Educational AI Auditing Across the Learning Lifespan: Concepts, Methods, Evidence, and Stage-Sensitive Governance

Educational AI Auditing Across the Learning Lifespan: Concepts, Methods, Evidence, and Stage-Sensitive Governance: Comparison

View Latest Version

Please note this is a comparison between Version 2 by Vicky Zhou and Version 1 by Evelyn Wu.

Educational AI auditing is the systematic, evidence-based examination of artificial intelligence systems used in formal, non-formal, and informal education to determine whether they are technically reliable, pedagogically valid, developmentally appropriate, equitable, accessible, safe, and institutionally accountable. Its object is not only a model but the situated arrangement through which models, data, interfaces, people, policies, and organizational routines redistribute educational opportunities, judgments, labor, and risk. Unlike benchmarking, an audit does not stop at task performance; unlike prospective impact assessment, it tests claims with empirical evidence; and unlike compliance review, it can ask whether a lawful use is educationally defensible. A stage-sensitive audit asks what educational work is delegated to AI, whether that delegation expands or displaces learners’ and educators’ capabilities, how effects differ across contexts and social positions, and who can contest or remedy harmful outcomes. The framework presented here combines four system levels, seven audit domains, five lifecycle gates, and a burden of evidence that rises with decision stakes, opacity, developmental dependence, and the persistence or irreversibility of consequences.

educational AI auditing
generative artificial intelligence
algorithmic auditing
educational technology
human development
educational equity
AI governance
institutional accountability

Artificial intelligence in education is often discussed as a contest between promise and peril. Adaptive tutors can personalize practice; generative systems can produce explanations, feedback, simulations, translations, lesson plans, and administrative summaries; predictive models can identify learners who may need support. Meta-analyses report positive average effects for some AI-supported interventions, but estimates differ substantially and heterogeneity routinely exceeds 90% [1,2,3]^[1][2][3]. The syntheses also disagree about which educational levels benefit most. The relevant policy question is therefore not whether AI works in general, but which configured system works, for whom, for what educational purpose, under what conditions, compared with what alternative, for how long, and at whose risk.

Model benchmarks cannot answer those questions. A fluent explanation may be factually wrong, instructionally mistimed, culturally narrow, inaccessible, or so complete that it removes productive struggle. A risk score may be calibrated yet stigmatize students or trigger an intervention without meaningful appeal. Evidence on AI-text detection illustrates the importance of testing the deployed configuration: one evaluation found high false-positive rates for essays written by non-native English writers ^[4], whereas an assessment-specific detector trained on well-sampled data reported high accuracy without evidence of such disadvantage ^[5]. Neither finding licenses a general claim about all detectors. Fairness depends on the model, training sample, task, threshold, population, and consequences of error.

Educational AI auditing should therefore be treated as a distinct field of inquiry and practice. Its unit is an educationally situated sociotechnical system: a model or algorithm configured for a population, task, interface, workflow, and governance regime. Its proposed criterion is capability formation, assessed by whether the arrangement helps people acquire, exercise, retain, and transfer valued capabilities, rather than by the quality of an immediate output alone. Its stage sensitivity recognizes that the same feature can have different consequences in early childhood, compulsory schooling, technical and professional education, higher education, and adult or community learning.

This Entry synthesizes established audit traditions and recent educational evidence into three intersecting views: educational stages and participation contexts; system levels from model to institutional ecosystem; and the lifecycle from problem formulation to retirement. Seven domains cut across these views: pedagogical validity; developmental appropriateness and capability formation; epistemic reliability; equity, accessibility, and cultural-linguistic justice; privacy, safety, and well-being; human agency, relational quality, and labor; and transparency, contestability, and accountability. The contribution is not the claim that sociotechnical or lifecycle auditing is new. It is an education-specific navigation rule that links the purpose and risk of a use to the levels, domains, evidence, thresholds, participants, and remedies an audit requires.

References

Deng, R.; Jiang, M.; Yu, X.; Lu, Y.; Liu, S. Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Comput. Educ. 2025, 227, 105224.
Han, X.; Peng, H.; Liu, M. The impact of GenAI on learning outcomes: A systematic review and meta-analysis of experimental studies. Educ. Res. Rev. 2025, 48, 100714.
Liu, X.; Guo, B.; He, W.; Hu, X. Effects of Generative Artificial Intelligence on K-12 and Higher Education Students’ Learning Outcomes: A Meta-Analysis. J. Educ. Comput. Res. 2025, 63, 1249–1291.
Liang, W.; Yuksekgonul, M.; Mao, Y.; Wu, E.; Zou, J. GPT detectors are biased against non-native English writers. Patterns 2023, 4, 100779.
Jiang, Y.; Hao, J.; Fauss, M.; Li, C. Detecting ChatGPT-generated essays in a large-scale writing assessment: Is there a bias against non-native English speakers? Comput. Educ. 2024, 217, 105070.