The study in one page.
Keywords — artificial intelligence · human–machine interaction · educational technology · basic education
Dialogue, reloaded.
The arrival of generative artificial intelligence (GenAI) in the educational ecosystem is no longer hypothetical — it is a daily reality. From class planning to the creation of teaching materials, thousands of educators across Latin America interact every day with models like ChatGPT in a dialogical process with a non-human interlocutor.
This new dynamic transcends the mere adoption of a technological tool: it forces us to question the very nature of dialogue in education. Historically conceived as a pillar of collaborative knowledge construction and critical thinking, dialogue now faces a crossroads.
Can this interaction become a catalyst for a liberating education — or does it risk consolidating as an instrumental monologue that prizes efficiency over pedagogical reflection?
Dialogue as an object of study has evolved from a social practice into a rigorous multidisciplinary field. Its analysis has been characterized as a complex system articulated around four key components: structural elements regulating turns; participant dynamics, defining roles and power; intention, orienting conversation toward specific goals; and interactive features such as feedback and the repair of communicative breakdowns.
How teachers dialogue with AI is therefore a privileged window into their practice and professional agency. This study aims to characterize the nature of dialogue between teachers and GenAI: identifying frequent patterns in real conversations, establishing predominant use profiles via cluster analysis, and describing interaction constellations by purpose, nature and thematic domain.
Coding at scale.
To characterize teacher-chatbot dialogues we used a mixed-methods, predominantly quantitative, sequential exploratory design — combining expert validation, emergent qualitative analysis, and automated coding at scale. The process was articulated in two phases: developing a robust analytical framework, and applying it to the corpus.
Data & corpus
Data come from comenio.ai, a teacher-assistance platform with six chatbots. The sample was drawn from the chatbot focused on the New Mexican School (NEM) for basic education. The corpus comprised 7,340 complete, anonymized conversations stripped of any personal identifiers. The unit of analysis was the "complete conversation" — every exchange between a teacher and the AI from initial request to conclusion.
Phase 1 — Analytical framework
Rather than starting from a predefined framework, we used an iterative process. A multidisciplinary expert panel engaged in human-in-the-loop calibration: Gemini 2.5 Pro was used as a synthesis and conceptual probing tool, generating code-grouping proposals that experts then critically evaluated, refined and validated. Final categorical decisions always rested with the panel.
Phase 2 — Automated coding
Once validated, a Python tool used the OpenAI API (gpt-4o-2024-08-06) with Structured Outputs to strictly adhere to the defined Pydantic schema. The resulting dataset records the presence (True) or absence (False) of each code per conversation.
Analysis proceeded in four phases: (1) descriptive frequencies; (2) Phi (φ) correlation matrix; (3) profile identification with K-Means (K=3), grounded in the Elbow Method, Silhouette Coefficient and theoretical criteria; (4) dimensional reduction via PCA, complemented by discursive analysis of representative cases.
Fig. 02Representative cases for each profile — turns reconstructed from the corpus.
A spectrum, not a monolith.
Descriptive analysis shows a clear tilt toward instrumental uses. basic_content_generation (57.3%) is the most prevalent category, and 54% of dialogues are classified as teaching-centered. But a notable interest in constructivist methods is also present: 46.6% are learning-centered, and 31.3% involve complex activity design.
Notably, metacognition about the AI itself is nearly absent (0.5%). In practical use, teachers treat the AI mainly as a functional "black box," with no visible interest in discussing its internal workings during interaction.
Fig. 03Frequency of each code across the full corpus (n=7,340). The 19 framework dimensions, grouped by category.
Correlation analysis · φ matrix
The correlation map reveals a strong dichotomy: basic content generation is negatively correlated with complex activity design (φ = -0.51) and positively with teaching-centered focus (φ = 0.41). Conversely, complex activity design has its strongest positive correlation with learning-centered focus (φ = +0.60).
Regarding interaction nature, there is near-total mutual exclusion between transactional and dialogical interactions (φ = -0.87). Iterative dialogue is positively associated with personalization (φ = +0.50) and prompt scaffolding (φ = +0.43): longer dialogues are functionally tied to task sophistication, not chance.
Fig. 04φ correlation matrix between the 13 most relevant codes. Red = negative correlation; blue = positive. The diagonal highlights the near-total exclusion between transactional and dialogical interaction (φ = −0.87).
Fig. 05PCA: 19 dimensions reduced to two components — Pedagogical Sophistication (PC1) and Dialogue Elaboration (PC2). Variance explained: 44.12%.
| Variable | Instrumental–Trans. | Codiseñador Ped. | Reparación reactiva |
|---|---|---|---|
| generacion_contenido_basico | 0.70 | 0.15 | 0.86 |
| diseño_actividad_compleja | 0.04 | 0.93 | 0.10 |
| solicitud_integracion_curricular | 0.02 | 0.28 | 0.06 |
| desarrollo_proyecto_completo | 0.00 | 0.30 | 0.01 |
| adaptacion_personalizacion | 0.07 | 0.62 | 0.19 |
| evaluacion_retroalimentacion | 0.07 | 0.28 | 0.20 |
| exploracion_conceptual_creativa | 0.02 | 0.13 | 0.11 |
| centrado_en_ensenanza | 0.66 | 0.31 | 0.57 |
| centrado_en_aprendizaje | 0.16 | 0.97 | 0.51 |
| interaccion_transaccional | 0.97 | 0.35 | 0.00 |
| interaccion_iterativa_dialogica | 0.00 | 0.52 | 0.96 |
| andamiaje_del_prompt | 0.00 | 0.40 | 0.17 |
| solicitud_clarificacion_ia | 0.03 | 0.06 | 0.12 |
| feedback_pedagogico_ia | 0.06 | 0.06 | 0.05 |
| fallo_del_chatbot | 0.29 | 0.20 | 0.53 |
| contenido_curricular_especifico | 0.74 | 0.92 | 0.77 |
| planificacion_didactica_general | 0.15 | 0.68 | 0.32 |
| metacognicion_sobre_ia | 0.01 | 0.00 | 0.01 |
Tabla 01Centroids of the three clusters across the 18 framework variables. Background intensity encodes each variable's prevalence within the profile.
Instrumental, Co-designer & Repair.
Instrumental–Transactional
Short, direct requests. AI as a productivity assistant: factual content, lists, definitions.
Pedagogical Co-designer
Collaborative design of learning experiences. Adaptation, scaffolding and multi-phase projects.
Reactive Repair
Iterations to repair bot failures. Extended dialogue by necessity, not by strategy.
Fig. 01Three emergent teacher-use profiles identified via K-Means (K=3) over 7,340 conversations. Bar length shows code prevalence within each cluster.
Three ways of being-with AI.
Results reveal that teacher–AI dialogue is not monolithic but a spectrum oscillating between instrumental efficiency and complex pedagogical co-creation. The identification of three profiles does not merely quantify patterns; it offers a lens to interpret the tensions, challenges and opportunities GenAI introduces into teaching practice.
More than half of the corpus concentrates in the Instrumental mode — AI, first, is a productivity assistant.
The prevalence of the Instrumental–Transactional profile — over half of interactions — confirms that AI is adopted first as a productivity assistant. While this answers a real need to optimize tasks, it raises a pedagogical risk: from a Freirean critical reading, this pattern can be interpreted as a form of "banking education" where the AI becomes a new repository of knowledge from which the teacher simply extracts resources.
The Pedagogical Co-designer profile, by contrast, demonstrates the tool's transformative potential. Nearly a third of teachers use AI not to answer questions but to think with it — co-constructing learning experiences, adapting content, scaffolding entire projects. This profile embodies what we might call the "augmented teacher."
The third profile, Reactive Repair, is perhaps the most revealing about the technology's current state. Its iterativity is not strategic but reactive: a fifth of all interactions occur because the system failed and the teacher persisted. This evidences both the technical limitations of current chatbots and the professional agency of teachers who refuse to accept the first inadequate answer.
From users to designers.
This study characterized teacher–AI dialogue and identified a spectrum of practices from instrumental consultation to complex co-creation. The three profiles offer both a snapshot of the current state and a conceptual framework for understanding its pedagogical implications — highlighting that AI mirrors the teacher's intentionality and agency: the quality of the pedagogical product depends on the quality of the dialogue conducted.
From users to interaction designers: AI literacy as a new teacher competence.
Teachers may shift from a role of "users" to one of "interaction designers," developing AI-literacy skills such as prompt engineering. One path forward is the formation of communities of practice to share strategies, alongside developers building pedagogically informed chatbots capable of understanding teacher intent and supporting the repair of communicative breakdowns.
Methodologically, the paper contributes a validated, scalable analytical framework useful for research, AI tool design and teacher formative assessment. The resulting typology can guide professional development toward the Pedagogical Co-designer profile.
APA reference.
Medina-Gual, L., Bazaldua-Huerta, D. F., Narváez-Serrano, L. R., & Degetau-Arsuaga, D. (2026). El diálogo docente–IA: análisis de los perfiles de interacción con inteligencia artificial generativa. Revista Latinoamericana de Estudios Educativos, LVI(2), 163–188. https://doi.org/10.48102/rlee.2026.56.2.798