Conference Program

When an algorithmic system classifies a student as "low risk" without that assessment being open to verification or discussion, it does not merely make a technical error: it forecloses the very possibility of educational dialogue. The teacher cannot open a conversation about what they cannot see, the student cannot regulate their learning on the basis of criteria they are unaware of, and the school cannot exercise its function if its decisions are delegated to an opaque procedure. This contribution argues that the problem is not the use of AI in education, but the conditions under which it is used: conditions which, when absent, make AI incompatible with any serious conception of critical learning.

Authentic learning requires that students have access to the criteria by which they are judged, that they can monitor and regulate their own progress, and that dialogue with tools and teachers remains open and contestable (Vygotsky, 1978; Flavell, 1979; Bakhtin, 1981). An AI system that produces predictions without explaining them violates all three conditions at once. The AI Act (EU 2024/1689), Arts. 9, 10 and 14, translates this into normative obligations: human oversight, data quality, and explainability of high-impact decisions, treated here not as bureaucratic constraints but as minimum pedagogical conditions.

The contribution adopts a two-level mixed design. At the macro level, an empirical analysis of territorial bias across 190 European NUTS2 regions uses Eurostat data on early school leaving rates and regional GDP per capita (2024) to show which students an algorithm calibrated on the European average renders structurally invisible. At the micro level, four case studies from the Educational Data Mining Conference 2023 are analysed for the centrality they assign to teachers and students.

At the macro level, a statistically significant correlation emerges between regional GDP per capita and the algorithm's prediction error (r = −0.194, p = 0.007): disadvantaged regions are systematically underestimated as "low risk", with a structural gap of 0.68 percentage points (t-test p = 0.044). The bias follows economic, not geographic, lines (Kruskal-Wallis p = 0.32): 82 false negatives appear because students live in low-income contexts the algorithm does not account for.

At the micro level, four recurring conditions emerge. Gabbay and Cohen (2023) show that metacognitive feedback in programming MOOCs produces more autonomous learners than systems providing solutions directly. An Accountable Talk case mediated by GPT-4 (EDM 2023) demonstrates that language models can sustain critical discourse when designed to elicit argumentation, not replace it. Wagner et al. (2023) show that a transparent recommendation system reduces dropout without a self-fulfilling prophecy effect. Karimov et al. (2023) document how an adaptive platform in a disadvantaged context in Azerbaijan improved grades for 68.6% of students, showing that algorithmic equity is a matter of design.

Both levels converge: AI is compatible with education only if verifiable — if teachers and students can interrogate it, challenge it, and correct it. Without explainability and human accountability, the school does not produce critical learners: it produces consumers of predictions who do not know they can refuse them.