Across learning and assessment, AI systems now generate instructional content, support personalized practice, provide feedback, and enable new forms of measurement. These systems create real opportunity. They also raise important questions about bias, accuracy, accountability, and trust.
In response, the field has turned to Responsible AI (RAI): principles and governance mechanisms designed to reduce harm and increase accountability (see Burstein, 2025; Burstein et al., 2025, for RAI practices at the Duolingo English Test) and to human-centered AI for a clear articulation of human values that need to be reflected in AI applications (von Davier & Burstein, 2024). This work is essential.
At the same time, we believe educational assessment requires a framing that goes beyond principles and checklists toward something more durable. We argue that AI in education requires custodianship.
Custodianship makes Responsible AI sustainable
Custodianship of AI (Dignum, 2025) frames responsibility as an ongoing practice of care, guidance, accountability, and reflection over time. It emphasizes shared authority, diverse perspectives, and human values throughout the design, development, and deployment of AI systems.
This framing matters because education is not a single domain with a single set of values. It is cross-disciplinary, culturally situated, and global. Legitimate disagreements exist about what effective learning and fair assessment look like. While RAI often focuses on ethical principles supported by procedural tools such as audits, documentation, and monitoring, custodianship adds a crucial dimension: who has voice, who holds power, and how responsibility evolves as contexts and technologies change.
Why education needs more than validity and fairness alone
AI is not new to educational assessment. For decades, natural language processing and machine learning have supported automated scoring of writing, short answers, and speech. These systems enabled scale and consistency for tasks that once relied entirely on human labor.
Generative AI extends those capabilities, particularly in content generation and personalized learning. It also expands the risk surface across a broader group of stakeholders: learners, educators, test takers, score users, institutions, and policymakers.
Custodianship helps address this expanded landscape by treating responsibility as shared work across the educational ecosystem, rather than a one-time compliance activity.
Four practices for custodianship of AI in education
To make custodianship actionable, we propose four foundational practices that support responsible, scalable AI use across learning and assessment.
1. Build cross-disciplinary collaboration into AI work
AI-enabled education is never purely technical. It involves experts from across disciplines, such as learning and assessment design, measurement, user experience, ethics, and policy. Custodianship therefore requires teams that include educators, applied linguists, learning scientists, psychometricians, engineers, human–computer interaction specialists, subject matter experts, and ethicists.
Custodianship also asks us to attend to power asymmetries, such as between developers and users, institutions and test takers, and across global contexts. Accountability to stakeholders is a requirement of responsible design. Cross-disciplinary collaboration supports shared authority, clearer decision-making, and collective responsibility.
2. Develop Responsible AI infrastructure that persists with the system
Custodianship depends on institutional memory and sustained governance. This requires infrastructure: standards, guidelines, documentation, auditing practices, monitoring plans, and feedback mechanisms that embed values into everyday decisions.
Effective RAI infrastructure is actionable, transparent, and designed to evolve as technologies and educational contexts change. It connects aspirational ethics—what we believe—to operational ethics—what we do in practice—and signals to stakeholders that AI use is guided by rigor and care.
3. Identify leverage points for responsible scaling
AI’s promise in education is often framed around scale: more content, faster feedback, broader access. Custodianship assumes that learning and assessment are complex systems, where scale can amplify both benefits and risks.
From a systems perspective, custodianship practice helps identify leverage points—places where targeted interventions produce outsized impact. Examples include human-in-the-loop review at critical stages, fairness checks for generated content, monitoring for model drift, and audit-ready documentation embedded into workflows. The goal is responsible scaling that preserves human values with regard to education.
4. Use AI for ideation ethically and thoughtfully
Custodianship also includes a responsibility to enable AI’s benefits. One area of growing importance is AI-assisted ideation in educational research and development, particularly in early stages where speed and breadth matter.
Rather than asking whether AI can be used for ideation, custodianship reframes the question: how should it be used, and under what obligations? Ethical use requires transparency, human responsibility for verification, and clear boundaries against over-reliance. When these conditions are met, AI-assisted ideation can accelerate research, broaden participation, and support innovation that benefits learners.
What custodianship makes possible
Custodianship of AI in education is not a slogan. It is a long-term stance that treats responsibility as practice, authority as shared, and reflection as continuous.
The four practices outlined here—cross-disciplinary collaboration, responsible AI infrastructure, leverage points for responsible scaling, and ethical AI-assisted ideation—offer a foundation that is both principled and practical.
As AI reshapes learning and assessment, custodianship provides a sustainable way to continue to innovate and expand opportunity, while protecting people, preserving human values, and maintaining trust.
This work, authored by Jill Burstein, Alina A. von Davier, and Nava Shaked, will be published in the forthcoming volume Artificial Intelligence Applications in Educational Learning and Assessment, edited by Alina A. von Davier and Duanli Yan.