We Don’t Want to Miss This
I want us to slow down. Not just to avoid danger, but so we don’t accidentally build something smaller than this moment allows.
Please allow me to explain.
The usual case for slowing down is framed as a warning about danger. That warning is necessary — but incomplete.
But I think there’s another good reason to slow down: opportunity.
If humanity builds AI too quickly around the assumptions of the most powerful actors, we may lose the chance to build systems that see more broadly than any of us.
The real danger isn’t only that AI might harm us. The deeper danger is that we might build something smaller than what this moment makes possible.
Thesis
Unexamined biases — whether liberal, nationalist, corporate, technocratic, or development-first — risk two things at once: missing the big transformative benefits of AI, and getting the real harms wrong. A better approach means deliberately listening to competing viewpoints, testing our assumptions, and making the value trade-offs visible instead of pretending any one side is neutral.
What This Argument Is, and Is Not
This is not a brief against caution. Some AI risks are real and serious:
- Cyber abuse and manipulation
- Ubiquitous surveillance
- Autonomous weapons
- Medical errors
- Labor displacement
- Biosecurity threats
- Dangerous concentrations of power
The target here is not caution itself. The target is unexamined caution—the kind that treats its own assumptions as universal truth while treating other societies’ needs as secondary.
It is also not an argument that one civilization’s way should replace another. The goal is not to swap Western bias for Chinese, Russian, or corporate bias. The goal is to build AI systems and governance that can see the blind spots in all of them.
A Crucible Moment
We are deciding right now whether the logic inside global AI systems will harden the biases of the last century or become the first infrastructure built with real diversity.
If we get this wrong, we don’t just get biased chatbots. We lock in blind spots for healthcare, law, education, and climate planning for generations. We risk turning AI into a global “McDonald’s” of thought — fast, uniform, and stripped of the local wisdom that makes us resilient.
But if we get it even partly right, we could build systems less biased than any single culture that created them.
No civilization, company, or ideology has full ownership of sound judgment. Liberal protections, Eastern relational thinking, Indian pluralism, African Ubuntu, Indigenous frameworks, and Global South development priorities each reveal blind spots in the others. The point is to keep AI from inheriting the narrowness of whichever group happens to own the servers.
The Nature of Bias
Biases aren’t just individual mistakes. They’re baked into institutions, media, and now AI systems themselves.
Training data, RLHF annotators, system prompts, and retrieval filters all push AI toward the worldview of whoever builds and funds it. A 2025 study found that even models from Russia and China often favored dominant English-language narratives on historical events. These biases proved hard to remove with simple prompts but flipped easily when given identity instructions like “answer as a Chinese patriot.” That shows how fragile assumed neutrality really is.
This isn’t a uniquely Western problem. Any powerful actor — corporate, Chinese, American, or otherwise — tends to naturalize its own perspective.
Why This Matters
Unchecked bias distorts our view of what’s possible. It makes contingent outcomes look inevitable, hides alternative paths, and focuses our attention on familiar threats while missing others. It narrows the range of futures we consider legitimate.
When AI systems reflect mostly one set of assumptions, we risk building tools that work well for some and poorly for most of humanity.
Whose AI for All?
Different cultures bring different strengths. Māori data governance treats data as a cultural treasure. Ubuntu emphasizes community well-being over pure efficiency. Global South voices push for development-first approaches that prioritize accessible tools over heavy regulation. These are not just abstract cultural philosophies; they are active, alternative models proving that the dominant approach is not the only operational path.
But states are not cultures, and governments are not whole peoples. “Community-led” can still mean elite capture if we’re not careful. Real pluralism needs guardrails: it must protect minorities, dissenters, and the vulnerable.
Distinguishing Strategic Competition from Undue Prejudice
Much of what looks like bias is actually normal great-power competition. Export controls, coalitions, and narrative battles are standard statecraft.
The problem arises when any side treats its own system as the neutral baseline and assumes the worst motives from everyone else. The real test is simple: Are we citing specific evidence and weighing impacts across many groups? Or are we relying on worst-case stories that mainly protect our own interests?
The “race” framing itself can blind us. Not every advance by another country is a loss for us. Many nations see AI primarily as a tool for development, infrastructure, and catching up — not as an existential contest.
Precaution vs. Progress
The precautionary principle is rarely neutral. For people with limited healthcare, delaying AI diagnostics to avoid rare errors can cost lives today. For others, rapid rollout risks reinforcing inequality.
All major powers are embedding their values into AI — some openly, others while claiming neutrality. The challenge is to move past pretending any one framework is universal and instead build an ecosystem where different value systems can coexist, compete, and be tested in practice.
A Better Path Forward
True pluralism doesn’t mean averaging everything into a bland compromise. It means building systems that can show the trade-offs clearly: “This approach protects individual liberty but may weaken community cohesion. That one ensures stability but limits personal freedom.” Let humans in different contexts make their own choices.
We can do this through practical steps:
- Model cards that honestly disclose cultural and value limitations
- Tools that let users analyze issues through multiple frames (liberal rights, Ubuntu, development priorities, etc.)
- Broader evaluation panels and local data governance
- Compute access that doesn’t just favor whoever got there first
The strategic advantage belongs to those who build systems trusted across real human differences. Markets may not reward wisdom immediately, but over time, bias-aware systems will prove more adaptable and legitimate.
Final Reflection
Questioning whether the loudest anxieties reflect genuine risk, strategic self-interest, or cultural conditioning isn’t about picking sides. It’s about recognizing that all of us see the world from somewhere. Rigorous thinking requires engaging perspectives that challenge our defaults.
Over a hundred countries currently have little real voice in AI governance. That narrow base inevitably produces narrow outcomes.
The best safeguard isn’t less caution — it’s broader epistemic humility. Build processes that make bias visible, testable, and open to revision.
I consider Greeks to be master restaurateurs, especially because they know better than to only serve Greek food. AI should be no different. Cultural diversity, in all its forms, makes our world richer and more resilient.
What are your thoughts?