Every time a group of governments, corporations, or experts gathers to draft ethical principles on artificial intelligence, that predictable moment arrives: someone insists they should be binding, and the tension builds. In the end, after endless debate, they opt to make them aspirational instead. It always plays out the same way.
The Chapultepec Principles, adopted under the OECD framework and debated in hemispheric forums with strong Latin American presence, follow the same tendency. They include solid ideas. They also omit aspects that say more about their limits than their strengths.
It's worth criticizing them seriously, especially because they represent a coherent regional contribution to the global debate on ethical AI. That means pointing out their failures and understanding why similar frameworks fail everywhere. Declarations tend to sound good. They rarely change anything.
These principles highlight key values: transparency, non-discrimination, accountability, privacy, and security. The flaw isn't in the content. It's in the silence on implementation, verification, and consequences for non-compliance. Just compare them to the OECD's 2019 AI Principles, the European Union's AI Act, the White House guidelines for responsible AI, or the Beijing principles. All of them mention transparency. All of them promote fairness. All of them, to a greater or lesser degree, remain aspirational without concrete enforcement tools. That convergence in language should reassure us, but it actually should raise alarm: if everyone repeats the same thing with minimal variation, chances are no one is truly moving forward.
Where Chapultepec shines is in its approach to data sovereignty and inclusive development from a Latin American perspective. It recognizes that AI doesn't emerge in a neutral vacuum, but in realities of inequality where data from vulnerable groups gets extracted without generating local value. That's convincing. The European AI Act, by contrast, starts from an economy with solid foundations and presupposes institutions that most countries in the region lack. Chapultepec at least makes that gap visible.
However, a good diagnosis doesn't guarantee effective solutions. That's where the structural obstacle that Le Bon and Dunbar, from different angles, had already anticipated comes into play.
Gustave Le Bon explained over a century ago how human crowds, when acting collectively, abandon individual reasoning and fall under charismatic leadership or simplified narratives. Ethical documents on AI are the fruit of institutional crowds: large committees, mass consultations, searches for consensus. The result tends to be the lowest common denominator. No one rejects transparency. No one opposes fairness. But in texts built on broad consensus, those concepts get diluted until they cover everything and nothing at once.
Robin Dunbar adds another layer. His studies on the cognitive limits of human networks show that effective coordination — the kind that produces measurable change — arises in small groups bound by trust. Not at conferences with hundreds of delegates. Not in agreements signed by dozens of nations with mismatched legal systems. The teams that achieve real progress operate within Dunbar's range: five to fifteen people, with clear goals, precise metrics, and direct accountability. A global ethical document that ignores local coordination violates that basic human dynamic.
This problem isn't unique to Chapultepec. Organizations with detailed ethical policies frequently show more serious biases in practice than those with no formal policy at all. The ethical text functions as a badge of good conduct, not as an operational guide. What's missing isn't declared principles; what's missing is the bridge between principle and practice.
The European AI Act represents, for now, the most concrete attempt to build that bridge. It defines risk categories, imposes clear prohibitions, and establishes fines with real figures. It has its flaws: it regulates the present while technology keeps advancing, and it demands audits that don't yet scale. Even so, it addresses enforcement. Chapultepec, like other hemispheric frameworks, stops at declared intentions.
What's missing is an intermediate level between the global and the local. A Dunbar-scale protocol for AI ethics would be a practical test: teams of no more than fifteen members, with technical, community, and legal profiles, turning each principle into verifiable indicators for a specific domain. Not for the entire hemisphere. For a city, a sector, or an institution. With reviews every six months and published results.
This approach accepts what Le Bon intuited: crowds don't execute, people do. And it validates what Dunbar proved: coordination has cognitive limits that declarations alone can't overcome. The challenge of AI ethics isn't a scarcity of principles. It's an abundance of them without protocols built at human scale.
Chapultepec offers a solid starting point for the region. But the second step — the one no one writes because it demands commitments, deadlines, and the risk of public failure — is what determines whether anything actually changes. Until that arrives, the principles will remain mere intentions on paper, ignored by the very AI systems they claim to regulate.
Stones don't lie, but historians sometimes do.
Sources:
1. OCDE Principles on Artificial Intelligence (2019) — oecd.org/digital/artificial-intelligence
2. European Union Artificial Intelligence Act (2024) — eur-lex.europa.eu
3. Dunbar, R. (1992). "Neocortex size as a constraint on group size in primates." Journal of Human Evolution
4. Le Bon, G. (1895). Psychologie des foules — modern editions available in Spanish as Psicología de las masas
5. NIST AI Risk Management Framework (2023) — nist.gov/artificial-intelligence
