Responding to AI-generated emotional alerts: teachers' intervention and students' engagement in the mathematics classroom

Synthesis: This qualitative case study asks what happens when a vision-based system tells a teacher that a student is struggling emotionally. Swidan followed one teacher and eight high-achieving high school students as they worked on a geometry problem in GeoGebra while Dash4Emotion, an AI-generated emotional alert system, displayed red-framed rectangles over students it flagged as experiencing negative emotions. Video, system logs, and stimulus recall interviews yielded twenty identified episodes; five are presented. What shifted engagement was the teacher's response to an alert, not the alert itself: reframing the task reopened Mira's work, brief affirmation stabilized Dema's persistence, three simultaneous alerts prompted a whole-class reframing, repeated alerts for Rusol pushed the teacher from explanation toward inquiry, and in one period the teacher deliberately did nothing. The paper's contribution is conceptual: affective AI is treated as a mediating resource inside instructional decision-making rather than as an autonomous instructional agent. The practical lesson: an emotional alert is not an instruction but a cue to be interpreted alongside what the teacher can see and what the mathematics demands.

Key Findings

1. One classroom, twenty episodes, five patterns. One teacher and eight high-achieving high school students worked on a geometry task in GeoGebra, with alerts logged alongside video and recall interviews.
2. Reframing reopened a stalled student. After a red-framed alert for Mira, who twirled a pen and answered minimally, the teacher shifted her task to the geometric locus of minimal points.
3. Affirmation stabilized without reframing. An alert for Dema drew supportive words rather than a new task framing: her engagement structure held, while the affective tone shifted briefly toward reassurance.
4. Three simultaneous alerts prompted a collective move. With three simultaneous red-framed rectangles displayed, the teacher addressed the whole class, moving from a numerical outcome to the geometric characterization of point D.
5. Repeated alerts changed the teacher's strategy, not the student's. After continued alerts for Rusol, a second intervention asking "What are you looking for, Rusol?" surfaced her intention to find the angle bisectors' intersection points.
6. Non-intervention was a pedagogical decision. Later, with fewer negative and more positive alerts, the teacher said students were working with less stress; he did not intervene.

How the study was conducted

Students worked in GeoGebra while Dash4Emotion analyzed facial expressions and issued alerts: a red-framed rectangle carrying a student's name signaled a negative emotional state. Data included video recordings of classroom interactions, logs of the AI-generated emotional system, and stimulus recall interviews. Analysis was an iterative, episode-based process conducted by the author: each episode was read holistically, then successive passes assigned informed codes. Descriptive codes captured alert type and timing, visible actions, verbal responses, gestures, and GeoGebra interaction; interpretive codes came from the framework (affective representation, motivating desire, engagement process, teacher response). Robustness rested on repeated data engagement, triangulation across video, AI logs, and recall interviews, and comparison across all twenty episodes.

The five patterns of teacher response

The episodes make the teacher's interpretive work visible. For Mira, the alert arrived as she sat quietly, having stopped exploring. Initial questioning did not reorganize her engagement, but reframing the task from "finding a minimum value" to the geometric locus of minimal points did: she began searching for additional locations, later recalling that she had been confused and too shy to ask. For Dema, the alert marked cautious, hesitant dragging of point D. The teacher watched her actions without speaking about 15 seconds before asking whether she was observing how the sum of distances changed, then affirmed that she was decreasing the values slowly, thoughtfully. Three simultaneous alerts redirected the whole class from whether students had 13.04 or 13.05 toward a deeper question about the location of point D. Rusol's repeated alerts, one visible at minute 33:20, first drew an extended explanation in which the teacher took control of the software; only the second, inquiry-based return, after a 17-second observation, changed her trajectory.

Why emotional alerts are cues, not instructions

The paper's central argument is that the pedagogical significance of AI-generated emotional alerts lies in how they reshape what teachers observe and how they respond, not in their capacity to detect affect. The alerts act as mediating components of the classroom environment: they do not determine instructional action, but they supply informational cues that draw attention to moments that might otherwise remain unnoticed. Teachers already decide moment to moment whether to intervene, when to allow productive struggle, and how to respond to difficulty, and a cue from a system has to be judged rather than obeyed. Hence the risk the paper names: an alert read as an instruction, rather than as one input among several, would displace the interpretation the paper identifies as decisive.

What this means for practice

• Instructors. Treat an alert as a question, not an answer: the same system's alerts preceded very different teacher moves, and the move is what reshaped engagement.
• Instructors. When alerts repeat for a student after an explanation, ask rather than explain; "What are you looking for, Rusol?" surfaced the student's own mathematical intention.
• Administrators and designers. What alerts did was direct attention; the value came from interpretation alongside observable behavior and the task's demands, so protect the time that takes.
• Developers and researchers. Dash4Emotion supplied aggregate class-level affective data as well as individual alerts, and the teacher used the aggregate to justify restraint.

Limitations

• This is a single qualitative case study in one classroom: one teacher and eight high-achieving high school students on one geometry task, so its patterns are not generalizable.
• The analysis was conducted by the study's single author, who also did the two-stage episode coding; interpretive codes were inferred through triangulation, not independent coding.
• Five episodes are presented out of twenty identified episodes, and the slice text ends before the discussion and any stated conclusion.
• The slice reports no validation of the facial-expression affect detection, and notes that a sadness alert was not itself treated as evidence of withdrawal.

Citation

Swidan, O. (2026). Responding to AI-generated emotional alerts: teachers' intervention and students' engagement in the mathematics classroom. ZDM - Mathematics Education, 58, 929–943.