereas the pathway to transfer intention ran through fit alone. Cognitive-load subscales diverged: adaptive training reduced extraneous load ( d = −0.34) but increased intrinsic load ( d = 0.28), partially supporting the load hypothesis and cautioning against treating cognitive load as a unitary construct. Exploratory moderation indicated that more experienced employees derived greater benefits, contrary to the compensatory pattern reported for novice workers. Findings extend cognitive load theory and personalization research to a high-reliability industrial workforce and suggest that the motivational returns of adaptive training are tied to employees' perceptions that the training fits them.