The colonial period didn’t just redraw maps. It rewired how hundreds of millions of people thought about time, property, personhood, and authority. The coercive tactics were real — the violence, the forced conversions, the deliberate erasure of languages — and that damage didn’t disappear when the flags came down. But coercion was the visible layer. The deeper rewiring was infrastructural: schools, legal systems, languages of commerce, the default medium for contracts and credentials and courts. You didn’t have to believe the ideology to be shaped by it. You just had to participate in the economy.
The new infrastructure runs on language models.
What Gets Embedded
A large language model is, at its core, a compression of its training corpus. It learns patterns of reasoning, not just vocabulary. Ask it to help draft a legal argument and it will think in terms of individual rights, precedent, and procedural legitimacy — because that’s the tradition most of its training data assumes. Ask it to mediate a community dispute and it will probably reach for adversarial framing before restorative ones.
This isn’t a flaw in the usual sense. The model is doing what it was trained to do. But training data isn’t culturally neutral. Roughly 46% of websites indexed by Common Crawl — the primary raw material for most major LLMs — are in English. Mandarin accounts for around 5%. Arabic, Hindi, Swahili, and hundreds of smaller languages contribute scraps. The epistemological assumptions embedded in decades of English-language text — individual agency, linear time, property as discrete ownership, consensus through explicit agreement — are baked into the weights.
Deploy that model globally and you’re not exporting a tool. You’re exporting a cognitive default.
The Optimization That Doesn’t See It
The companies building these systems are not trying to homogenize global thought. That’s not their goal, and it’s not a useful frame for understanding why it might happen anyway.
Their targets are capability (does the model perform well on benchmarks?), safety (does it avoid harmful outputs?), and commercial viability (do users find it useful?). Cognitive diversity doesn’t appear on that list. Neither does preserving the epistemic distinctiveness of oral cultures, or minority legal traditions, or the embedded timekeeping logic of societies that organize time around seasons rather than calendars.
These things aren’t being targeted for erasure. They’re just not being optimized for survival. That distinction matters less than it sounds.
What Actually Gets Lost
The easy version of this concern is linguistic: endangered languages die faster when a powerful, fluent alternative is available in your pocket and there’s no economic incentive to maintain the original. That’s real. It’s not the deepest problem.
Different cultures have developed genuinely different solutions to hard coordination problems. Some Indigenous legal traditions encode long-run resource management constraints that Western property law structurally cannot. Several East Asian philosophical frameworks approach risk and obligation through relational webs rather than individual contracts — which produces different, not obviously worse, outcomes in collective action problems. Certain oral traditions carry embedded ecological knowledge that has no written form and cannot survive disruption of the social structures that transmit it.
None of this is romanticization. Some traditional practices are worth losing. The point is that cognitive diversity — the breadth of available frameworks for thinking about time, risk, fairness, and belonging — functions like biological diversity. A world that processes difficult problems through one epistemological dialect is less resilient, not just less interesting. When the standard framework fails, you want other frameworks available.
The Historical Parallel Has Limits
The colonial analogy is useful but imperfect. Physical colonization required territory and enforcement. Cognitive homogenization via AI infrastructure requires neither — it happens through adoption, because the tools are genuinely useful.
That’s harder to resist, not easier. You can decline a missionary’s invitation. Declining a tool that writes your legal filings, tutors your children, and handles your medical triage is a different calculation.
There’s no clean solution here. Localized models trained on specific cultural corpora help at the margins; so do multilingual training pushes. But these are downstream mitigations for an upstream structural condition: the organizations building the most capable AI systems are concentrated in a handful of cities, optimizing for a handful of metrics, drawing on a corpus that is not — and probably cannot be — a representative sample of human thought.
The second colonization isn’t coming from bad intentions. That’s what makes it worth thinking about carefully.