UNESCO's "Recommendation on the Ethics of Artificial Intelligence", adopted in November 2021 by its 193 member states, holds Google's top spot for almost every Spanish search related to AI ethics. Ten principles. All ten lovely. All ten empty. Transparency, fairness, human autonomy, harm prevention, responsibility, privacy, social benefit, sustainability, accountability, inclusion. Who's going to be against any of this? Nobody. That's why, for exactly that reason, they mean nothing operationally: they don't land on a concrete case, they don't set thresholds, they don't prioritise when two principles clash, they don't force anyone to do anything in particular. They're a document made to be signed, not to govern.
Today we're talking about the pontificate of ignorance. It isn't that it takes the chair, as we saw in the previous article on the broken public conversation. It's that it pronounces from an institutional pulpit, marks a route to salvation and fills our SERP with edifying bullets without understanding the sector it claims to govern. Today we examine that route to salvation, to see whether it leads anywhere.
What the Recommendation actually says
It's worth reading. Not the executive summary, not the pretty leaflet UNESCO publishes for journalists, but the whole text. It's 38 pages. You'll find it in Spanish at unesdoc.unesco.org. And when you read it calmly, what you find is a perfectly drafted document, with an obvious genealogy in earlier human-rights documents, with careful diplomatic vocabulary and an internal structure in four parts: scope of application, aims and objectives, values and principles, and areas of policy action.
The part quoted in headlines is the values-and-principles one. The four great values: respect for human dignity, the flourishing of ecosystems, diversity and inclusion, and peace. The ten operational principles: proportionality, safety and no-harm, fairness and non-discrimination, sustainability, privacy and data protection, human oversight and decision, transparency and explainability, responsibility and accountability, awareness and education, and adaptive multilateral governance.
Read them one by one. Is there a single one you object to? Hard. They're statements so general that disagreement becomes impossible. Whoever says "I'm against fairness and non-discrimination in AI" is automatically out of the conversation. It's what in formal logic is called an empty proposition: so broad it excludes nothing, and therefore informs nothing.
And that's the first trap. A document signed by 193 countries unanimously can only have that content. If it had operational content — if it said, for instance, "all models trained with more than 10²⁵ floating-point operations must undergo mandatory adversarial evaluation before deployment" — one of the 193 countries would have blocked it, because that would complicate life for its national industry or its intelligence services. Unanimity is only possible at the price of vagueness. And vagueness is what we have.
The format is useless and it's worth saying so
Abstract universal principles don't resolve trade-offs. And trade-offs are what, in practice, determine whether an AI tool is ethically acceptable or not.
When a company wants to deploy a facial-recognition system at the airport, there's an immediate conflict between "privacy" and "security". Both are UNESCO principles. The Recommendation says something like "they must be weighed appropriately". It doesn't say who weighs, how it's weighed, what happens if the two weights are equal, what court resolves the discrepancy, what quantitative threshold of privacy justifies what level of security. It doesn't say it because it can't: if it did, it would stop being a document signable by 193 countries and become a regulation, with all the national objections that generates.
When an AI-based medical system produces a potentially useful but opaque diagnosis — there's no way to explain how it produced it — there's a conflict between "transparency" and "social benefit". The Recommendation says both are important. It doesn't say what to do when you can only have one.
When a model is trained on massive data scraped from the internet without individual consent, there's a conflict between "privacy and data protection" and "the economic sustainability of the sector". It's not trivial. The Recommendation doesn't resolve it.
The function of an operational ethical framework is exactly that: to say what to prioritise when principles clash, who decides and under what procedure. The UNESCO Recommendation doesn't do that. It enumerates principles without hierarchy, without thresholds, without procedures and without sanctions. It's an aesthetic document, not a normative one.
The contrast with the AI Act
Let's compare it with a text that does govern: Regulation (EU) 2024/1689, also known as the AI Act, approved by the European Union in 2024 and applicable progressively between February 2025 and August 2026. It isn't perfect. It has gaps, it has justified criticisms, it has optimistic deadlines. But it's an operational text, and it's worth seeing why.
The AI Act classifies AI systems into four levels of risk: unacceptable risk, high, limited and minimal. The classification isn't rhetorical: each level has different obligations and different sanctions. Unacceptable-risk systems — social scoring, subliminal manipulation, real-time remote biometric identification in public spaces with limited exceptions — are prohibited. Full stop. Not "they must be weighed appropriately". Prohibited.
High-risk systems — medical diagnosis, credit scoring, recruitment systems, critical infrastructure — must meet concrete requirements before deployment: conformity assessment, technical documentation, transparency toward users, operational human oversight, robustness and cybersecurity, registration in a European database.
General-purpose models — the large foundation models, GPT, Claude, Gemini, Llama, the Chinese ones — have a separate regime. Obligations of technical documentation, instructions for use, compliance with the Copyright Directive, publication of a summary of the training data. When the model is trained with more than 10²⁵ floating-point operations (the current technical frontier), it's considered to present "systemic risk" and then there are additional obligations: model evaluation, adversarial testing, notification of serious incidents, reinforced cybersecurity protection.
And, above all, there are sanctions. The most serious infringements — prohibited practices, breach of the GPAI regime — can reach 35 million euros or 7% of annual worldwide turnover, whichever is greater. Other infringements, up to 15 million or 3%. Incorrect information to the authority, up to 7.5 million or 1%. That's no longer aesthetics. That's administrative law with real economic consequences.
Why doesn't the AI Act appear at the top of Google when a user searches for "AI ethics"? Because it doesn't fit in bullets. It's 113 articles, 180 recitals and thirteen technical annexes. Precise legal definitions. It doesn't boil down to ten pretty principles. It requires expert reading. UNESCO offers ten lines that fit on a corporate infographic. The AI Act offers an entire normative edifice that requires specialised lawyers. The SERP chooses the infographic.
The real function of the UNESCO document
Here it's worth honestly asking what the Recommendation is for, since it isn't for governing the sector. Because it is for something. If 193 countries signed it unanimously, someone got something.
It serves, first, for internal legitimation. Any signatory state can now cite the Recommendation when its citizens ask what it's doing on AI ethics. "We've signed the UNESCO Recommendation" is an administratively valid answer that implies almost nothing operationally but offers the appearance of action. It's institutional marketing applied to government. It replaces a real policy with a symbolic signature.
It serves, second, to fill the international discursive gap without having to commit to anything binding. The great technological powers — the United States, China, the EU members — didn't want, in 2021, a binding international treaty on AI. They wanted to keep competing without restrictions imposed from outside. A non-binding recommendation lets them sign without losing degrees of freedom. Diplomatically it's perfect. Substantively it produces nothing.
It serves, third, as a symbolic reference in corporate documents. When a SaaS company publishes its "AI code of ethics", it can cite the UNESCO Recommendation as a reference framework, which lends an appearance of seriousness without obliging the company to anything operationally. It's like when a fashion brand cites respect for human rights on its website without explaining anything about the conditions of its production chain.
It serves, fourth, to feed the ethics-consulting ecosystem. An entire sector of consultancies offers "ethical AI audits" based on the UNESCO principles. Those audits produce reports, the reports produce invoices, the invoices produce jobs, the jobs produce defenders of the framework. It's a self-sustaining ecosystem around a document with no operational content.
What it isn't for is preventing a concrete system from being deployed harmfully, sanctioning whoever deploys it, resolving a real conflict between principles, giving the affected citizen tools of defence, or forcing the developer to do something they don't want to do. None of that is in the Recommendation. And since none of that is in it, the Recommendation doesn't produce any of it either.
What would actually move the debate
Let's imagine what a text of ethical AI governance that actually worked would look like. It doesn't have to be perfect. It only has to be operational. A few identifiable features.
It would have an explicit hierarchy of principles. When "privacy" and "security" clash, the text would say which prevails, under what conditions, with what limited exceptions and with what review procedure. It wouldn't say "they must be weighed appropriately". It would say something uglier and more useful.
It would have numerical thresholds. "A model trained with more than X FLOPs must undergo a safety evaluation before deployment." "A facial-recognition system with an error rate above Y percent in demographic categories Z cannot be used in domain W." Numbers are what separate operational regulation from generic ethical discourse.
It would have defined and proportionate sanctions. Administrative sanction, economic sanction, deployment suspension, market ban. Without sanctions, no principle is met. With sanctions, principles produce behaviour.
It would have complaint and defence procedures accessible to the citizen. When a citizen believes an AI system has harmed them — they've been denied a loan, a visa, a job, a treatment — they should have a clear administrative procedure to claim, with deadlines, with defined burdens of proof, with a named competent authority.
And it would have mandatory periodic review with data. Every two years, signatory states would have to submit a report with concrete compliance metrics. If the metrics don't improve, there are consequences. Peer review is what keeps any normative framework alive.
None of this is in the UNESCO Recommendation. It is, partly, in the AI Act — which, not by chance, is far more unpopular on the Spanish SERP. And it is, in more mature form, in the national regulatory regimes emerging in the United States, the United Kingdom, Brazil, Singapore. Those texts matter operationally. The Recommendation matters symbolically. The difference between the two things, in public policy, is everything.
Why there's no real debate about ethical governance
It's worth closing by speculating on why the public conversation about ethical AI governance stays on the UNESCO Recommendation instead of discussing the AI Act, US regulation, the Chinese governance models, the African and Latin American proposals. There are several hypotheses.
One: the Recommendation is available, it's translated into multiple languages, it's actively promoted by its issuing body. The operational texts are less accessible, worse translated, more buried.
Two: the Recommendation is what SaaS companies can cite without risk. The AI Act, by contrast, obliges them. A company prefers the public debate to centre on the document that praises it (signing it feels like good citizenship) instead of the document that fines it.
Three: signatory states prefer their citizens to believe the problem is solved by the Recommendation, so they don't have to do tough national regulation. Inaction shelters under the signature.
Four: the ethics-consulting ecosystem lives off the Recommendation. It has no interest in operational regulation that would reduce its market to measurable technical audits, where its added value would be smaller.
Five: the public conversation about AI in Spanish, as we saw, is broken at the source. Whoever should translate the AI Act, comment on it, explain it, criticise it, do rigorous science communication — specialised journalism — doesn't have the capacity. Whoever does have the capacity — the corporate offices — prefers to talk about UNESCO.
The result is a scenario where the Spanish-speaking reader searching for "AI ethics" runs into bullets about dignity, autonomy and diversity. They come away convinced they've understood the debate. And they miss the real conversation, which happens in another language, in another register and with other consequences.
The UNESCO Recommendation isn't a fraud. It's an honest document within what international diplomacy allows to be produced. The fraud is presenting it as governance when it's aesthetics. Knowing how to read the difference is the first step to not confusing good intentions with regulation. And, without operational regulation, UNESCO's good intentions are going to get as far as its earlier recommendations on cultural heritage, education or science got: to producing a pretty document everyone cites without almost anything changing.
Quick definitions
- UNESCO Recommendation on the Ethics of AI (2021): a document adopted by the 193 member states stating four values and ten principles for the ethical development and use of AI. Non-binding, without sanctions, without operational thresholds.
- AI Act (Regulation EU 2024/1689): European regulation, operational between 2025 and 2026, classifying AI systems into four risk levels with specific obligations and sanctions. Directly applicable in the 27 member states. Sanctions of up to 35 million euros or 7% of worldwide turnover.
- GPAI (General Purpose AI): a category introduced by the AI Act for the large foundation models. They have a specific regime, stricter when they exceed 10²⁵ FLOPs of training compute.
- Soft document vs hard document: in international law, soft law (recommendations, declarations, principles) versus hard law (binding treaties, regulations, statutes). The UNESCO Recommendation is soft law. The AI Act is hard law. The difference between the two is the difference between aesthetics and government.
Referencias
- UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. Adopted by the 193 member states in November 2021. https://www.unesco.org/en/articles/recommendation-ethics-artificial-intelligence
- Regulation (EU) 2024/1689 of the European Parliament and of the Council, laying down harmonised rules on artificial intelligence. Consolidated text (113 articles, 180 recitals) in the Official Journal of the EU. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
- AI Act, Article 99 (penalties). Bands of 35 M €/7%, 15 M €/3% and 7.5 M €/1% of annual worldwide turnover. https://artificialintelligenceact.eu/article/99/
- Stanford HAI (2026). AI Index Report 2026, chapter on global regulation.
- Floridi, L. (2023). The Ethics of Artificial Intelligence. Oxford University Press.
- Wallach, W. & Allen, C. (2009). Moral Machines. Oxford University Press.
- Bostrom, N. (2014). Superintelligence. Oxford University Press.
También te interesa
- Why the critiques of AI aren't made by real experts
- How did the public conversation about AI break?
- Why reading IBM on AI is a waste of time
No comments yet
No comments yet. Be the first.
Leave a comment