From Substitution to Scaffolding: Breaking the Self-Reinforcing Harm Cycle of AI in Education (and Beyond)

Synthesis: Favero, Pérez-Ortiz, Käser, and Oliver (2026) argue that the central risk of AI in education is not technological failure but misalignment — AI that substitutes for human effort erodes the very capacities education is meant to build. They organize this risk into an integrative framework of four interrelated dimensions — cognition, Learner Agency, emotional Well-Being, and Ethics — linked by a self-reinforcing harm cycle in which offloading reduces effort, weakens agency, and compounds emotional and ethical harm. Grounding the framework in an exploratory analysis of 49 International Baccalaureate essays, they find learners themselves perceive these risks (80% report AI reliance reduces thinking) while converging on the AI they want: systems that withhold immediate answers, prompt recall, and encourage reflection through questions. From this they derive a single design principle, scaffold, do not substitute, argued to extend beyond education to any system that mediates human thinking, and outline a research agenda for building AI that fosters enduring human capacity.

The self-reinforcing harm cycle

The paper's central conceptual contribution is a unifying mechanism that treats the scattered harms of AI — usually discussed in isolation — as one interconnected, self-reinforcing cycle:

• Cognition: Cognitive offloading reduces the effortful reasoning that builds durable understanding; large-scale studies link high AI dependence to lower performance on Critical Thinking assessments, mediated by offloading. This connects directly to the knowledge base's over-reliance literature (see Cognitive Offloading and the associated articles on the speedup illusion and synthesis writing).
• Agency: convenience and persuasive outputs erode Learner Agency; dependency undermines independent Problem Solving, and overtrust leads students to accept AI output uncritically. The authors warn of intellectual conformity — ready-made answers nudging learners toward normative reasoning and standardized thought.
• Emotion: reliance produces technostress, digital fatigue, AI guilt (feeling both assisted and inauthentic), and — for some — AI entitlement, a normative breakdown around effort and authorship.
• Ethics: continuous data collection creates surveillance and power asymmetries; students' fear of being wrong suppresses experimentation and intellectual risk-taking. Notably, students in the corpus raised ethics mainly as authorship and integrity ("how can the school know they are originals and not just AI generated?") rather than Privacy.

The cycle is self-reinforcing because each dimension feeds the next: cognitive offloading weakens effortful reasoning, which increases overtrust and reduces agency; diminished agency intensifies emotional harms; and structural ethical factors (surveillance, opaque data practices) amplify all of these, leaving learners less confident, less critical, and increasingly dependent on algorithmic authority.

The students' own words

The framework is grounded in an exploratory qualitative analysis of 49 argumentative essays written by International Baccalaureate students (median age 17) from three German-speaking Swiss schools, responding to "Does using AI change what it means to learn?" Findings are descriptive, not generalizable:

• 80% (39/49) explicitly linked AI reliance to reduced thinking — "the knowledge arrives without being earned, and unearthed knowledge tends not to stay."
• 41% endorsed effortful retrieval or effort as valuable ("retrieving information from memory, even imperfectly, strengthens understanding").
• 88% (43/49) raised agency, most often as erosion of independent effort.
• 53% (26/49) named at least one behavior that supports rather than replaces thinking — withholding the solution, asking critical questions, prompting active recall, offering alternative explanations.
• Yet 48 of 49 framed AI in terms of replacement, not scaffolding — the AI students say they need is not the AI they typically encounter.

This convergence is striking because students' desiderata align almost exactly with established learning-science principles: effortful retrieval, productive struggle, delayed Feedback, and questioning. The authors note these are "not consumer preferences but pedagogically sound intuitions."

Scaffold, do not substitute

The single design principle that emerges — scaffold, do not substitute — positions Scaffolding as a first-class capability for AI systems: knowing when to withhold an answer, ask a question, surface uncertainty, or present alternative perspectives. The paper cites Maike, a Privacy-preserving, environmentally sensitive educational chatbot that guides learners through critical questioning and self-reflection via the Socratic method, as an early illustration.

For different stakeholders, the principle translates into distinct obligations:

• Researchers: move beyond benchmarks of what AI can produce toward measures of what AI helps humans develop — understanding, judgment, agency, and independent thinking. The central question becomes whether systems strengthen human capacity over time or quietly replace the effort through which that capacity is built.
• System builders: scaffolding must become a designed capability, not an afterthought.
• educators, institutions, policymakers: preserve the conditions under which learning occurs — productive struggle, reflection, intellectual agency — through tasks, assessments, procurement, and AI Governance structures.

The paper's broader claim is that this challenge extends beyond education to the entire AI ecosystem: any system that mediates human thinking can either weaken human capabilities through substitution or strengthen them through scaffolding, with consequences not only for learners but for democratic societies.

Relationship to existing research

The framework integrates and extends themes already present in the knowledge base. It provides a unifying theoretical lens that connects the over-reliance / Cognitive Offloading literature, the learner-agency and dependency literature, Reducing AI Misuse, and AI ethics in education — synthesizing what prior work has tended to treat as separate concerns. Its "scaffold, do not substitute" principle operationalizes the Pedagogies and Teaching Strategies-vs-technology distinction that underlies work on pedagogically aligned AI, and its call to measure human capacity development rather than machine output aligns with the knowledge base's critiques of overly narrow AI benchmarks. Its evidence base (student essays) adds a rare learner-perspective dimension to a literature dominated by researcher and system perspectives.

What this means for practice

• Instructors. Require an unaided first attempt before any AI use on a task: 48 of 49 essays framed AI in terms of replacement rather than scaffolding and 80% (39/49) linked reliance to reduced thinking, so the pre-AI draft is where the effort that produces learning happens.
• Instructors. Configure the tools you assign to scaffold rather than solve — withholding the answer, prompting recall, asking critical questions, offering alternative explanations — since 53% (26/49) of students named exactly those behaviors as the AI they want, and their preferences match retrieval practice and delayed Feedback.
• Learners. Ask AI for questions instead of answers and treat an unearned answer as a warning sign: 41% of the essays endorsed effortful retrieval as valuable ("retrieving information from memory, even imperfectly, strengthens understanding").
• Designers. Build Scaffolding as a designed capability with explicit rules for when to withhold an answer, surface uncertainty, or present alternative perspectives, rather than layering a policy on top of a system that answers on demand.
• Administrators. Write the conditions for productive struggle, reflection, and intellectual agency into procurement and Assessment policy, because substitution risk is set by task design and tool choice, not by access alone.

Limitations

• The empirical basis is 49 argumentative essays from three German-speaking Swiss schools (n = 19, 11, 19; median age 17) answering a single prompt; the authors call the cohort "illustrative rather than representative" and make no claims of statistical generalizability.
• Evidence is descriptive only: the analysis codes students' first, unaided drafts with a deductive codebook and reports share-of-essay counts (80%, 41%, 88%, 53% of 49), so there is no control condition, intervention, or outcome measure of learning.
• The framework itself is untested: the four dimensions and their self-reinforcing cycle are a conceptual synthesis, with the figure marking the closing link to consequences beyond education as weakly evidenced (dashed) and grounded only in the 49-student exploratory analysis.
• The "scaffold, do not substitute" principle is argued from one cited illustration (Maike) rather than from evaluated interventions, so the paper supports a design agenda and research questions, not demonstrated effects.

Citation

Favero, L., Pérez-Ortiz, J. A., Käser, T., & Oliver, N. (2026). From Substitution to Scaffolding: Breaking the Self-Reinforcing Harm Cycle of AI in Education (and Beyond). [cs.HC]. https://doi.org/10.48550/arXiv.2608.17451