Self-Directed Generative AI Use and Academic Development Among Secondary School Students: A Large-Scale, One-Year Longitudinal Study

Synthesis: Most secondary students now encounter generative AI on their own, without pedagogical guidance — and whether it supports or undermines academic development remains contested. Analyzing TOSCAneo data from 4,062 students in 136 secondary schools in Germany, this one-year longitudinal study separates two factors of self-directed Generative AI use: task-completion use and learning-oriented use. Both increased across the 2023/2024 school year (d = 0.35 and d = 0.33), and only 5% of students never used GenAI, making Large Language Models (LLMs) tools a near-universal presence in K-12 schooling. Lower prior achievement predicted more subsequent use, while neither form predicted later achievement; a task-completion-weighted profile predicted lower subsequent effort. Guidance therefore matters most for lower-achieving students, a finding that reframes Digital Divide and Differential Effects Across Learner Groups concerns around purpose and Student-AI Interaction rather than access. Policy should shift from restricting tools toward Framing AI Use for Students and the Self-Regulated Learning and Self-Directed Learning skills behind productive use.

Key Findings

1. Sample and design. TOSCAneo data cover 4,062 students in 136 secondary schools in Germany, with standardized tests at the beginning and end of the 2023/2024 school year.
2. Use grew sharply. GenAI use increased for task-completion (d = 0.35) and learning-oriented use (d = 0.33), with only 5% of students never using GenAI.
3. Achievement predicts use. Lower prior achievement predicted more subsequent GenAI use in all six subjects (task-completion) and in five of six (learning-oriented) (-.28 ≤ β ≤ -.12, all ps < .05).
4. Use does not predict achievement. Neither form predicted subsequent achievement in any subject; item-level growth ranged from d = 0.21 to d = 0.65.
5. A task-completion profile predicts lower effort. It predicted lower subsequent effort in four of six subjects; learning-oriented use predicted higher effort in three (physics β = .39, p < .001).
6. Motivation effects were isolated. No comparable pattern emerged for self-concept or value beliefs.

Does unguided GenAI use help or harm academic development?

Generative AI tools such as ChatGPT, Claude, and Gemini are transforming how students learn, yet research has concentrated on pedagogically designed systems implemented in supervised settings. Less is known about self-directed use, which occurs without predefined instructional designs or systematic guidance. The authors argue that effects depend less on whether students use GenAI and more on the purpose for which it is used: forms that reduce active cognitive engagement may limit practice and knowledge construction, whereas uses that support explanations, examples, and exam preparation may promote deeper processing. The study measures self-directed use rather than inducing it, uses standardized tests rather than self-reported grades, and models both directions.

Two forms of self-directed use

Students rated how often they used GenAI for seven academic activities on a 5-point scale from never to every school day. The task-completion factor captures four items — translating texts, doing homework, preparing presentations and essays, and solving tasks during class — while the learning-oriented factor captures three: better understanding of class material, further learning, and preparing for exams or tests. A two-factor model fit better than a one-factor model (Δχ²[6] = 267.51, p < .001), with factors correlated at r = .86 (T1) and r = .81 (T2). The distinction connects to Motivation theory and the engagement learning-oriented use supports.

Achievement predicts use; use does not predict achievement

Cross-lagged panel models for each subject revealed two main patterns. Lower achievement at the start of the year predicted more frequent GenAI use at the end, after controlling for prior use and all other variables — in all six subjects for task-completion use and in five of six for learning-oriented use. Neither form of use predicted subsequent achievement anywhere. Effort differed: task-completion use predicted lower later effort in mathematics (β = -.24, p = .002), chemistry (β = -.25, p = .003), physics (β = -.42, p < .001), and English (β = -.25, p = .008), while learning-oriented use predicted higher effort in mathematics (β = .21, p = .011), chemistry (β = .19, p = .023), and physics (β = .39, p < .001). The opposite-sign coefficients suggest the balance between completion and learning matters more than overall frequency.

Guiding GenAI use in schools

The authors conclude that guidance should focus on how students engage with GenAI, not whether they use it, and that support is particularly important for lower-achieving students, who reported more frequent use of both forms; they stop short of recommending less use, arguing instead that schools should help these students sustain productive engagement. Developing metacognitive judgments about when GenAI facilitates learning and when it substitutes for effort should become part of GenAI literacy, and effort changes may precede achievement changes over longer periods. Task-completion use is not inherently maladaptive — offloading a subtask can free cognitive resources — but distinguishing purposes is the first step for guidance.

What this means for practice

• Instructors. Teach students to use GenAI for explanation, clarification, and revision rather than effort-minimizing task completion, and aim this guidance especially at lower-achieving students.
• Administrators. Shift policy from whether GenAI is used to how it is used: only 5% of students never use these tools, so restriction is no longer a realistic lever.
• Policymakers. Treat metacognitive judgment about when GenAI supports versus substitutes for learning as a core component of AI literacy in secondary curricula.
• Researchers. Pair longitudinal designs with process measures of student–AI interaction to test whether task-completion use erodes effort beyond one year.

Limitations

• GenAI use measures were not subject-specific, while achievement, motivation, and effort were assessed across six subjects.
• The two use factors were strongly correlated (r = .86 at T1; r = .81 at T2), exceeding their cross-time stability (.62 and .51), so the effort findings rest on limited unique variation.
• Cognitive engagement was not measured directly, and with only two measurement occasions the model cannot separate stable between-person differences from within-person change.
• Analyses rely on self-reported use from students averaging 16.14 years (SD = 0.84) in one German state; the study was not preregistered.

Citation

Scherer, R., Kizilcec, R. F., Bäulke, L., Hübner, N., Meyer, J., Trautwein, U., Wagner, W., & Nagengast, B. (2026). Self-Directed Generative AI Use and Academic Development Among Secondary School Students: A Large-Scale, One-Year Longitudinal Study. PsyArXiv preprint (osf.io/m93n6).