PaperlayerView exact revision on arXiv

Source-linked AI summary

Addressing Trust in AI Systems through Education: A Didactic Perspective

Pierre Haritz, Hendrik Krone, Thomas Liebig

arXiv:2609.02453v1cs.CYcs.AI

TL;DR

ML education faces a gap in teaching students to understand black-box systems and form calibrated trust. The paper presents ICE-T, which integrates intermodal transfer, Use-Modify-Create computational thinking, and explanatory thinking, and argues that existing practice rarely supports these facets together. It concludes that trust calibration should be an explicit educational objective, while noting that ICE-T’s connection to trust calibration remains theoretical and unevaluated.

  • Problem

    ML education remains challenged by opaque tools that provide superficial understanding, while teaching ML to novices in a way that treats trust as a primary concern remains unresolved.

  • Method

    The paper develops ICE-T by integrating intermodal transfer, Use-Modify-Create computational thinking, explanatory thinking, and a process model, then connects the framework to trust and K-12 activity literature.

  • Results

    The paper argues that ICE-T’s three facets supply representational richness, graduated process control, and error contextualization, while reviews show these mechanisms are rarely integrated in current practice.

  • Takeaways & Limitations

    Trust calibration should be treated as an explicit educational objective of ML education.

  • Takeaways & Limitations

    The ICE-T–trust-calibration connection is theoretical, has not itself been evaluated, and is empirically grounded only in the K-12 landscape.

Abstract

from arXiv · show

Machine learning (ML) education faces two persistent and connected obstacles: many educational tools present ML as an opaque black box, which leaves learners with a superficial understanding, and this same opacity prevents users from forming the calibrated trust that appropriate reliance on AI systems requires. We present ICE-T, a didactic framework that integrates three mutually reinforcing facets: intermodal transfer grounded in Bruner's enactive, iconic, and symbolic modes of representation, computational thinking operationalized through the Use-Modify-Create progression, and explanatory thinking supported by a process model. Connecting the framework to the empirical literature on algorithm aversion, AI literacy, and mental model formation, and to systematic reviews of the K-12 ML activity landscape, we argue that the three facets supply the cognitive mechanisms that the trust calibration literature identifies as drivers of appropriate reliance: representational richness, graduated process control, and the capacity to contextualize errors. On this basis, we propose that trust calibration be treated as an explicit educational objective, with ICE-T as a principled and scalable means of achieving it.

1 Introduction

ML education is increasingly important, but teaching students with little programming experience while treating trust as a primary concern remains unresolved. Existing tools often expose only surface-level interactions, and ICE-T integrates three didactic facets while arguing for trust calibration as an explicit objective.

  • Motivation: AI and ML education is urged across educational levels because these technologies increasingly mediate decisions in everyday domains.The paper links this need to understanding, critically evaluating, and shaping AI systems.
  • Research problem: Teaching ML effectively to students with little or no programming experience while treating trust as a primary concern remains an open research problem.
  • Research problem: Many educational tools support surface-level model interaction without developing understanding of data pipelines, algorithmic mechanisms, or evaluation criteria.This black-box treatment risks leaving students with a superficial and potentially misleading understanding of ML systems.
  • Contributions: ICE-T integrates intermodal transfer, computational thinking, and explanatory thinking as a multi-faceted framework for teaching ML.
  • Contributions: The paper connects ICE-T to trust-calibration literature and systematic reviews, arguing that trust calibration should become an explicit educational objective.It also argues that current K-12 practice systematically undersupports the mechanisms needed for appropriate reliance.

2 Preliminaries on Trust

Trust in AI concerns willingness to accept vulnerability under uncertainty, but productive use requires confidence calibrated to actual capabilities and limitations. The literature distinguishes overtrust and undertrust and frames trust formation as dependent on the system, the user, and their context.

  • Trust calibration: Overtrust accepts systems that do not warrant confidence, whereas undertrust rejects systems that are safe and reliable.Both extremes can prevent socially beneficial use of AI.
  • Trust calibration: Calibrated trust is the state in which users’ confidence tracks an AI system’s actual capabilities and limitations.
  • Trust formation: Trust formation depends on characteristics of the AI system, the human user, and the context in which they meet.
  • Trust formation: Technical safety alone does not produce calibrated trust, because poorly understood systems can prompt blind acceptance or blanket rejection.Conversely, communication about an unsafe system can produce unjustified confidence.
  • Definitions: AI trust is defined as a user’s willingness to be vulnerable to an automated or intelligent system under uncertainty.
  • Definitions: Trust research distinguishes cognitive trust based on rational appraisal from affective trust based on emotional responses, while also separating subjective trust from system trustworthiness.

3 Theoretical Foundations of Didactics

ICE-T draws on complementary educational theories to make ML learning more representationally rich, progressively autonomous, and explanatory. Its facets address how learners represent concepts, engage with computational artifacts, and reason about system behavior.

  • Intermodal transfer: Bruner’s enactive, iconic, and symbolic modes support learning through action, visual representation, and increasingly abstract forms of understanding.
  • Intermodal transfer: Intermodal transfer deliberately moves learners between multiple representations of the same concept to strengthen conceptual understanding and retention.The modes coexist rather than forming rigidly sequential stages.
  • Computational thinking: Computational thinking includes decomposition, pattern recognition, abstraction, and algorithm design for formulating computationally tractable problems and solutions.
  • Computational thinking: The Use-Modify-Create progression scaffolds increasing student autonomy and creative ownership in computational activities.It begins with pre-built artifacts and progresses toward modification and novel creation.
  • Explanatory thinking: Explanatory thinking requires coherent accounts of how and why ML systems behave, including the roles of training data, algorithms, metrics, and error or bias.
  • Explanatory thinking: Black-box tools can engage learners and demonstrate practical utility but fall short of enabling deeper critical reasoning about AI systems.

4 Evaluation of Existing Educational Tools and Platforms

Existing ML education resources form a heterogeneous landscape in which tools usually support one or two didactic dimensions rather than integrating all three. Larger reviews corroborate that supervised-learning activities remain concentrated at low abstraction levels and rarely combine technical and societal perspectives.

  • Evaluation approach: The tool analysis evaluated educational resources against intermodal transfer, computational thinking through UMC, and explanatory thinking.
  • Findings: Individual tools typically excel in one or two didactic dimensions but rarely integrate intermodal transfer, computational thinking, and explanatory thinking.
  • Findings: Visual programming platforms support structured computational activities but often provide limited intermodal transfer and explanatory engagement with ML mechanisms.
  • Findings: Web-based training tools provide compelling enactive and iconic experiences but typically do not expose algorithmic details needed for explanatory thinking.
  • Systematic-review evidence: Almost 60% of coded supervised-learning entries in a 126-activity review fell within invisible use, button interaction, or model deployment.Open-ended creation tasks appeared only twice in the corpus.
  • Systematic-review evidence: Technical and societal perspectives on the same algorithm co-occurred in only about 3% of activities in a complementary review of 133 activities.

5 PETSP-ML: A Process Model for Explanatory Thinking

PETSP-ML adapts the CRISP-DM workflow into a four-stage classroom process that prioritizes explanatory understanding before model training. Its distinct Model Understanding stage builds conceptual foundations and links data, algorithmic choices, evaluation, and outcomes.

  • 5 PETSP-ML: A Process Model for Explanatory Thinking: PETSP-ML adapts CRISP-DM into four classroom stages aligned with the cognitive and didactic requirements of ML education.The stages are Task Understanding, Data Phase, Model Understanding, and Model Training and Evaluation.
  • 5 PETSP-ML: A Process Model for Explanatory Thinking: Students examine task goals, data quality and bias, algorithmic principles, and model performance across the four stages.The process connects task suitability, data representativeness, labeling accuracy, model limitations, metrics, and observed outcomes.
  • 5 PETSP-ML: A Process Model for Explanatory Thinking: Model Understanding precedes training so students learn how the algorithm makes decisions, where it is limited, and whether it fits the intended goals.This ordering deliberately prioritizes conceptual foundations before procedural model-training skills.

6 The ICE-T Concept: Integrating the Three Facets

ICE-T unifies intermodal transfer, computational thinking, and explanatory thinking into one ML-learning framework. Its facets structure representation, learner autonomy, and instructional sequence as complementary dimensions that prompt students to explain and justify their reasoning.

  • 6 The ICE-T Concept: Integrating the Three Facets: ICE-T combines Intermodal Transfer, Computational Thinking, and Explanatory Thinking through EIS, UMC, and PETSP-ML, respectively.The three components are presented as mutually reinforcing facets of the framework.
  • 6 The ICE-T Concept: Integrating the Three Facets: A learning unit can move from hands-on classification to visual representation and formal ML reasoning while following the PETSP-ML process.The example uses decision-tree classification of animals and integrates multiple representational forms with the process model.
  • 6 The ICE-T Concept: Integrating the Three Facets: EIS structures representation, UMC structures student autonomy, and PETSP-ML structures temporal sequence without requiring every possible combination.Educators deliberately decide which combinations of modes, autonomy stages, and process phases to realize.
  • 6 The ICE-T Concept: Integrating the Three Facets: The integrated approach gives students multiple cognitive perspectives, procedural and conceptual knowledge, and repeated prompts to explain and justify their reasoning.

7 From Understanding to Trust

Trust in AI is both necessary for productive adoption and risky when poorly calibrated. The reviewed literature links calibrated reliance to accurate mental models, domain-specific understanding, and literacy that makes users more discerning rather than simply more trusting.

  • 7 From Understanding to Trust: Trust supports productive AI adoption but can produce harmful over-reliance when it is poorly calibrated.
  • 7 From Understanding to Trust: Calibrated trust requires users to understand what an AI system can do, when it succeeds, and where its failure modes lie.Without this functional mental model, trust responses tend to rely on heuristics, prior attitudes, or superficial cues.
  • 7 From Understanding to Trust: Literacy interventions can recalibrate reliance by encouraging users to trust models when they perform well and withhold trust when outputs are uncertain or flawed.The cited study reports increased discernment rather than simply increased trust.
  • 7 From Understanding to Trust: Familiarity and trust show a non-linear relationship: novices may reject algorithms, moderately literate users show more calibrated profiles, and highly familiar users may over-rely.The reported pattern is an inverted-U relationship rather than a simple increase in trust with familiarity.
  • 7 From Understanding to Trust: Domain-specific explanations improve understandability and willingness to act more than generic explanations, while teacher AI understanding and self-efficacy are associated with more calibrated trust.

7.2 The Black Box as a Shared Obstacle to Understanding and Trust

The paper identifies opaque black-box treatment as a shared obstacle to deep ML understanding and calibrated trust. It therefore connects graduated process control and explanatory learning to the construction of functional mental models that can contextualize errors.

  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: Treating ML as an opaque black box prevents students from building conceptual knowledge and users from building mental models for calibrated reliance.
  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: People may reject an algorithm after one error when they lack a mental model that contextualizes the error as a consequence of its design.
  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: Giving users slight control to modify algorithmic predictions substantially reduces algorithm aversion.
  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: Educational activities that let students enact algorithms, modify parameters, and observe downstream effects can build functional mental models for understanding and calibrated trust.
  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: A review of 126 activities asks whether current K-12 practice provides the process control and explanatory depth required by UMC and PETSP-ML.
  • 7.2 The Black Box as a Shared Obstacle to Understanding and Trust: Of 89 supervised-learning activities, 69 assigned only the User role, compared with 17 reaching Creator and 8 reaching Machine.The paper interprets this imbalance as leaving trust-calibrating process-control potential largely untapped.

The Critical-Technical Gap.

The paper identifies a critical-technical gap in ML education: activities rarely connect algorithmic mechanisms with societal consequences. ICE-T addresses this gap by combining explanatory thinking with intermodal representations and graduated learner control to support calibrated reliance.

  • The Critical-Technical Gap.: Across 133 activities, technical and societal perspectives on the same algorithm co-occur in roughly 3% of cases.This separation limits opportunities to connect model mechanisms with their consequences.
  • The Critical-Technical Gap.: Learners can trace biased or unfair outcomes to training data and evaluation choices only when mechanisms and consequences are learned together.Current practice typically examines either the mechanism or its consequences, rather than both.
  • The Critical-Technical Gap.: ICE-T responds by integrating intermodal transfer, computational thinking, and explanatory thinking as complementary facets.The framework links these facets to cognitive mechanisms associated with appropriate reliance.
  • The Critical-Technical Gap.: Intermodal transfer uses multiple representations to support richer and more flexible mental models of ML concepts.The facet is intended to deepen understanding beyond exposure to a single representation.
  • The Critical-Technical Gap.: The UMC progression increases learner agency from observing systems to modifying and creating algorithmic processes.The paper presents the resulting increase in calibrated reliance as a testable hypothesis, not an established result.

Explanatory Thinking and Error Contextualization.

Explanatory thinking supports trust calibration by enabling learners to interpret model errors as consequences of identifiable data, algorithmic, and evaluative choices. Together with intermodal transfer and UMC, it contributes to a mental model that supports context-sensitive reliance and informed skepticism.

  • Explanatory Thinking and Error Contextualization.: Explanatory thinking requires students to construct causal explanations of system behavior rather than merely observe outputs.PETSP-ML connects outputs to training-data properties, algorithmic partitions, and evaluation metrics.
  • Explanatory Thinking and Error Contextualization.: Contextualizing errors turns an apparently opaque failure into a consequence of identifiable design choices and data limitations.This capacity is presented as the mechanism through which explanatory thinking supports calibrated trust.
  • Explanatory Thinking and Error Contextualization.: The three facets jointly provide representational understanding, control over design choices, and the ability to assess likely failure modes.These components form the mental model that calibrated trust requires.
  • Explanatory Thinking and Error Contextualization.: ICE-T supports context-sensitive trust decisions by helping students rely on systems under appropriate conditions and remain skeptical when design assumptions do not fit.The framework links this stance to informed skepticism rather than indiscriminate acceptance.
  • Explanatory Thinking and Error Contextualization.: The framework treats trust calibration as an explicit educational objective alongside conceptual understanding and procedural skill.Its three facets are presented as an integrated mechanism rather than an add-on.

8 Conclusion

The conclusion presents ICE-T as an integrated response to pedagogical and trust-calibration challenges in ML education. It argues that the framework’s three facets address cognitive requirements for calibrated trust, while acknowledging that this connection remains theoretical and unevaluated.

  • 8 Conclusion: ICE-T combines Bruner-based intermodal transfer, the Use-Modify-Create progression, and PETSP-ML-supported explanatory thinking.The framework integrates these components to address both ML pedagogy and trust calibration.
  • 8 Conclusion: Systematic reviews of 126 and 133 K-12 ML activities identified black-box treatment, passive learner roles, and separation of technical and societal perspectives as major gaps.These findings provide the empirical landscape that motivates the framework.
  • 8 Conclusion: The paper argues that ICE-T supplies representational richness, graduated process control, and error contextualization as cognitive mechanisms required for calibrated trust.The conclusion connects these mechanisms to evidence on algorithm aversion and AI literacy.
  • 8 Conclusion: The paper proposes treating trust calibration as an explicit educational objective of ML education.This proposal follows from the framework’s account of how learners develop appropriate reliance.
  • 8 Conclusion: The connection between ICE-T and trust calibration remains theoretical because ICE-T has not itself been evaluated and its empirical grounding is limited to K-12.Future work should test classroom learning units, the predicted UMC reliance gradient, and transfer across ages and ML paradigms.
Loading 2609.02453v1…