Student Epistemic Agency in AI-Mediated Practices: Output-Directed and System-Directed Dimensions Diverge

Synthesis: Bergdahl (2026) surveyed 47 Swedish year 7 students who had spent a semester in an "AI on the schedule" pilot, testing whether epistemic agency directed at AI output extends to the systems producing it. It does not. Students showed strong agency on source evaluation and governance but weak agency on data privacy and algorithmic manipulation — and the split held within individual students: 79% had a score range of three or more across nine items, and a within-person dissociation index favoured output-directed over system-directed engagement in 35 students against 4 in reverse (sign test, p < .001). Among the 34 most source-critical students, 21 still endorsed "Good, then the answers get better" on chatbots reading their writing. Most strikingly, the low scores on data privacy were recorded one month after a lesson in which students watched their own logged behaviour projected live and called it "scary" — then stayed in through recess to keep using the platform. The study's contribution is that epistemic agency in AI-mediated practice is not a unified construct, and instruction that raises one register does not raise the other.

Study Design

The setting is a Swedish compulsory school that had launched a pilot called "AI on the schedule" — relevant because Sweden's national curricula include AI training only from upper secondary upward, so these year 7 students were receiving instruction that their curriculum does not require or specify. Three phases of training preceded the survey: fundamental AI concepts, ethics, prompting and a train-your-own-model exercise; a comparison of chatbots plus a tool visualising how datapoints track behaviour; and finally AI-supported game and story development. Those phases deliberately covered source evaluation, responsibility, data privacy, algorithmic manipulation and governance.

A nine-item survey (n = 47) was distributed in November, during phase three. Items were grouped a priori by the epistemic demand they place on the student:

• Output-directed (Q1–Q3) — evaluate AI output against external criteria.
• System-directed (Q4, Q6, Q7) — awareness of data practices, transparency, and algorithmic influence in everyday technology.
• Use-directed (Q5, Q8, Q9) — responsibility and ownership in AI-mediated work.

Each question offered four ordered alternatives (A lowest to D highest agency). Analysis used cross-tabulations of question pairs rather than scale scores, a choice that follows from the study's premise: if agency is domain-specific, an aggregate score destroys the finding. Classroom observation during the October lessons supplied group-level behavioural data to sit beside the November self-report.

Key Findings

• Agency is strong where it is easiest to perform and weak where the system is least visible. As a group, students scored highest on governance (Q9: Mdn = 4, 57% selecting the highest alternative) and on source evaluation and responsibility (Q2, Q3, Q5: all Mdn = 3), and lowest on data privacy (Q4: Mdn = 1, 66% at the lowest alternative) and algorithmic manipulation (Q7: Mdn = 2, with the largest single group at 47% on the lowest alternative). Grouped by domain, the median of output-directed medians was 3 against 2 for system-directed.

• The dissociation holds within students, not just between them. 79% (37 of 47) had a score range of three or more across the nine items, meaning the same student who scored high on one aspect scored low on another. A per-student dissociation index (median of output-directed minus median of system-directed) was positive for 35 students, reversed for 4, and flat for 8 — exact two-sided sign test, p < .001. Responses within the output-directed items were comparatively consistent, so the variation is between domains rather than noise.

• Source criticism and system criticism are separate registers. Among the 34 most source-critical students (Q3 = C or D), 21 (62%) chose Q4 = A — "Good, then the answers get better" — on chatbots reading their writing. Among students selecting Q3 = B, 82% chose Q4 = A, the highest concentration in any single cell. Only one student in the whole sample selected the highest-privacy alternative, and that student was also at the top on source comparison. The same pattern appeared on algorithmic influence: among the 11 who take "the answer that sounds best," 82% reported indifference to prolonged social-media scrolling.

• Direct instruction did not transfer, and the observation makes this vivid. During the October lessons the teacher demonstrated algorithmic data collection by having students enter a simulated social platform; while the teacher was still helping students log in, several were already clicking through the feed, unaware every action was being logged. The back end was then projected so the class could watch which students viewed what, how likes reshaped the feed, and how network analysis clustered them by inferred interests. Reactions included it being "scary how much data these collect," and one student still scrolling the demonstration feed remarked "it is hard to put it down." Several stayed in through recess to keep experimenting. One month later, 66% selected the lowest alternative on data reading and 47% the lowest on algorithmic influence. The observation and the survey are group-level and cannot be linked to individuals, but the juxtaposition is the paper's central provocation: the aspect made most directly visible became the aspect students reported least engagement with.

• Students double-check output when it seems wrong, which structurally shifts the burden onto prior knowledge. Bergdahl reads the reasoning data as showing students check AI output only when it is perceived as implausible — easier when they already know the topic. Students with less content knowledge may instead assume the AI knows more and accept output uncritically. This connects the agency finding to offloading behaviour: the trigger for verification is a mismatch the student must already be equipped to notice.

• Data collection is reframed as a benefit rather than a cost. Even after training on recommender systems, data collection and privacy, students tended to view it as something they gain from, since it presents content matching their preferences. On responsibility for flaws in human-AI output, many were indifferent, treating the current state of AI technology as an acceptable explanation for mistakes.

What the Evidence Can and Cannot Support

The author is unusually explicit about the design's limits, and they matter for how far this travels.

• Single school, single time point, no causal claim. One Swedish school, one survey administration, so no causal claim about the relationship between instruction and agency is available. The design cannot say whether the lesson failed to transfer or whether these attitudes preceded it.
• Self-report is reasoning about practice, not practice. A student who selected "I check where they got the information" may not actually do so. The classroom observation partly mitigates this, but at group level only.
• Social desirability runs in a specific direction, and it strengthens rather than weakens the headline. The ordered alternatives may have inflated the questions where a "correct" answer is socially visible — responsibility and governance — so those high scores may overestimate agency. The low scores on data privacy and algorithmic manipulation, where no socially expected answer exists, are therefore the more reliable indicators. The limitation cuts toward the finding rather than against it.
• Alternative explanations cannot be excluded. Aspects where students scored high may be inherently easier to grasp or addressed in other subjects, and the low scores may reflect developmental factors rather than instructional gaps — this is year 7, and abstract understanding of data economics is plausibly age-related.

What this means for practice

• Do not treat AI literacy as one outcome. Source evaluation and system awareness are separate capacities, and a lesson that lands one does not land the other: 62% of the most source-critical students still accepted chatbots reading their writing, and only one student in the sample chose the highest privacy stance.
• Assume that making a mechanism visible is not the same as making it matter. The class watched its own logged behaviour projected live, called it "scary," stayed through recess to keep using the platform — and a month later two-thirds chose the least protective option on data reading. Visibility without consequence is not instruction.
• Build verification triggers into the task rather than relying on students to notice implausibility. Students checked AI output when it seemed wrong, which works only if prior knowledge can flag the error. The students least equipped to notice are the ones who most need a structural prompt to check.
• Frame data collection as a trade, concretely. Students read it as benefit because the benefit is visible and the cost is not; instruction has to supply the missing side rather than assert that privacy matters.
• Treat measurement as part of any AI pilot. This study exists because a school built a pilot and someone tested it separately. A pilot that puts AI "on the schedule" without curriculum specification produces instruction whose effects are otherwise unknown.

Limitations

• Single school, single time point. One Swedish compulsory school and one November survey, so no generalisability claim and no causal inference about instruction — the design cannot say whether the lesson failed to transfer or whether these stances preceded it.
• Self-report measures reasoning about practice, not practice. A student who selected "I check where they got the information" may not actually do so; the group-level classroom observation only partly compensates.
• Social desirability runs toward inflating the high scores. Ordered alternatives may have overstated responsibility and governance, where a correct answer is socially visible. The low scores on data privacy and algorithmic manipulation, where no expected answer exists, are the more reliable indicators.
• Alternative explanations remain open. High-scoring aspects may be inherently easier to grasp or covered in other subjects, and the low scores may reflect developmental readiness rather than instructional gaps in year 7 students.
• Cross-tabulations show association, not causation, and the sample size limits statistical power in individual cells.
• The instrument was not piloted with a separate sample; comprehensibility rested on wording items in language observed in the classroom.

Citation

Bergdahl, N. (2026). Student epistemic agency in AI-mediated practices. Smart Learning Environments, 13(38).