CoChat

The safety gap: restoring productive struggle through pedagogically aligned generative AI

Open access
Wang, Hong, Shan, Wenhui · 2026 · Frontiers in Education

The DOI matches its Crossref record; 15 of the 15 references it cites were found in a registry or an index; none of the 4 sources its claims cite is checked for support yet.

Abstract

The integration of Generative Artificial Intelligence (GenAI) into education presents a paradox: the more helpful the tool, the more it risks eroding the cognitive processes essential for deep learning. Educational research has long shown that meaningful learning depends not only on correct answers, but also on productive struggle, schema construction, and forms of scaffolding that preserve rather than bypass learner effort. This challenge is especially acute for novice learners, whose understanding is still developing and for whom premature answer delivery may replace the generative processes through which durable knowledge is formed. This paper introduces the concept of the “Safety Gap”—the divergence between a student's AI-assisted performance and their internal capability to verify that output, leading to potential epistemic risks when the tool fails. While GenAI offers unprecedented opportunities for personalized support and content generation, current Oracle-style models often eliminate the productive struggle necessary for schema construction. Drawing on recent literature, we argue that educational AI should be designed not merely to maximize convenience or answer completion, but to support cognitive engagement and independent judgment. We therefore propose a pedagogical framework centered on Socratic and Adversarial AI architectures that withhold direct solutions, introduce constructive cognitive friction, and prioritize process-based assessment over final outputs.

Overview

CoChat Summary

TL;DR

This perspective article introduces the "Safety Gap"—the divergence between a student's AI-assisted performance and their unassisted internal capability to verify that output—and argues that frictionless, Oracle-style generative AI erodes the productive struggle needed for schema construction. The authors propose a pedagogical framework (Restrict, Realign, Require, Refine) centered on Socratic and Adversarial AI architectures, process-based assessment, and preserved AI-free cognitive spaces.

Key findings

  • The Safety Gap is defined as the "Epistemic Discrepancy"—the inability of the user (Line A, independently reproducible reasoning without AI support) to verify the output of the tool (Line C, AI-assisted output quality); it is not high performance itself.
  • Current AI architectures are categorized into three types with differing risk: the Oracle (generic ChatGPT/Gemini/Grok) carries "High" risk because it eliminates struggle; the Socratic tutor (e.g., SocraticLM/Khanmigo) carries "Low" risk by supporting struggle; the Adversarial agent carries "Minimal" risk by building resilience (Table 1).
  • Cognitive offloading is not inherently detrimental: for advanced learners with established schemas, AI may reduce routine processing demands, but for novices the same convenience "may short-circuit the generative effort through which foundational competencies are built."
  • Automation bias and the "fluency heuristic" ("If it sounds this coherent, it must be true") lead to a surrender of epistemic agency, risking an "epistemic collapse where knowledge becomes a recursive loop of AI-generated content unchecked by human critical thought."

Methods

  • Design: a PERSPECTIVE article presenting a conceptual argument and proposed pedagogical framework, not an empirical study; the authors write from a medical education standpoint (Cancer Center, The First Hospital of Jilin University).
  • Data/sample: none; the paper draws on recent literature (e.g., Kapur's Productive Failure, Sweller's Cognitive Load Theory, Misiejuk et al.'s review of GenAI in learning analytics) and presents Figure 1 as "a conjectural conceptual model rather than an empirical plot."
  • Analysis: conceptual synthesis and framework proposal organized around four strategies—Restrict (AI-free cognitive sanctuaries, inverted scaffolding), Realign (Socratic withholding protocols), Require (process-based assessment via metacognitive reflection), and Refine (using LLM hallucinations as organic adversarial examples).

Limitations

  • Stated in the paper: the authors note that Misiejuk et al. "also note that evidence on the effectiveness of GenAI on actual learning is still lacking"; Figure 1 is explicitly labeled a conjectural conceptual model rather than empirical data.
  • Evident from the text: the proposed four-part framework (Restrict, Realign, Require, Refine) is not empirically tested or validated in this paper, and no outcome data, sample, or measurement of the Safety Gap is reported.
  • Evident from the text: the "Oracle" label is the authors' own analytic terminology "rather than as terminology taken from any single cited source," and the architectural risk ratings in Table 1 are asserted rather than empirically derived.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/feduc.2026.1757622

Cites

15 of 15 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 2 Not valid · 5 Not checked · 19 Own results and interpretations

5 more claims

Create a free account to see all 26 claims and to verify the 2 references still open.

See all claims — free account
  • Probably not a substantial claim. 12
  • It may state two separate claims as one. 9

Create a free account to see these claims.

Citations

Cites

Cited by

See all 2

Metadata

Authors
Wang, H., & Shan, W.
Affiliations
First Hospital of Jilin University (CN)
Corresponding author
Wang, Hong
Venue
Frontiers in Education
Published
2026-04-22
Volume
11
ISSN
2504-284X (ISSN-L)
OpenAlex
W7155211039
Cited by
2
References
14
Citations by year
2026: 2
FWCI
16.52
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Education (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Jilin University (2023XZD098, 20250PSFX6P002Y)
  • Division of Undergraduate Education
Topics
  • Social Sciences › Education › Artificial Intelligence in Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Intelligent Tutoring Systems and Adaptive Learning
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
generative AI in education, productive struggle, safety gaps, schema construction, scaffolding, novice learners, epistemic risk, adversarial AI, process-based assessment, cognitive engagement, deep learning, AI-assisted learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.