How to Evaluate Learning Resources: An Evidence-Based Framework for Professionals

Most professionals do not suffer from a shortage of learning material. They suffer from an inability to tell, quickly and reliably, which material is worth their evenings.

That problem has sharpened for three reasons. First, the underlying skill requirements are moving: in the World Economic Forum's employer survey, respondents expected 39% of workers' core skills to change by 2030, and half the workforce surveyed had already completed some form of training as part of a long-term learning strategy (World Economic Forum, 2025). Second, the supply of plausible-sounding material has expanded far faster than the supply of vetted material. A corpus-level study published in Nature Human Behaviour estimated that up to roughly 22% of computer-science abstracts showed signs of large language model (LLM) modification by September 2024 (Liang et al., 2025); in October 2025 arXiv stopped accepting unreviewed review articles and position papers in its computer-science category, citing an unmanageable influx (arXiv, 2025). Third, in some jurisdictions, workplace learning has become a compliance artefact rather than a private matter — Article 4 of the EU AI Act has required providers and deployers of AI systems to act on AI literacy since 2 February 2025 (European Commission, 2026a).

This article is written for practitioners who choose learning resources for themselves or for others: engineers, analysts, clinicians, researchers, managers, and learning-and-development or compliance leads. It sets out where professional learning resources come from, what the research does and does not support about how adults learn from them, an original five-part evaluation framework (PROOF), comparison tables you can reuse, a verification workflow for AI-assisted answers, and a dated summary of recent regulatory and infrastructure changes. It does not recommend specific commercial products.

---

Executive summary

• Resource selection is now a risk-management activity, not a matter of taste: employers surveyed by the World Economic Forum (2025) expect 39% of core skills to shift by 2030, and EU rules oblige organisations to take AI-literacy measures.
• The strongest, most replicated findings in learning science are unglamorous: practice testing and spaced practice were rated high-utility by Dunlosky et al. (2013), while highlighting and rereading — the techniques learners rely on most — were rated low-utility.
• Learner satisfaction is a weak proxy for learning. In a randomised classroom study, students taught actively learned more but felt they had learned less than peers in polished lectures (Deslauriers et al., 2019). Course ratings inherit this bias.
• Completion is a weak proxy for capability. Analysis of six years of MITx and HarvardX data found that completion rates did not improve over time and that most registrants never returned after their first year (Reich & Ruipérez-Valiente, 2019).
• The PROOF framework proposed here — Provenance, Recency, Openness, Outcome evidence, Fit — turns those findings into a repeatable five-minute screen.
• Machine-assisted writing is now measurable in the scholarly record (Liang et al., 2025; Kobak et al., 2025). This does not make sources untrustworthy, but it does make provenance and independent corroboration more important than surface polish.
• Credential infrastructure has matured: Open Badges 3.0 was finalised in 2024 and rebuilt on the W3C Verifiable Credentials Data Model, which became a W3C Recommendation on 15 May 2025 (1EdTech, 2024; W3C, 2025). Cryptographic verifiability confirms who issued a claim, not that the underlying learning was rigorous.
• Openly licensed material is a policy priority, not a fringe option: UNESCO's Recommendation on Open Educational Resources was adopted on 25 November 2019 and remains the reference instrument for member states (UNESCO, 2019).
• The regulatory picture shifted again on 27 July 2026, when Regulation (EU) 2026/1744 amended the AI Act and reframed the Article 4 literacy duty as an obligation to take measures supporting AI literacy (European Commission, 2026b).
• Practical conclusion: build a small, layered resource stack anchored in primary documentation and standards, verify against at least one independent source, and measure yourself on production of work rather than consumption of content.

---

Why resource selection became a professional risk

For most of the last century, the professional learning stack was short and institutionally curated: a textbook, a vendor manual, a professional body's syllabus, a trainer. Selection was mostly delegated to publishers, universities and standards bodies.

Three shifts dismantled that arrangement.

Abundance. Open courseware, preprint servers, vendor documentation, video platforms and community forums removed distribution costs. The choice problem moved from acquisition to triage.

Volatility. Tool-level knowledge now depreciates in months. The Future of Jobs survey data suggest employers see this as a persistent, if slightly stabilising, condition — 39% of core skills expected to change by 2030, down from 44% reported two years earlier (World Economic Forum, 2025). Forecasts of this kind are survey-based expectations, not measurements, and should be read as indicative.

Machine-generated text at scale. Two independent corpus studies using different methods converge on the same direction of travel: Liang et al. (2025) estimated LLM-modified content in up to ~22% of computer-science abstracts by September 2024, and Kobak et al. (2025) estimated a lower bound of 13.5% of 2024 biomedical abstracts, rising to about 40% in some subcorpora. Both are population-level estimates, not detectors of individual documents, and neither establishes that assisted text is inaccurate. What they do establish is that fluency and formatting no longer carry the signal they once did.

Callout — What these studies do not say. They do not say that a fifth of computer-science papers are fabricated, that AI-assisted writing is misconduct, or that detection tools can identify individual offenders. They estimate corpus-level vocabulary shifts. Treat them as evidence about the information environment, not about any specific author.

---

A short history of the professional learning stack

Two milestones anchor the middle of that table. MIT's OpenCourseWare initiative, announced in 2001, established that elite institutions would publish teaching material openly; UNESCO's Recommendation on Open Educational Resources, adopted on 25 November 2019, made openly licensed educational material an explicit policy objective for member states and remains the first international normative instrument in this area (UNESCO, 2019). The Dubai Declaration on OER, adopted on 20 November 2024, extended that agenda to AI and emerging technologies (UNESCO, n.d.).

---

What the evidence actually supports

Techniques with the strongest support

In a systematic review of ten common study techniques, Dunlosky et al. (2013) assigned high utility to only two: practice testing (retrieval practice) and distributed practice (spacing). Elaborative interrogation, self-explanation and interleaved practice were rated moderate. Summarisation, highlighting, keyword mnemonics, imagery for text, and rereading were rated low — despite being the techniques learners most often report using.

The practical consequence for resource selection is direct: a resource that forces recall and spreads exposure over time is doing something the evidence supports. A beautifully produced video series that you watch once is not.

Why your judgement of a resource is unreliable

Deslauriers et al. (2019) randomly assigned university physics students to active instruction or to a polished lecture covering identical content. Students in the active condition scored higher on a test of learning but rated their own learning lower. The authors concluded that evaluating instruction by learner perception risks favouring inferior methods.

Two caveats matter. The study was conducted in a single institution and discipline over two class sessions, and it measured short-term outcomes. Its finding should be treated as a strong caution about satisfaction metrics rather than a universal law. Still, it is the most direct evidence available that star ratings and "this was so clear!" comments are contaminated signals.

Why completion tells you little

Reich and Ruipérez-Valiente (2019) analysed MITx and HarvardX course data on edX from October 2012 to May 2018 and reported three patterns: most learners never returned after their first year, growth concentrated in wealthier countries, and completion rates showed no improvement across six years. Their data are platform-specific and now several years old, but the structural point survives: enrolment and completion measure funnel behaviour, not competence.

A myth worth retiring

Matching instruction to a learner's self-reported "learning style" remains widely believed and poorly supported; Pashler et al. (2008) found that the evidence base lacked the experimental designs needed to justify the practice. Choosing a resource because it is "visual" or "auditory" is not an evidence-based criterion. Choosing it because it requires retrieval, spacing and feedback is.

---

The PROOF framework: a five-minute screen

The framework below is original to OneWise and is designed to be applied before you commit time, not after.

How to use it. Score each dimension 0 (absent), 1 (partial) or 2 (strong). A resource scoring 0 on Provenance or Outcome evidence should generally be rejected regardless of its total. A resource scoring 8–10 is worth a serious time commitment. Between 4 and 7, use it as a supplementary explainer and verify its claims elsewhere.

Figure 1 — Diagram brief for the OneWise design team. Title: The PROOF screen: from candidate resource to committed study time Purpose: Show PROOF as a sequential filter with two hard gates, ending in one of three outcomes. Components and sequence (left to right): (1) Rounded rectangle "Candidate resource" → (2) five stacked hexagons labelled P, R, O, O, F, each with a one-line caption (Provenance / Recency / Openness / Outcome evidence / Fit) → (3) diamond decision node "Provenance or Outcome evidence = 0?" → (4) three terminal rounded rectangles: "Reject", "Supplementary use + verify", "Commit study time". Arrows: Solid arrows left to right through the hexagons; a red dashed arrow from the diamond down to "Reject"; two solid arrows fanning right to the remaining terminals, labelled "score 4–7" and "score 8–10". Visual hierarchy: Hexagon band is the visual centre; terminal states are colour-coded (grey, amber, green); the decision diamond is the only red element. Caption: PROOF is a screen, not a scorecard: two dimensions act as hard gates, the remaining three set the depth of commitment.

---

Comparing resource types

No single resource type wins. The table below compares the categories most professionals draw on, using criteria that matter for time allocation.

---

A verification workflow you can actually run

Verification fails when it is framed as a research project. Framed as a two-minute habit, it is sustainable.

1. Separate the claim types. Distinguish (a) syntax and configuration details, (b) causal or empirical claims, (c) legal or regulatory claims, and (d) opinions about best practice. Each has a different authoritative source.
2. Route each claim to its primary source. Configuration → official documentation for the exact version. Empirical claim → the underlying study, not the press release. Regulatory claim → the legal instrument or the regulator's own guidance. Best practice → at least two independent practitioner sources, ideally with disagreement visible.
3. Check the date against the artefact's own version. A correct answer for version 2 is a wrong answer for version 5. Undated advice is unusable advice.
4. Corroborate independently. Two sources that both derive from the same upstream post are one source. Follow citations upward until you reach something primary.
5. Test it. For technical claims, the cheapest verification is execution: run the command, reproduce the calculation, apply the rule to a case where you already know the answer.
6. Record the outcome. A short note recording the claim, the source, the date and the version converts one-off verification into a reusable personal knowledge base — and, in regulated settings, into evidence.

Figure 2 — Diagram brief for the OneWise design team. Title: Triangulating an AI-assisted answer Purpose: Show the loop between a generated answer, primary sources and practical testing. Components: Central node "Generated answer / summary"; three surrounding nodes: "Primary documentation (versioned)", "Peer-reviewed or regulatory source", "Local test or reproduction". A fourth node below: "Verified note (claim + source + date + version)". Relationships: Bidirectional arrows between the central node and each of the three surrounding nodes, labelled "check", "corroborate" and "execute". A single downward arrow from the centre to "Verified note". A dashed feedback arrow from "Verified note" back to the centre labelled "reuse". Visual hierarchy: Central node in a neutral tone; the three verification nodes in a stronger accent colour to signal that they, not the generated answer, carry authority. Caption: A generated answer is a hypothesis. The three surrounding checks are what make it usable.

---

Credentials: verifiable is not the same as valuable

Credential infrastructure has quietly become the most standardised part of the learning stack. Open Badges 3.0 was finalised by 1EdTech in 2024 and rebuilt on the W3C Verifiable Credentials data model, allowing achievements to be cryptographically signed, checked without contacting the issuer, and aligned to skills frameworks (1EdTech, 2024). The underlying Verifiable Credentials 2.0 family became a W3C Recommendation on 15 May 2025, defining the issuer–holder–verifier model that these credentials rely on (W3C, 2025).

This solves forgery. It does not solve meaning.

For hiring managers and L&D teams, the practical rule is to read the assessment design behind a credential, not the badge: what was assessed, by whom, under what conditions, and with what pass standard. For individuals, a portfolio of inspectable work — a repository, a published analysis, an internal post-incident review — remains the highest-signal artefact available, because it is difficult to obtain without the underlying capability.

---

Latest developments

Dated, verifiable changes affecting professional learning resources.

• 2 February 2025 — Article 4 of the EU AI Act became applicable, requiring providers and deployers of AI systems to address AI literacy among staff and others operating systems on their behalf (European Commission, 2026a).
• 15 May 2025 — The W3C published the Verifiable Credentials 2.0 family as W3C Recommendations, including the Verifiable Credentials Data Model v2.0 (W3C, 2025).
• May 2025 — The European Commission's AI Office published questions and answers on the AI literacy obligation and launched a public repository that now holds more than 40 AI-literacy practices contributed by companies and public bodies. The Commission explicitly notes that replicating a listed practice does not confer presumption of compliance (European Commission, 2026a).
• 2 July 2025 — Science Advances published Kobak et al.'s excess-vocabulary analysis of more than 15 million PubMed abstracts, estimating that at least 13.5% of 2024 abstracts had been processed with LLMs (Kobak et al., 2025).
• 31 October 2025 — arXiv's computer-science category began requiring documented prior peer review for review articles and position papers, citing an unmanageable influx of such submissions (arXiv, 2025).
• 2025 — Nature Human Behaviour published Liang et al.'s analysis of 1,121,912 papers and preprints, estimating LLM-modified content in up to ~22% of computer-science papers by September 2024 (Liang et al., 2025).
• 17 June 2026 — The European Commission and the OECD published the final AILit Framework, Empowering Learners for the Age of AI, organised around four domains: engage with, create with, manage and shape AI. Its scope is primary and secondary education, but it is the clearest signal yet of a shared international vocabulary for AI literacy (European Commission, 2026c).
• 1 July 2026 — arXiv spun out from Cornell University to operate as an independent nonprofit, a governance change relevant to anyone whose learning depends on preprint infrastructure (arXiv, 2026).
• 24–27 July 2026 — Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July and entered into force on 27 July 2026. It amends the AI Act, deferring most stand-alone Annex III high-risk obligations to 2 December 2027 and reframing Article 4 as a duty to take measures supporting AI literacy rather than to guarantee a sufficient level (European Commission, 2026b).

Interpretive note: the Omnibus lowers the documentary burden of Article 4; it does not remove the obligation, and Article 50 transparency duties were not deferred. Organisations that had built role-based literacy programmes have lost nothing. Those that had built none still have an obligation of effort to discharge. Legal interpretation of the amended wording is still developing; this article is not legal advice.

---

Documented examples and illustrative cases

Documented — MIT OpenCourseWare (United States, university). MIT's decision to publish course materials openly, announced in 2001, created a durable, freely reusable teaching archive and helped establish the norm that institutional teaching material can be an open public good. It also illustrates the core limitation of open courseware: materials without assessment support exposure rather than measured capability.

Documented — UNESCO OER Recommendation (global, intergovernmental). Adopted on 25 November 2019, the Recommendation commits member states to support the creation, use and adaptation of openly licensed educational materials, with periodic reporting on implementation (UNESCO, 2019). For corporate L&D teams, its practical relevance is licensing: openly licensed material can be legally adapted into internal training, whereas most commercial course content cannot.

Documented — European Commission repository of AI literacy practices (EU, public and private sector). More than 40 organisational practices have been collected and published, gathered through surveys of AI Pact signatories and a wider call (European Commission, 2026a). It is a rare public corpus of what organisations actually did, with the important caveat that inclusion implies neither endorsement nor evaluation.

Documented — standards as curricula (global). The NIST AI Risk Management Framework (AI RMF 1.0), released on 26 January 2023, and its Generative AI Profile, released on 26 July 2024, are widely used as structuring documents for internal governance training, alongside the certifiable management-system standard ISO/IEC 42001 (NIST, 2023). Standards make unusually good curricula because they are versioned, publicly reviewed and stable enough to build assessments against.

Illustrative (hypothetical) — a 40-person fintech engineering team. A team lead must upskill engineers on a new payments API and on AI-assisted coding review. Applying PROOF, the vendor's versioned API documentation scores highly on Provenance and Recency but zero on Outcome evidence; it is paired with an internal lab in which each engineer ships one instrumented change against a sandbox. A popular video course scores well on Fit but poorly on Recency; it is used only for conceptual grounding. The credential decision is deferred until the team can articulate what the badge would prove to whom. This example is constructed to demonstrate the method, not drawn from a reported deployment.

---

Frequently misunderstood ideas

• "Free means low quality." Licensing describes rights, not rigour. Some of the highest-authority resources available — standards, regulator guidance, national statistics — are free.
• "Peer-reviewed means correct." Peer review filters; it does not guarantee. Single studies, especially in education, generalise poorly across contexts.
• "AI-assisted means untrustworthy." The corpus studies cited here measure prevalence, not error. Undisclosed assistance is a transparency problem; the accuracy question still has to be settled claim by claim.
• "Newer is better." Recency matters for tooling and law. For foundations — statistics, distributed systems, physiology — a 2012 text may be better than a 2026 summary of it.
• "If I understood it, I learned it." Comprehension during instruction and durable retrieval afterwards are different outcomes, and the first predicts the second poorly (Deslauriers et al., 2019).

---

Barriers, trade-offs and unresolved questions

Time is the binding constraint, not access. For most professionals, the marginal cost of another course is attention, not money. Any framework that adds evaluation overhead has to save more time than it consumes — which is why PROOF is designed as a five-minute screen with hard gates.

Equity and language. Evidence of concentration in wealthier countries in large-scale online learning (Reich & Ruipérez-Valiente, 2019) is a reminder that open access is necessary but not sufficient. Bandwidth, English-language dominance, working hours and childcare shape who benefits.

Cognitive offloading. Delegating explanation to an assistant may reduce the retrieval effort that produces durable learning. Direct evidence on long-term effects in professional settings is still thin, and claims in either direction should be treated as preliminary. The prudent position: use assistants to generate practice questions and counter-arguments rather than to supply finished understanding.

Measurement and privacy. Learning analytics can improve programme design and can also become surveillance. Organisations should ask what decision each metric informs before collecting it, and apply the same data-minimisation logic they would to any other personal data.

Environmental cost. Streaming video and inference-heavy tooling carry energy costs that are rarely visible to the learner. Published per-query figures vary widely and depend on model, hardware and workload; treating any single figure as settled is unwise.

Open research questions. How well laboratory findings on retrieval and spacing transfer to self-directed adult learning at work; whether verifiable micro-credentials change hiring behaviour at scale; and what quality-assurance mechanisms can survive machine-generated content at volume — none of these is settled.

---

Practical takeaways

If you are an individual professional

1. Anchor each learning goal in one primary source — a specification, official documentation or a standard — then add explanatory material around it.
2. Apply PROOF before you start, not after you have spent four hours.
3. Convert every resource into retrieval: after each session, write from memory what changed in your understanding, then check.
4. Space revisits at expanding intervals rather than binge-consuming.
5. Keep a dated verification log with source, version and outcome; it compounds.
6. Ship an artefact per topic — a script, a memo, a diagram, a review. Artefacts survive; certificates decay.

If you lead a team

1. Define the capability you need in observable terms before you shortlist any resource.
2. Buy assessment, not content: prefer resources with graded work, or pair free material with an internal exercise.
3. Budget the working hours, not just the licence fee. Unbudgeted learning is unlearned.
4. Maintain a small, curated internal list of primary sources per domain, with named owners and review dates.

If you are responsible for L&D or compliance

1. Map training to role-specific risk rather than to headcount coverage.
2. Document your reasoning: under the amended Article 4, the defensible question is what measures you took and why (European Commission, 2026b).
3. Prefer openly licensed material where you need to adapt content internally, and record the licence terms.
4. Review supplier claims against assessment design and issuer identity, using verifiable credential standards where available (1EdTech, 2024; W3C, 2025).
5. Re-examine satisfaction-based evaluation: high scores may indicate comfort rather than learning (Deslauriers et al., 2019).

---

Key insights

1. Evaluate resources before consuming them; triage is the scarce skill, not access.
2. Practice testing and spaced practice have the strongest support of the common study techniques (Dunlosky et al., 2013).
3. Learner satisfaction and actual learning can move in opposite directions (Deslauriers et al., 2019).
4. Completion rates measure funnels, not competence (Reich & Ruipérez-Valiente, 2019).
5. Fluency is no longer a quality signal; provenance and corroboration are (Liang et al., 2025; Kobak et al., 2025).
6. Primary sources — standards, specifications, regulator guidance — are usually free and usually underused.
7. Cryptographic verifiability authenticates the issuer, not the rigour (W3C, 2025; 1EdTech, 2024).
8. Open licensing is a workflow advantage for internal training, not a quality claim (UNESCO, 2019).
9. In the EU, learning provision is now partly a documented compliance activity (European Commission, 2026a, 2026b).
10. Artefacts you produce are the most reliable evidence of capability you can offer or assess.

---

Frequently asked questions

What is the fastest way to evaluate a learning resource? Apply a five-part screen: Provenance, Recency, Openness, Outcome evidence and Fit. Reject anything scoring zero on provenance or outcome evidence; commit time only to resources that score strongly on at least four of the five.

How can I tell if a technical article is out of date? Look for a visible publication or revision date, an explicit product version, and whether the interfaces or commands shown match the current release. Undated technical writing should be treated as unverified until checked against versioned documentation.

Are free learning resources as good as paid ones? Price and quality are largely uncorrelated. Standards bodies, regulators, statistical agencies and universities publish authoritative material at no cost. What paid resources more often provide is structured assessment, feedback and pacing — which is precisely what free material tends to lack.

Is it safe to learn from AI assistants? They are useful for drafting, summarising and generating practice questions, and unreliable as final authorities. Treat any generated answer as a hypothesis to be checked against versioned documentation, a primary study or a local test.

Do online course certificates help careers? They can signal exposure and initiative, and vendor certifications carry weight in some specific labour markets. Evidence that generic completion certificates predict on-the-job capability is weak, and completion itself has historically been low in large-scale online courses (Reich & Ruipérez-Valiente, 2019).

What is the difference between a micro-credential and a badge? "Micro-credential" describes the scope of the achievement; "badge" usually describes the technical format used to express it. Open Badges 3.0 expresses achievements as verifiable credentials aligned to the W3C data model (1EdTech, 2024).

What are Open Badges 3.0 and why do they matter? They are cryptographically signed digital credentials, finalised in 2024, that can be verified without contacting the issuer and aligned to skills frameworks. They reduce forgery risk and improve portability; they do not certify the quality of the assessment behind the claim.

What are verifiable credentials? A W3C standard, published as a Recommendation on 15 May 2025, for expressing claims made by an issuer about a subject in a tamper-evident, machine-checkable form, exchanged between issuers, holders and verifiers (W3C, 2025).

What are open educational resources (OER)? Teaching, learning and research materials that are in the public domain or released under an open licence permitting no-cost access, reuse, adaptation and redistribution, as defined in UNESCO's 2019 Recommendation (UNESCO, 2019).

Does the EU AI Act require me to train my staff on AI? Article 4 requires providers and deployers to take measures relating to AI literacy for staff and others operating AI systems on their behalf. It applied from 2 February 2025 and was amended on 27 July 2026 into an obligation to take measures supporting AI literacy. No specific course or certificate is mandated. Consult qualified counsel for your obligations (European Commission, 2026a, 2026b).

How much time should I allocate to learning each week? There is no evidence-based universal number. What the evidence supports is distribution: several shorter, spaced sessions with active recall generally outperform a single long block of the same total duration (Dunlosky et al., 2013).

How do I verify a statistic I found in an article? Follow it upward to its origin — the study, dataset or official report — and check what was actually measured, over what period and with what population. Secondary coverage frequently drops the qualifiers that determine what a number means.

Why did arXiv restrict review articles in computer science? From 31 October 2025, arXiv required documented prior peer review for review and position papers in its CS category, citing an unmanageable influx of such submissions in an environment where generative tools make them fast to produce (arXiv, 2025).

Should I trust preprints? Preprints are valuable for tracking fast-moving fields and carry no quality guarantee. Check whether the work has since been peer-reviewed, whether data and code are available, and whether independent groups report similar results.

What should I do if two authoritative sources disagree? Identify whether the disagreement is empirical, definitional or normative. Empirical disputes are resolved by looking at the underlying data and methods; definitional ones by stating which definition you are using; normative ones cannot be resolved by evidence alone and should be reported as contested.

How should I evaluate an internal corporate training programme? Judge it on assessment design and on transfer to work — observable changes in output, error rates or decision quality — rather than on satisfaction scores, which can favour comfortable instruction over effective instruction (Deslauriers et al., 2019).

---

Glossary

AI literacy — Skills, knowledge and understanding that allow providers, deployers and affected persons to make informed use of AI systems and to recognise the associated opportunities and risks; defined in the EU AI Act.

Comprehensive Learner Record (CLR) — A 1EdTech standard for bundling multiple verifiable achievements into a single learner record.

Creative Commons licence — A family of standardised public copyright licences permitting defined reuse, from attribution-only to more restrictive variants.

Distributed (spaced) practice — Spreading study of the same material across separated sessions rather than massing it.

Interleaving — Mixing different problem types or topics within a practice session instead of blocking them.

Micro-credential — A credential covering a narrow, defined set of learning outcomes, typically smaller than a formal qualification.

MOOC — Massive open online course; a large-enrolment online course, usually free to access with paid certification options.

Open Badges — A 1EdTech specification for portable digital credentials; version 3.0, finalised in 2024, is built on verifiable credentials.

Open educational resources (OER) — Educational materials in the public domain or openly licensed for no-cost access, reuse, adaptation and redistribution.

Preprint — A research manuscript posted publicly before, or without, formal peer review.

Provenance — The documented origin, authorship and chain of custody of a piece of information.

Retrieval practice (practice testing) — Deliberately recalling information from memory as a study technique, rather than rereading it.

Verifiable credential (VC) — A tamper-evident, cryptographically signed digital claim made by an issuer about a subject, defined by the W3C Verifiable Credentials Data Model.

---

References

Academic papers

Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom. Proceedings of the National Academy of Sciences, 116(39), 19251–19257. https://doi.org/10.1073/pnas.1821936116

Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266

Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27), eadt3813. https://doi.org/10.1126/sciadv.adt3813

Liang, W., Zhang, Y., Wu, Z., Lepp, H., Ji, W., Zhao, X., Cao, H., Liu, S., He, S., Huang, Z., Yang, D., Potts, C., Manning, C. D., & Zou, J. Y. (2025). Quantifying large language model usage in scientific papers. Nature Human Behaviour, 9, 2599–2609. https://doi.org/10.1038/s41562-025-02273-8

Pashler, H., McDaniel, M., Rohrer, D., & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119. https://doi.org/10.1111/j.1539-6053.2009.01038.x

Reich, J., & Ruipérez-Valiente, J. A. (2019). The MOOC pivot. Science, 363(6423), 130–131. https://doi.org/10.1126/science.aav7958

Standards and official documentation

1EdTech Consortium. (2024). Open Badges specification, version 3.0. https://www.imsglobal.org/spec/ob/v3p0

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.100-1

World Wide Web Consortium. (2025, May 15). Verifiable Credentials Data Model v2.0 (W3C Recommendation). https://www.w3.org/TR/vc-data-model-2.0/

Government and intergovernmental sources

European Commission. (2026a). AI talent, skills and literacy. Shaping Europe's Digital Future. https://digital-strategy.ec.europa.eu/en/policies/ai-talent-skills-and-literacy

European Commission. (2026b, July 27). AI Omnibus enters into force [News release]. Shaping Europe's Digital Future. https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force

European Commission & Organisation for Economic Co-operation and Development. (2026c). Empowering learners for the age of AI: An AI literacy framework for primary and secondary education. https://ailiteracyframework.org/

UNESCO. (2019). Recommendation on Open Educational Resources (OER). https://www.unesco.org/en/legal-affairs/recommendation-open-educational-resources-oer

UNESCO. (n.d.). Open Educational Resources. Retrieved 4 August 2026, from https://www.unesco.org/en/open-educational-resources

Industry reports and institutional sources

arXiv. (2025, October 31). Attention authors: Updated practice for review articles and position papers in arXiv CS category. arXiv blog. https://blog.arxiv.org/2025/10/31/attention-authors-updated-practice-for-review-articles-and-position-papers-in-arxiv-cs-category/

arXiv. (2026, June 30). arXiv's next chapter: Updates on our spin out from Cornell University. arXiv blog. https://blog.arxiv.org/2026/06/30/arxivs-next-chapter/

World Economic Forum. (2025). The future of jobs report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025/

---

Editorial transparency note

This article is based on published research, official documentation and primary regulatory sources, all cited above and accessed on 4 August 2026. It does not draw on undisclosed first-hand experience, proprietary datasets or confidential organisational data. The PROOF framework, all tables, diagram briefs and the hypothetical fintech example were created for OneWise. Where evidence is preliminary, contested or context-limited, this is stated in the text. Regulatory summaries are provided for orientation and are not legal advice.