Building sustainable futures: Responsible development of AI-enabled open educational resources as digital public goods

Synthesis: Atenas, Farrow, Caldwell, Costello and Havemann (2026) ask how open educational resources should be governed if they are to function as digital public goods in the age of generative AI, and answer with policy analysis rather than a study. No data were collected and there is no sample, only a reading of United Nations and UNESCO instruments: the Common Agenda, the 2024 Global Digital Compact, the 2019 Recommendation on OER, the 2024 Dubai Declaration, the ROAM-X framework with its 303 Internet Universality Indicators, and the FAIR principles. The diagnosis is fragmentation: disconnected repositories, inconsistent metadata, uneven connectivity and weak cross-border collaboration. The prescription is integrated, interoperable, non-commercial and publicly governed platforms, plus openly licensed AI rather than dependence on systems such as ChatGPT. No platform is evaluated, so the paper shows that the problem is described consistently across policy documents, not that the remedy works.

Key Findings

1. The paper's organizing move is taxonomic: OER are classified as digital public goods, defined through UNDP as building blocks for digital public infrastructure, which shifts the policy object from the individual resource to the platform that holds it.
2. It describes fragmentation as structural rather than transitional: resources sit on disconnected platforms, often in formats that resist reuse, without consistent metadata, and the gap widens where internet access is least reliable.
3. It adds a distributional claim: the global OER landscape is uneven, with Global North dominance of material in English, and limited collaboration between institutions and between countries.
4. Its remedy is a standards layer, ROAM-X (rights, openness, accessibility, multi-stakeholder participation, cross-cutting issues) for governance, described via UNESCO's 303 Internet Universality Indicators, and FAIR (findable, accessible, interoperable, reusable) for indexing and reuse.
5. On AI it holds that benefits and risks arrive together: translation, speech output, automated quizzes and rubrics, adaptive pathways and lower production cost, against market capture by venture capital, blurred consent and attribution, energy demand and flooded low-quality content.
6. It argues that copyright currently works against OER creators, since United States rulings suggest AI-generated content cannot be copyrighted because the law assumes human authorship, leaving such material in the public domain.
7. Its recommendations are legal and organizational rather than technical: open licensing of publicly funded materials, disclosure when content is machine generated, informed consent and institutional audit of learner data, independent academic review, and capacity building in data and AI literacy.

Fragmentation as the problem the paper names

The paper sets OER inside a policy lineage that starts with dissatisfaction with profit-driven academic publishing and the founding of Creative Commons and the OpenCourseWare consortium in 2001. It notes, citing Creative Commons' 2019 annual report, that over 2 billion works had been released under open licenses by that point. UNESCO's 2019 Recommendation supplies the operative definition: materials in any format and medium that sit in the public domain or under a license permitting no-cost access, reuse, repurposing, adaptation and redistribution.

The digital public goods frame carries the analytic weight. Such goods are defined, through UNDP, as building blocks that let countries build safe, trusted and inclusive digital public infrastructure, and infrastructure is defined in turn as the foundational systems through which people, businesses and governments interact. Classifying OER that way makes interoperability a condition of the category rather than an aspiration attached to it. The authors argue that resources dispersed across unconnected platforms, in formats that resist reuse and without consistent metadata, cannot do infrastructural work, and that the failure is sharpest where connectivity is worst.

ROAM-X and FAIR are then assembled as the operating standards, with two tables translating principle into practice. Findable means standardized metadata and persistent identifiers, with a DOI per item. Reusable means a clear open license, modular design and documentation of how the content may be modified. Interoperability is what lets an AI-driven platform exchange material with systems it did not build, which the paper ties to flexibility and to adaptation to local contexts and languages.

What AI adds and what it takes away

The benefits are listed concretely: translation into multiple languages, text to speech and other assistive functions, automated production of quizzes, rubrics and case studies, pathways that adjust to progress, and updating of material as research moves. Cost and time fall, which the authors say returns educator attention to facilitation.

The risks get equal space, and the sharpest is market capture. Venture capital could centralize educational resources in a few corporations, and the paper calls the resulting control oligarchic because it prioritizes private profit over the availability of diverse and inclusive material. It notes efforts to build large language models for low-resource languages such as Finnish and Basque, then qualifies the hope: whether those initiatives are themselves open and democratic is uncertain.

The remaining risks are of a piece. Generative models blur consent and attribution when trained without attention to copyright. Their computation and energy demand raises carbon concerns. Cheap generation can flood the landscape with incorrect, biased or misleading content that fails quality standards and produces mistrust. Models perform best in well-documented domains, so niche topics are underrepresented, a pattern the authors compare to Wikipedia coverage. Fabricated sources and citations undermine the markers of quality that OER users rely on. Platforms advertising open or free AI may operate on a quid pro quo in which learner data are the currency.

Authorship, licensing and the limits of the commons

Copyright is treated as a live obstacle rather than a settled framework. The paper reports that United States rulings suggest AI-generated content cannot be copyrighted, because copyright assumes human authorship, so such content falls into the public domain. That leaves creators of AI-assisted OER without the ownership the licensing system presupposes, even as training on copyrighted works is contested separately, in litigation the paper cites as The New York Times vs. OpenAI.

The harder case is traditional and indigenous knowledge. Licensing built on the Creative Commons model assumes a rights holder free to license, and handles knowledge subject to cultural restrictions poorly. The paper asks for culturally sensitive licenses that respect those restrictions while allowing controlled access, and concedes that licenses supporting traditional knowledge are still under development. Its ethical test for the rest is data justice, in three pillars: visibility, engagement with technology, and non-discrimination.

A framework of levers, never tested

The recommendations are layered rather than ranked. At the level of principle: promote responsible use of AI, publish ethical guidelines for integrating it into OER creation, and fund research on its effects in open education. At the level of policy: capacity building through AI competency frameworks, informed consent mechanisms, disclosure that content is AI generated, plain-language brokering to stakeholders, independent academic review of AI-enabled OER, and an equity guarantee. A closing table sets out six elements, from skills development through privacy safeguarding, collaboration, ethical integration and evaluation to a policy framework.

What is missing is any test of these levers. The paper concedes that substantial innovation aligning AI with open education remains elusive, and that standards for technology and intellectual property in open education remain underdeveloped. Its account of what is going wrong is well sourced and consistent across the instruments it cites. Its account of what would fix it is a design proposal, and the design is not compared against an alternative, costed, or piloted anywhere in the text.

What this means for practice

This paper addresses institutions and funders more directly than it addresses teachers. What carries over to a classroom is the test it applies to any tool: whether the material is openly licensed rather than merely free, whether the metadata would let someone else find and rebuild it, and whether the learner data it collects are consented to and audited. The paper's warning about platforms that are free in exchange for data applies to classroom tools as directly as to national repositories, and its point about English-language dominance applies to any instructor choosing what to translate and adapt.

Limitations

• No data were collected and the paper's own data availability statement reads "Not applicable". No platform, policy or country is examined as a case, so the claim that OER infrastructure is fragmented is supported by citation rather than by any inventory, audit or metadata sample.
• The central prescription has no implementation account. The paper does not say who would fund publicly governed platforms, how interoperability would be enforced, or what a compliant platform would look like in operation, and it concedes that standards for technology and intellectual property remain underdeveloped.
• Legal claims are drawn from United States material, chiefly rulings on authorship and a pending case, then applied to a global open licensing system whose other jurisdictions are not discussed.
• The paper argues from international instruments that already share its premises, so agreement among them is treated as corroboration. Documents that contest the digital public goods framing are not engaged.

Citation

Atenas, J., Farrow, R., Caldwell, N., Costello, É., & Havemann, L. (2026). Building sustainable futures: Responsible development of AI-enabled open educational resources as digital public goods. International Review of Education.