Open educational resources and generative AI: towards a framework for responsible integration

Synthesis: Rodés-Paragarino and Atenas (2026) asked how generative AI and open educational resources can be responsibly integrated in higher education and lifelong learning, and answered with a framework rather than a test. The design is a qualitative framework-development study in two phases: an integrative analysis of international AI and OER policy instruments, then stakeholder co-creation with five international experts, a thematic panel of national informants, and 81 workshop participants in Bilbao in December 2025. No intervention was evaluated and no outcome was measured. The result is a seven-pillar architecture covering public value, collective governance, human agency, quality verification, rights and provenance, metadata and interoperability, and accessibility, localization, and monitoring. The study can show that the two policy spheres govern different objects and leave shared gaps; it cannot show that the framework changes institutional practice.

Key Findings

1. The documentary analysis found that AI and OER frameworks converge on equity, inclusion, lifelong learning, teacher empowerment, international cooperation, and responsible innovation, but diverge on what is being governed: AI guidance centers ethics, competencies, data governance, and deployment, while OER guidance centers openness, licensing, access, reuse, and the knowledge commons.
2. The authors identify the absence of a coherent lifecycle perspective as the most significant gap. AI-mediated resources move through prompt design, model selection, source selection, generation, editing, validation, attribution, metadata enrichment, repository publication, recommendation, localization, analytics, and adaptation, and no reviewed instrument covered that sequence.
3. Seven structural gaps organize the diagnosis: declarative equity, metadata and interoperability, provenance and licensing, quality and trust, legal and institutional accountability, fragmented capacity-building, and infrastructure dependency.
4. Coding of expert, informant, and workshop evidence produced six cross-cutting dimensions, which the authors then used to inform the seven interdependent pillars of the OE-AI Framework rather than a staged procedure.
5. Experts refused to equate human oversight with the formal presence of a reviewer. They defined agency as responsibility, integrity, professional judgment, and answerability, and argued that responsibility cannot be delegated to the individual educator alone.
6. Provenance was reframed as a workflow issue rather than a citation issue, one that documents the origin of content, the nature of AI assistance, the human decisions that shaped a resource, the license status of source materials, and the validation steps taken before publication.
7. Workshop participants asked for disclosure checklists, quality-review protocols, accessibility criteria, metadata fields, and repository workflows, and described responsible integration as a gradual institutional transition rather than an end state.

Two policy spheres that govern different objects

The diagnostic claim is structural, not rhetorical. AI frameworks and OER frameworks share broad educational ambitions, but they distribute responsibility differently and assume different objects of governance. AI instruments treat the object as a technology to be deployed responsibly; OER instruments treat it as a resource to be licensed, shared, and sustained. Once generative models enter the production and circulation of resources, the same artifact is simultaneously a technological output, a pedagogical artifact, a legal object, a data-mediated entity, and a component of institutional infrastructure. The paper's comparison matrix organized documents across four dimensions: normative orientation, operational components, methodological logic, and scope or level of action. That structure is what allows the authors to claim divergence in mechanism rather than merely in emphasis, and to argue that separate governance reproduces the opacity and enclosure both fields say they oppose.

How the framework was built

Phase 1 was a purposive documentary corpus of international policy instruments, institutional guidelines, technical and legal resources, and recent scholarship on AI, OER, digital education, ethics, licensing, and responsible innovation. Phase 2 was held in Bilbao in December 2025 and combined expert validation with a thematic panel and a co-creation workshop. Five international experts came from France, Spain, Portugal, Brazil, and the United Kingdom; the panel drew on national informants with experience in educational technology, public support systems, adult literacy, and open knowledge advocacy; the workshop involved 81 primary, secondary, and university teachers alongside public bodies and civil society.

Two author-built matrices, one for the documentary review and one for stakeholder synthesis, preserved traceability from source to pillar. Analysis was abductive, moving between categories derived from documents and themes emerging from stakeholders. The paper reports triangulation and reflexivity as its rigor strategies but gives no inter-coder agreement statistic, so the framework's reliability rests on documented procedure rather than on measured consensus.

The seven pillars as conditions, not steps

The framework's central design decision is that the pillars are interdependent rather than sequential. Weak provenance weakens quality verification, weak metadata limits discovery and reuse, and inadequate governance removes the conditions for meaningful human agency. The seven areas are normative orientation and public value; collective governance and accountability; human agency in AI-assisted authoring; quality verification and pedagogical validation; rights, provenance, and licensing; metadata, interoperability, and discovery; and accessibility, localization, and sustainable monitoring. Read together, they extend the classical OER freedoms of retaining, reusing, revising, remixing, and redistributing, each of which the paper argues can be constrained by unclear provenance, platform terms, incompatible licenses, or opaque training data. The framework is offered as a diagnostic tool for reviewing workflows, a planning tool for policy, and a reflective tool for professional development. That is a claim about usefulness, not evidence of it.

What the framework cannot establish

The authors describe the framework as a structured synthesis of evidence and professional judgment, and state that it has not been tested through institutional implementation. Nothing in the paper measures whether institutions that adopt the pillars produce more open, better attributed, or more accessible resources than institutions that do not. The stakeholder phase establishes that experts and practitioners recognize the gaps and want the instruments; it does not establish that the instruments work.

A second boundary is conceptual. The claim that AI and OER frameworks are governed separately is supported by the document set the authors assembled, and the document set was selected purposively to build an integrative framework, which makes it well suited to mapping positions and poorly suited to estimating how common those positions are. The paper also concedes that its corpus privileged international policy and recent literature in higher education and lifelong learning, leaving non-dominant linguistic contexts, Global South settings, accessibility domains, and non-formal education for later work.

What this means for practice

Instructors should read the pillars as questions to ask before adopting an AI-assisted resource: who wrote the prompt, what source material entered the model, what license the output carries, who reviewed it, and what metadata will let a colleague find and adapt it. Provenance is the pillar most often missing in practice, because it requires documenting a workflow rather than citing a source. Faculty developers can use the seven areas to audit existing OER workflows and to combine AI literacy with OER literacy in one program rather than two. Administrators hold the levers the paper identifies: procurement criteria, repository metadata fields, defined review roles, and disclosure policy. The framework locates responsibility in institutional arrangements, not in individual caution.

Limitations

• No implementation, longitudinal, or comparative test is reported. The framework was built from documents and stakeholder deliberation and has not been applied in repositories, courses, or institutional policy, which the authors name as future work.
• The stakeholder component is small and purposive: five experts from five countries, a panel of national informants, and 81 workshop participants in one city in December 2025. The authors state it should not be read as representative of all educational contexts.
• The paper is internally inconsistent about the panel: the abstract reports four national informants, while the methods section reports five. Thematic coding was carried out by the author team, and no inter-coder reliability coefficient is reported.
• The documentary corpus privileged international policy instruments and recent literature in higher education and lifelong learning, so coverage of non-dominant linguistic contexts, Global South systems, accessibility domains, and non-formal education is not established.

Citation

Rodés-Paragarino, V., & Atenas, J. (2026). Open educational resources and generative AI: towards a framework for responsible integration. Open Praxis, 18(4), 761-776.