Research Article
Addressing Trust in AI Systems through Education: A Didactic Perspective
Synthesis: Haritz, Krone, and Liebig (2026) argue that two problems in machine-learning education are actually one: educational tools that present ML as an opaque black box leave learners with superficial understanding, and that same opacity prevents them from forming the calibrated trust that appropriate reliance on AI demands. They propose ICE-T, a didactic framework uniting intermodal transfer (grounded in Bruner's enactive, iconic, and symbolic modes), Computational Thinking (operationalized through Use-Modify-Create), and explanatory thinking (supported by a process model). Connecting these facets to research on algorithm aversion, AI Literacy, and trust formation, they argue that ICE-T supplies the cognitive mechanisms the trust-calibration literature identifies as drivers of appropriate reliance — representational richness, graduated process control, and the capacity to contextualize errors — and that trust calibration should therefore be treated as an explicit objective of ML education.
Opacity Breeds Both Misunderstanding and Miscalibrated Trust
The paper starts from two connected obstacles in ML education. Pedagogically, many tools let students interact with models at a surface level — for instance, training an image classifier through a web interface — without exposing the data pipelines, algorithmic mechanisms, or evaluation criteria underneath. This black-box treatment risks leaving learners with a superficial and potentially misleading picture of what ML systems actually do. Consequentially, those same learners cannot build trust that is calibrated to a system's real capabilities and limits: they veer toward either blindly accepting the output or being overly dismissive. As the paper notes, the EU AI Act defines AI literacy precisely as the skills and knowledge to make informed deployment decisions and stay aware of opportunities and risks — which requires understanding how a system works in order to know when to rely on it and when to be skeptical.
The ICE-T Framework
ICE-T integrates three mutually reinforcing facets into a single didactic framework for teaching ML:
- Intermodal transfer — grounded in Bruner's enactive, iconic, and symbolic modes of representation, giving learners multiple, complementary ways to grasp a model rather than a single opaque route.
- Computational thinking — operationalized through the Use-Modify-Create progression, so learners move from using a working system to modifying it and finally to creating their own.
- Explanatory thinking — supported by a process model (PETSP-ML) that lets learners walk through how a prediction or system behavior comes about.
The paper connects these facets to the trust-calibration literature and to systematic reviews of the K-12 ML activity landscape, arguing each facet supplies a mechanism the literature identifies as driving appropriate reliance: representational richness, graduated process control (the ability to exercise increasing agency over the system), and the capacity to contextualize errors rather than treat them as random or authoritative.
Treating Trust Calibration as an Educational Objective
The central proposal is normative: trust calibration should be an explicit, assessable objective of ML and AI education, not an incidental byproduct of technical fluency. Drawing on reviews of existing K-12 ML activities, the authors contend these three mechanisms are systematically undersupported in current practice — tools teach surface interaction far more than they build graduated control or error-contextualization. Because curricula at the school level increasingly carry the burden of AI literacy, making calibration an explicit goal offers a principled and scalable route to appropriate reliance on AI across healthcare, transport, finance, and other consequential domains. For educators and curriculum designers, ICE-T supplies a coherent sequence for teaching ML that keeps trust — rather than mere tool use — as the organizing educational aim.
Connected Concepts
- AI Literacy
- Trust Calibration
- Computational Thinking
- K-12
- Curriculum Design
- Pedagogies and Teaching Strategies
- AI in Education
Connected Articles
- Trust and Reliance on AI in Education: AI Literacy and Need for Cognition as Moderators — Trust and Reliance on AI in Education: AI Literacy and Need for Cognition as Moderators
- Trust-utility gap in introductory physics education: Students' adoption, domain-specific skepticism, and preferences for AI integration — Trust-Utility Gap in Introductory Physics Education
- AI Literacy Interventions in Education: A Meta-Analysis of Effects and Moderators — AI Literacy Interventions in Education: A Meta-Analysis
- A Case Study of Integrating AI Literacy Education in a Biology Class — A Case Study of Integrating AI Literacy Education in a Biology Class
- Using LLMs to Detect Growth in Computational Thinking in Introductory Physics — Using LLMs to Detect Growth in Computational Thinking in Introductory Physics
Citation
Haritz, P., Krone, H., & Liebig, T. (2026). Addressing Trust in AI Systems through Education: A Didactic Perspective. arXiv preprint arXiv:2609.02453.