Scaffolding Students-AI Dialogue: A Framework for Safe Educational Interactions

Synthesis: Adolescents are the fastest and largest age group adopting large language models: 84% of Swiss 14–19-year-olds report regular generative AI use against 60% of 30–49-year-olds, yet these systems were never designed for their educational, emotional or developmental needs. Content filters can block harmful output but cannot make an interaction pedagogically sound, and prompts cannot reliably steer a probabilistic model across a multi-turn conversation. Muss, Leisten and Bardyn (Université de Neuchâtel, ETH Zurich and AI Swiss) answer with SCAFFOLD, the Steered Contextual AI Framework for Orchestrating Learning Dialogue: a model-agnostic, Open Source system of frames that surrounds generation with external verification, targeted repair and pre-approved safe fallback at every turn. The paper specifies how educators design a students-AI interaction, then the six-stage frame engine, then a classroom pilot with 12–16-year-olds; it is offered as shared infrastructure for the field so that pedagogical safety and learning design become auditable properties rather than vendor promises.

Key Findings

1. Adolescents dominate adoption while carrying the least design attention. 84% of Swiss 14–19-year-olds use generative AI regularly, the highest of any age group, on platforms central to their day.
2. Neither existing safety layer is enough alone. Filters block harmful output without making an interaction pedagogically sound, and pedagogical frameworks cannot steer an unpredictable model.
3. SCAFFOLD is a modular system of frames, not a better prompt. It sits between student and model, analyzing input, shaping prompts with safeguards, verifying output and falling back.
4. Six stages run every turn, and failure never falls through. Deterministic checks enforce code-writable rules such as turn counting; LLM-as-a-judge checks carry calibrated confidence; an exhausted repair budget returns a pre-approved safe fallback.
5. A classroom pilot ran the framework on a social robot. 24 analyzed students aged 12–16 in a rural German comprehensive school co-created a microcontroller mnemonic with the robot Marty.
6. Students became contributors, and co-creation predicted post-test knowledge. Off-topic turns fell from 38% to 1%, and co-creation level, consistent with retrieval practice, predicted post-test knowledge (b = 0.93, p = .011), unlike prior knowledge.

What SCAFFOLD is made of

A frame is a coded layer between the user and the Large Language Models (LLMs) that handles tasks and talks to a memory. Generation is surrounded by external verification, targeted repair and safe fallback, since instructions dilute as dialogue lengthens. Risk handling alerts a teacher to persistent misuse or dark thoughts and tracks engagement, progress and critical thinking. The pilot carried five frames, among them Balanced Turns and a Comprehension Tracker of per-student misconception profiles, on gpt-4.1-mini with text-to-speech on the social robot. Turn balance varied less (F = 5.84, p = .02), turn equity did not differ significantly, and comprehension Assessment failed.

The risks it is built to address

The risks are not merely factual: Privacy concerns and unsafe responses, plus attachment, cognitive atrophy and over-trust. For adolescents they are relational and developmental, so topic bans are insufficient: dependence, unsafe disclosure, misinformation and shrinking productive effort through cognitive offloading are one connected problem, answered by preserving the effort learning requires, with misuse and harm treated as systemic. Hallucination is structural, since LLMs produce likely continuations rather than verified truths, so verification splits between machine checks, guaranteed in formal domains like mathematics or code, and human verification for open tasks such as analysis, creative expression and ethical reasoning. Trust is a prerequisite, so monitoring must be transparent, benefit-limited and revocable.

Scaffolding the dialogue: mechanisms and implementation

SCAFFOLD starts from the learning activity, not the model. Educators define the context (setting, goals, interlocutors), the desired LLM behavior (pedagogical strategy, engagement), the background analysis and the learning material: enforceable requirements. When a student asks for an answer, should the model state it, ask for reasoning, or pose a metacognitive question? Leaving that to the model forfeits the Scaffolding and self-regulation benefits. Mechanically it gathers constitutional prompting, post-hoc validation and a non-regression correct–reverify loop. Memory is a knob, re-initialized each interaction or persisted for personalization, and AI judges need validation, while multi-agent evaluation may be too slow.

What teachers and designers are meant to do with it

The framework is a translation layer: appropriate interactions can be defined without coding expertise. Teachers bring the goals, context, strategy, content, flow, rules and assessment of a lesson plan, plus explicit desired model behavior, and consider students' zone of proximal development; the paper treats that work as productive in itself. The co-creation system is open-source with data, analysis code, a codebook and an interactive Simulation, seeding an ecosystem built with educators and child-development experts. Model-agnosticism allows scrutiny, and local deployment keeps data in schools. It also speaks to AI Regulation in Education: the EU AI Act classes educational AI as high-risk, demanding conformity assessment, human oversight and transparency.

What this means for practice

• Instructors. Decide before the lesson what the model does when a student asks for an answer, and specify setting, goals, behavior and analysis.
• Instructors. Budget more than ten minutes: the pilot's session ran out before groups reached the mnemonic practice phase.
• Learning designers. Type every check by reliability: enforce time limits and turn counts deterministically, reserve LLM-as-a-judge checks with calibrated confidence, validated first.
• Designers. Guarantee failure never reaches the student with a bounded correct–reverify loop and a pre-approved safe answer.

Limitations

• A position and methodology paper: one short exploratory pilot offering signal, not an effect size, and a correlational learning association; it also circulates as an MA thesis under another title (osf.io/dwp4t).
• The ten-minute session was too short to reach the third phase, all groups met the unframed model first without counterbalancing, and the LLM's own turns were ignored.
• The sample was small and homogeneous: 27 children from one rural German school, 24 analyzed after a data loss, 4 female; the probabilistic checks still need validation.

Citation

Muss, O., Leisten, L. M., & Bardyn, C. E. (2026). Scaffolding Students-AI Dialogue: A Framework for Safe Educational Interactions. PsyArXiv preprint.