UNESCO AI Ethics Recommendation and Course Evaluation

Bottom line up front: Most European quality-assurance teams weighing AI for course evaluation are watching two instruments — the EU AI Act and, more recently, the Council of Europe's AI treaty. There is a third that reaches almost every public university on the planet and is largely missing from the conversation: the UNESCO Recommendation on the Ethics of Artificial Intelligence, adopted by acclamation by all 193 Member States in November 2021. It carries no fines and imposes no direct legal duty. But it is the closest thing the world has to a shared ethical baseline for AI, it names education explicitly, and it reframes AI in your feedback process as a matter of human rights, human oversight and public trust — not a compliance checkbox.

What the Recommendation actually is

In November 2021, UNESCO's General Conference adopted the Recommendation as, in the organisation's own description, the first global standard-setting instrument on the ethics of AI. All 193 Member States endorsed it (UN News; UNESCO). Because it is a Recommendation rather than a convention, it is not legally binding — but adoption "by acclamation" signals a rare consensus, and it functions as the reference values that national AI strategies, funders and regulators increasingly cite.

Its core is a set of values and principles: proportionality and do-no-harm; safety and security; fairness and non-discrimination; sustainability; the right to privacy and data protection; human oversight and determination; transparency and explainability; responsibility and accountability; awareness and literacy; and multi-stakeholder governance. Read that list against a course-evaluation pipeline that quietly runs student comments through a large language model, and several principles light up at once.

Why a non-binding text reaches your evaluation office

The Recommendation is addressed to Member States, and through them to the public bodies they run — which is most European universities. It is not abstract. UNESCO built two implementation tools that give it teeth in practice:

The Readiness Assessment Methodology (RAM), which asks a country to evaluate its AI ecosystem across five dimensions — legal and regulatory, social and cultural, economic, scientific and educational, and technological and infrastructural. Education is not an afterthought in this instrument; it is one of the five axes on which national readiness is judged.
The Ethical Impact Assessment (EIA), a structured, system-level review of a specific AI application's ethical impacts before and during deployment — a natural fit for a university asking whether its feedback-analysis tooling is fair, transparent and overseen.

The Recommendation also contains a dedicated policy action area on education and research. In other words, an instrument that most evaluation teams have never read already contemplates exactly the kind of decision they are making when they decide whether — and how — to let AI touch student voice.

The three-instrument stack

It helps to see the UNESCO text as the top layer of a three-part stack, each with a different job:

UNESCO Recommendation (2021) — global, soft-law, values-based. It sets the ethical floor and applies wherever a government has endorsed it, inside or outside the EU.
Council of Europe Framework Convention on AI (2024) — a binding international treaty that regulates the deploying public authority in human-rights terms. We unpack it in what the Council of Europe AI treaty means for public-university course evaluation.
EU AI Act — binding EU law that regulates the product and the deployer by risk tier. When course-evaluation AI appraises staff, the relevant duties sit with the university as deployer under Article 26, and the high-risk classification can bite.

The AI Act tells you what you must do. The UNESCO Recommendation tells you what "good" looks like even where the law is silent — and it covers institutions in the many jurisdictions the EU and Council of Europe do not reach.

What a compliance-first reading misses

The Recommendation's real contribution is that it insists on principles that a narrow legal audit can skate past. Three matter most for course evaluation.

Human oversight and determination. The text is explicit that final responsibility and accountability must remain with humans; AI systems should not displace ultimate human decision-making. A dashboard that ranks instructors on an automated sentiment score, unread and unchallenged, is precisely the pattern this principle warns against — the same automation-bias risk that turns a summary into an unaccountable verdict.

Transparency and explainability. Students and staff should know when AI is involved and be able to understand, in meaningful terms, how a conclusion was reached. That is not satisfied by a black-box theme label; it requires provenance back to the underlying quotes.

Awareness and literacy. The Recommendation frames AI literacy as an ethical obligation, not a nice-to-have — echoing, and predating, the EU's own AI-literacy duty. A committee that cannot interrogate how a feedback tool works cannot exercise the oversight the same document demands.

But isn't a non-binding recommendation just aspirational?

This is the strongest objection, and it deserves a straight answer: yes, the Recommendation creates no enforceable obligation and no fine, and it is not a legal shield. If your only question is "what can we be penalised for?", the AI Act and national law are where you look.

But three things make it matter anyway. First, scope: it is the only one of the three instruments that follows your institution outside the EU and the Council of Europe's membership, which is decisive for globally networked universities and transnational provision. Second, gravity: because all 193 states adopted it, it is the vocabulary that national strategies, research funders and procurement frameworks reach for — and increasingly what accreditation reviewers expect to see reflected in your governance. Third, direction of travel: soft law here has repeatedly hardened into binding law, and building to the higher ethical standard now is cheaper than retrofitting later. Treating it as "just aspirational" is the same mistake as treating an unread satisfaction mean as evidence — technically defensible, practically negligent.

Where Koji fits

Koji was built for exactly the posture the Recommendation describes: AI that assists human judgement rather than replacing it. Its AI-moderated conversational interviews probe beyond a number, but the interpretation and every consequential decision stay with people — human oversight by design, not by policy memo. Automatic thematic analysis of open text is traceable to the exact quotes behind each theme, so a panel can see why a conclusion was drawn, satisfying explainability. AI moderation is disclosed to respondents, addressing transparency, and the platform is built for GDPR/AVG-appropriate, data-minimising handling of EU student data. The same AI interview engine powers general user and customer research on koji.so, so teams that also run stakeholder or employer research work to one consistent, oversight-first standard.

Crucially, Koji is framed to mitigate and surface, never to eliminate: it reduces reliance on a single biased global score by returning structured, quote-anchored evidence a human can weigh — which is what human oversight and determination actually require in practice.

The UNESCO Recommendation will not send you a fine. But it is the clearest published statement of what responsible AI in education looks like, agreed by the entire world. If your evaluation process can be explained, is overseen by people, and treats students and staff as rights-holders rather than data points, you are already most of the way there.

Frequently asked questions

Is the UNESCO Recommendation on the Ethics of AI legally binding? No. It is a Recommendation, not a convention, so it creates no directly enforceable obligation or fine. Its force is ethical and political: it was adopted by all 193 Member States and functions as the reference baseline that national AI strategies, funders and regulators cite.

When was it adopted and by how many countries? It was adopted by acclamation at UNESCO's General Conference in November 2021 by all 193 Member States, and UNESCO describes it as the first global standard-setting instrument on the ethics of AI.

Does it actually mention education? Yes. Education and research is one of its policy action areas, and "scientific and educational" is one of the five dimensions of the Readiness Assessment Methodology countries use to gauge their AI readiness.

How is it different from the EU AI Act? The AI Act is binding EU law that regulates AI products and their deployers by risk tier; the UNESCO Recommendation is a global, non-binding statement of values and principles. The Act tells you what is mandatory; the Recommendation defines responsible practice even where the law is silent, and applies far beyond the EU.

What does it require for AI used in course evaluation? It has no line-item requirements, but its principles imply: meaningful human oversight of any consequential decision, transparency that AI is involved, explainability of how conclusions are reached, and AI literacy among the people using the outputs.

Can a tool make us "UNESCO-compliant"? No tool can, because the Recommendation is a set of principles, not a certification. What software can do is make the principles easier to honour — keeping decisions with humans, making AI involvement transparent, and making conclusions traceable to evidence.

Ready to run AI-assisted course evaluation that keeps humans in charge and every conclusion traceable? See how Koji for Education works.