Human-Centered Design of LLM-Powered Educational Chatbots: A Study with Secondary Teachers

Synthesis: Reichert, Briceno, Tabarsi, and Barnes (2026) report a participatory design study in which six secondary teachers enrolled in a Research Experiences for Teachers summer program created low-fidelity paper prototypes of generative AI chatbots for their own classrooms. Through inductive qualitative analysis of a structured focus group and affinity diagramming of the prototypes, the authors show that teachers consistently conceptualized AI as a "bounded expert": a system with specialized capability confined to a strictly defined domain and operating under human supervision rather than as an autonomous agent. Boundedness operates along two dimensions — authority boundaries, in which teachers retain ultimate control because professional responsibility for student learning and safety cannot be delegated, and expertise boundaries, in which teachers see AI as lacking their contextual knowledge of individual students, classroom dynamics, and institutional norms. The teachers articulated a multi-layered safety architecture with four interconnected components — content scoping, content presentation, student adaptation, and teacher oversight — that operationalizes bounded authority through three protective layers: domain boundaries that restrict scope, content filtering that enables safe personalization, and teacher override that preserves human judgment in ambiguous cases. Mapping the designs onto Gagné's nine events of instruction revealed selective delegation: teachers welcomed AI assistance for presenting content, supplying practice problems, and offering formative Feedback, yet refused to delegate objective-setting or summative Assessment.

Key Findings

• Teachers designed "bounded experts," not autonomous agents: All six participants described AI that holds specialized capability within a strictly defined domain while remaining under human supervision. As P1 put it, the system should "free up our time to focus on students that need that extra help" without displacing pedagogical authority.
• Boundedness had two distinct dimensions: Authority boundaries — teachers keep ultimate control because professional and legal responsibility for student learning and safety cannot be delegated to a machine — and expertise boundaries — AI lacks teachers' contextual knowledge of individual students, behavioral histories, classroom dynamics, and institutional norms.
• Four interconnected components made up the bounded design (Fig. 1): content scoping (P1's chatbot was confined to Emperor Qin Shi Huang in an ancient China unit; P6's to Python variables, data structures, and functions), content presentation (visual aids in P4, structured tables in P1), student adaptation, and teacher oversight.
• Adaptation was deliberately bounded: P3's chatbot offered to present penguin facts as scary stories and adapted when a student named both scary stories and pop music, then solicited feedback; P6's let students rename themselves and the bot. Students could choose how to learn a topic but could not redirect conversation to unrelated content — personalization without loss of educational control.
• Three protective safety layers framed the architecture (Fig. 2): domain boundaries enforcing lesson-specific scope, including an "information quota" requiring a minimum number of facts or problems before progressing; content filtering with standardized refusals such as "Sorry, this is not part of my knowledge base" plus teacher alerts; and teacher override for ambiguous cases — P4 noted that questions on human reproduction could be legitimate in that unit and should route to a teacher rather than be auto-rejected.
• Transparency was behavioral rather than algorithmic: teachers prioritized visible limits and uncertainty cues over model explanations — P4's chatbot asked "Is the visual aid helpful?" while P5's math chatbot guided students to analyze "the skills in order to solve" problems without handing over answers.
• Oversight was framed as professional responsibility, not distrust of AI: teachers universally requested complete conversation logging, real-time alerts for inappropriate queries, and override capability, reading transcripts both for accuracy checking and for behavioral supervision of student-AI interactions.
• Delegation was selective across instruction: mapped onto Gagné's nine events, AI was welcomed for gaining attention, remediation, presenting content, Scaffolding, practice, corrective Feedback, and real-world transfer — but not for informing students of objectives or assessing performance, activities teachers treat as core professional judgment.

Study Design & Method

• Participants: six of twelve secondary teachers enrolled in a Research Experiences for Teachers summer program at a public university in the southeastern United States. Subjects spanned social studies, math, animal science, science, and computer science & robotics across grades 6–12; self-reported experience ranged from N/A to 16–20 years.
• Workshop context: all twelve teachers completed a three-day workshop series based on a publicly available, previously evaluated curriculum on LLMs, modified by the researchers to cover recent models such as GPT-4o and to integrate GPT-4o into in-workshop activities. Teachers were invited into the study on the fourth day; six agreed and completed a final project.
• Design session: participants worked in two groups of three but designed individually. They first completed a worksheet prompting reflection on the chatbot's content and scope, its "freedom," its role, its tone, intended student takeaways, and how it would differ from a conventional chatbot such as ChatGPT (~15 minutes), then built low-fidelity paper prototypes with note cards, pens, colored pencils, tape, and long sheets of paper (~25 minutes). Each prototype had to include at least one example student–chatbot conversation. Low-fidelity prototyping was chosen for accessibility, rapid iteration, and focus on conceptual rather than technical exploration.
• Focus group: a single one-hour structured focus group covered how the design process went, what was difficult, which parts were most and least important, how the chatbot would be used in class, its usefulness to the teacher, and desired new functionality. Two researchers facilitated; the session was recorded on Zoom and the auto-generated transcript was corrected and edited by one researcher.
• Analysis: two researchers performed inductive thematic analysis on the transcript in four sections, coding independently with private copies to avoid mutual influence, then reconciling tags by discussion and grouping them against the research questions. Final projects were analyzed with affinity diagramming on a digital whiteboard, with each researcher preparing individual insights on scope, tone, and appearance before grouping tags into themes. No inter-rater reliability was computed because only mutually agreed tags were carried forward.
• Limitations: a sample of six teachers limits generalizability; most participants taught STEM subjects, with only P1 representing the humanities; IRR was not calculated; and rapid Large Language Models (LLMs) change warrants replication, though the authors argue principles rooted in real classroom experience will persist.

What this means for practice

• Instructors. Scope an educational chatbot to lesson content before optimizing its accuracy: every teacher in the study made domain restriction a first requirement, treating it as an instructional need rather than a hallucination patch.
• Instructors. Keep objective-setting and summative Assessment under your own authority while delegating presentation, practice and formative feedback, the only instructional events these teachers were willing to hand over.
• Designers. Build teacher override and routed escalation rather than auto-rejection, so that ambiguous but legitimate questions reach a human instead of triggering a blanket refusal, and let teachers customize AI behavior — tone, purpose, scope, domain, modalities, presentation modes and conversational control — from a teacher-facing interface where they upload their own documents as prompting context. Expect the bounded-authority pattern to travel beyond education: the authors offer it as guidance for other high-stakes professions where AI augments rather than replaces expertise, such as diagnostic AI under physician authority or legal retrieval under lawyer judgment.
• Designers. Provide complete conversation logging with real-time alerts and a way for students and teachers to rate answer relevance, since teachers wanted transcripts for both accuracy checking and supervision of student–AI interactions. Support Multimodal AI, accessible presentation — speech input and multilingual input and output — for students with disabilities such as dyslexia and for multilingual students, an Accessibility and inclusion requirement rather than a feature upgrade.
• Designers. Prioritize behavioral transparency — visible limits, uncertainty cues, standardized refusals — over algorithmic explanation, which is what these teachers actually designed. Engage teachers as domain experts early and throughout design rather than as recipients of finished tools: participants valued being designers and urged tool builders to consult people "on the ground," making this usability work rather than outreach.

Limitations

• Six of the twelve teachers enrolled in a Research Experiences for Teachers summer program at one southeastern US public university agreed to take part, so all findings rest on that self-selected group.
• The systems studied were low-fidelity paper prototypes built from note cards and paper in about 25 minutes, so the results describe teachers' stated designs, not the behavior of deployed chatbots.
• No inter-rater reliability was calculated; the authors state that disagreements were resolved through discussion because of dataset size.
• Most participants taught STEM subjects, with only one of the six representing the humanities, and the authors note that rapid Large Language Models (LLMs) change warrants replication.

Citation

Reichert, H., Briceno, D., Tabarsi, B., & Barnes, T. (2026). Human-centered design of LLM-powered educational chatbots: A study with secondary teachers. In HCI International 2026: Human-Computer Interaction (Lecture Notes in Computer Science, pp. 516–535). Springer.