Einfluss der Passung des KI-gestützten Scaffoldings auf den Lerneffekt: Zusammenhänge mit Angst, Scham und dem Interaktionsverhalten

There is general scientific consensus in educational psychology regarding the influence of emotional
and social factors on learning outcomes. The increasing adoption of artificial intelligence including in
educational institutions has made the transferability of classical learning theories to AI-supported
learning environments relevant. Against this backdrop, this study analyses the significance of
achievement emotions and interaction behavior using an AI-supported tutor.
The goal of this bachelor’s thesis was to explore whether the relationship between the fit of AI-supported
scaffolding and learning outcomes is mediated by negative achievement emotions such as anxiety or
shame, as well as interaction behavior with the AI tutor. The study employed an online repeatedmeasures
design with N = 33 participants. The self-regulated chatbot condition provided adaptive
support based on individual needs, whereas in the over-regulated condition a consistently high level of
support was provided, whether or not it was actually needed.
The hypothesis-driven analyses revealed no significant differences in anxiety and shame between the
conditions. Similarly, no significant indirect effects were found using LMMs and multilevel mediation
analyses. No significant effects consistent with the hypotheses were found regarding the relationships
between performance emotions and interactive behavior either. Even in the combined model for anxiety
and shame, there were no significant effects on interaction behavior when the corrected significance
level was taken into account (anxiety: b = -0.370, p = .042; shame: b = 0.405, p = .029). Exploratory
analysis showed a significant difference between the experimental conditions. In the self-regulated
condition, the subjective learning effect was comparatively higher when prior knowledge was taken into
account (b = 0.338, t(97) = 3.47, p < .001, 95%-CI [0.145, 0.530]). However, no significant difference
was found between the conditions in terms of objective learning outcomes (t(32) = 0.13, p = .900, d =
0.022). An exploratory analysis also revealed that significantly more messages were written in the selfregulated
condition (t(32) = 3.33, p = .001)
Overall, the results do not provide sufficient empirical evidence for the postulated mediation
assumptions but they point to a possible relevance of the scaffolding condition for learning outcomes as
well as for emotional constructs and interactional aspects. Because of the small sample size, low
statistical power, and other methodological limitations, the results should be interpreted with caution.
Additional research with larger samples and standardized AI interactions is needed.