The Offloading Effect
Opens the original study.
- Study
- Quantitative manuscript, cross-sectional survey
- Who
- 302 engineering students and scholars
- Method
- PLS-SEM with serial-parallel mediation
- Checks
- Bootstrapping, MICOM, multi-group, endogeneity, curvilinear, NCA
- When
- July 2026
01 Question
Does heavier AI use weaken self-reflection and lower grades?
The concern is specific: growing AI dependency erodes metacognitive awareness, and weaker performance follows. 302 engineering students and scholars answered validated five-point scales on six constructs.
- Respondents302
- Constructs6
- Hypotheses8
02 Method
Eight hypotheses were tested as a chain through confidence and offloading.
The model is PLS-SEM with serial-parallel mediation. Bootstrapping, MICOM, multi-group, endogeneity, curvilinear and necessary-condition analyses then checked it.
- ModelPLS-SEM
- Checks on top6
- Respondents302
03 Findings
Seven of the eight hypotheses are supported.
Dependency drives both offloading (beta 0.46) and confidence (beta 0.53). Offloading raises metacognitive awareness (beta 0.33), yet awareness does not predict performance (beta 0.05). Two of the eight were data-informed revisions, disclosed in the paper.
- Supported7 of 8
- Dependency to confidence0.53
- Awareness to performance0.05
03 Findings
Confidence and offloading carry the effect on performance.
Confidence (beta 0.40) and offloading (beta 0.35) are the two strongest routes, and together carry about three-quarters of dependency's effect. Dependency helps at the margin and is never necessary, since the analysis finds no minimum level of AI use.
- Confidence to performance0.40
- Offloading to performance0.35
- Of the effect carriedAbout 75%
03 Findings
The model explains 60% of the variance in performance.
R squared for performance is 0.60 and Q squared is 0.59, so the model predicts out of sample. The confidence pathway flattens at high dependency, and in heavy users confidence stops predicting performance.
- R squared0.60
- Q squared0.59
Respondents
302 people
Each dot is one respondent. Dots show group sizes, not individual answers.
04 Directions
- 01Design for the confidence plateauReduce AI scaffolding as competence signals rise, and surface moments when a student solved something unaided.
- 02Make offloading a decisionShow what the AI carried and what the student carried, and ask for an explain-back step.
- 03Build verification into the workflowReflection does not transfer on its own, so test-before-accept gates replace the hope that students check.
05 Limits
- H4 and H8 replaced an original pair that failed. The paper discloses the revision, and the results carry that weight.
- Performance is self-reported and AI-attributed, and the direct path carries an endogeneity flag, so the direct effect needs caution.
- The survey is cross-sectional, so it shows associations and not change over time.
- Heavy and light users behave differently, and the split follows usage rather than seniority.
The research site
Get in Touch