AI Availability Reshaped Students' Writing Motivation Profiles Without Moving the Average

Synthesis: Four hundred and six students in Grades 5–8 at three German secondary schools wrote two comparable creative-narrative tasks — one with no digital tools, one with GenAI optionally available — and reported their Motivation regulation after each. Mean levels barely moved, which is what a conventional analysis would have concluded. Person-centered latent profile analysis found five motivational configurations at both occasions, substantial redistribution between them, and one psychological need that did drop: autonomy satisfaction fell by 0.28 on the scale. How much a student used the AI did not predict which profile they belonged to.

Key Findings

1. The sample comprised 406 students from three German secondary schools nested in 21 classrooms, in Grades 5–8, mean age 11.85 years (SD = 1.21), 46.1% female, participating within regular German language instruction.
2. Both writing occasions were individual creative narrative tasks of about 150 words and 60 minutes, differing in prompt and in AI availability; Time 1 permitted no digital tools, and Time 2 offered GenAI through fobizz, a protected classroom environment.
3. A five-profile solution was retained at both Time 1 and Time 2 on the joint criteria of information criteria, likelihood-ratio tests, classification quality, profile sizes and interpretability; profiles were labeled high-quality, high-quantity, low-quality, low-quantity and moderately motivated.
4. Mean levels of intrinsic, identified, introjected and external regulation were stable across contexts, while person-centered analysis showed heterogeneity in how the regulations combined and differences in the prevalence of comparable profiles.
5. Autonomy was the only basic psychological need to decline between occasions, Δ = −0.28, p < .001; need satisfaction — autonomy and competence in particular — predicted membership in more self-determined profiles.
6. End-of-term German grades differentiated profile membership at both occasions, tying motivational organization to academic functioning rather than to task enjoyment alone.
7. Cross-context transitions were mostly diagonal (0.59 to 0.74) but not fully so: 41% of the low-quality profile moved to low-quantity, and a quarter of the high-quality profile moved to moderately motivated, which is why the authors treat the pattern as redistribution rather than as movement between equivalent states.
8. Degree of AI use at Time 2 did not significantly predict profile membership — a student's regulatory organization was not explained by how much they used the tool.

Why the average hides the change

This is the methodological argument of the paper, and it is worth taking seriously beyond AI research. If you ask whether self-determined motivation changed when AI became available, the answer from mean comparison is essentially no. Every regulation mean was statistically indistinguishable across the two contexts, and the authors are explicit that a variable-centered reading would have concluded stability.

The person-centered reading tells a different story. The same five configurations describe both occasions, but their relative sizes shift, and students do not stay put: the transition matrix shows a majority remaining in an equivalent profile while meaningful minorities move. Whether that movement reflects the writing task, the AI's presence, the passage of time or simply a different prompt cannot be settled by this design, which is why the authors interpret the pattern descriptively instead of as latent-state transition.

For writing education the practical translation is that averaging over a class will report stability while individuals are reorganising. A teacher who sees no change in "motivation" on a class survey may still be teaching a room in which a third of students have rearranged how they regulate the work.

Autonomy was the need that moved

The basic-needs result gives the paper its cleanest mechanism claim. Autonomy satisfaction declined at Time 2 while competence and relatedness did not, and need satisfaction — autonomy and competence especially — predicted who sat in the more self-determined profiles. Grades predicted membership too, at both occasions.

That combination matters for how AI is introduced rather than whether. If the resource arrives in a form that takes decisions away — what to write, how to open, what counts as good — the need most sensitive to that loss is the one that declined here. A design that keeps choice over topic, opening, structure and whether to consult the tool at all is not a soft preference; on this evidence it is the condition under which the more self-determined configurations stay populated.

The authors also separate two things that are routinely conflated. AI availability is a feature of the environment; AI use intensity is a behavior. Only the first produced detectable differences here, and use intensity predicted nothing. A classroom where the tool is offered and declined is a different psychological situation from one where it is absent, and a policy that measures only usage counts will miss that distinction entirely.

What a two-wave design can and cannot claim

The two tasks were close in time, with a break of about an hour between the Time 1 questionnaire and the introduction of the AI-available context, and the authors state that the design compares motivational organization across two contexts rather than serving as a pretest–posttest intervention. Carry-over from the first task is therefore plausible and unmeasured.

Measurement invariance for motivational regulation held up to strict invariance across occasions, which is a real strength: the construct was measured the same way in both contexts. For need satisfaction only partial scalar invariance was established, so mean comparisons of needs carry more uncertainty than the motivational regulation comparisons do — relevant when reading the autonomy decline as a quantity rather than a direction.

Everything is self-reported, by children, in classroom sessions, at two points. No behavioral trace of AI use was collected, and no task-quality measure anchors the motivational data to what students actually wrote.

What this means for practice

• Instructors. Judge a class-level motivation survey as an average that can mask redistribution. In this sample the means held steady while a meaningful minority of students changed regulatory profile, so individual check-ins and work samples carry information a whole-class score does not.
• Curriculum designers. Protect autonomy when you introduce GenAI. The need that declined was autonomy, and autonomy predicted more self-determined profiles, so choice over topic, opening, structure and whether to consult the tool is a design constraint rather than a courtesy.
• School and district decision-makers. Do not evaluate an AI rollout by usage counts. Availability and use intensity behaved differently in this study, and usage was unrelated to motivational profile membership.
• Tool designers. Build for optionality. A classroom product that makes the tool available without making it the default path preserves the condition under which the more self-determined profiles stayed populated here.

Limitations

• The design compares two contexts rather than testing an intervention; carry-over between the two tasks, which were separated by about an hour, is not modeled.
• Motivational regulation and need satisfaction are self-reports collected in classroom sessions, with no behavioral measure of AI use and no assessment of the texts students produced.
• Partial scalar invariance for basic psychological needs limits how far the autonomy decline can be read as a precise magnitude, though strict invariance held for motivational regulation.
• The sample is 406 students in Grades 5–8 at three German secondary schools, so generalization to other systems, ages and writing genres is untested.
• Because full profile similarity across contexts was unsupported, the transition estimates describe redistribution in this sample rather than movement between equivalent latent states.
• End-of-term grades were the only achievement indicator, taken from school records rather than from the writing tasks themselves.

Citation

Schweder, S., & Grahl, L. (2026). Motivational organization in AI-supported writing: A self-determination perspective. Computers and Education Open, 11, 100430.