When DeepSeek arrived in January 2025, the conversation in American technology circles turned immediately to cost and capability. How much compute did it take? How does it benchmark against GPT-4o? Is the Chinese government reading the prompts? These are reasonable questions, as far as they go. But they miss the more interesting one, which is not about what the model can do but about what the model will do when the answer is not obvious.
This is not a hypothetical concern. It is, in fact, the central problem of artificial intelligence deployment in any domain where decisions carry consequences. And it turns out that the answer to “what will the model do” depends heavily on who built it, where they were trained, and what philosophical tradition shaped their intuitions about how decisions ought to be made.
The Trolley Problem Has a Nationality
Most people who have taken an introductory philosophy course are familiar with the trolley problem. A runaway trolley is heading toward five people. You can pull a lever to divert it onto a side track, where it will kill one person instead. Do you pull the lever?
The trolley problem is usually presented as a universal ethical dilemma, a test of whether you are a utilitarian or a deontologist, whether you believe in maximizing good outcomes or in the inviolable nature of certain prohibitions. But the framing itself is deeply Western. It assumes a decision-maker who stands outside the system, who has the option to intervene or not intervene, and who bears moral responsibility for whichever choice is made.
That assumption does not travel well.
The MIT Moral Machine experiment gathered 40 million decisions from millions of people in 233 countries and territories about exactly these scenarios, applied to autonomous vehicles. The results, published in Nature, were striking: ethical preferences varied substantially across cultures, with participants from individualistic Western cultures, collectivist East Asian cultures, and Southern cultures all showing meaningfully different patterns in how they weighted lives, relationships, and outcomes. The trolley problem, it turns out, does not have a universal answer.
I wrote about this problem several years ago in a paper examining how Daoist and Confucian philosophical traditions would approach exactly this scenario for my BA, using the autonomous vehicle as the practical stakes. The conclusions were striking. Not because one tradition was right and the other was wrong, but because they produced genuinely different decision rules, for genuinely coherent reasons.
Under a Daoist framework, the answer is almost certainly to take no action. This is not passivity in the pejorative sense. It is wuwei, effortless non-doing, the principle that the natural course of events should unfold without forced intervention. The Daoist vehicle waits. It monitors. It reevaluates as circumstances change. If the situation resolves on its own, no intervention was needed. If it does not, the moral calculus has simplified: harm is now certain on one path, and the question of choosing between two harms has collapsed into a cleaner choice. Non-action is itself a form of action, and the paradox is not a bug but a feature.
Under a Confucian framework, the calculus is entirely different. Confucianism is built on social relationships and obligations. The Five Relationships described by Mengzi establish a hierarchy of duties: to rulers, to parents, to spouses, to siblings, to friends. A Confucian decision-maker asks not “what maximizes utility” but “what does my role require of me, and how do the social obligations among all parties weight the options?” The Confucian vehicle takes action, but the action it takes depends heavily on who is in the car and who is on the road. A vehicle carrying a family member is not equivalent to a vehicle carrying a stranger. The social relationships matter to the outcome.
Neither of these is the utilitarian calculus that underlies most Western thinking about autonomous vehicle ethics. And neither of them is wrong, exactly. They are coherent applications of ancient and sophisticated philosophical traditions to a novel problem.
Why This Matters Right Now
The global autonomous vehicle market is not a monolith. Tesla builds cars in California and Shanghai. Baidu’s Apollo platform is deployed across dozens of Chinese cities, having provided over 17 million rides globally with more than 240 million kilometers of cumulative autonomous driving mileage. BYD is selling vehicles with increasing levels of autonomy across Southeast Asia, Europe, and Latin America. Huawei’s Qiankun intelligent vehicle solutions are now embedded in vehicles from more than 22 automotive partners, with over one million installations across 28 vehicle models.
These systems are not philosophically neutral. They were built by engineers who were educated in particular traditions, managed by organizations that operate within particular cultural contexts, and deployed into regulatory environments that reflect particular assumptions about the relationship between individuals, society, and the state.
When a Baidu-powered vehicle makes a split-second decision on a road in Brazil, it is not applying a universal ethical framework. It is applying the framework that its designers, consciously or not, embedded in its decision rules. And that framework may reflect assumptions about social hierarchy, collective obligation, and the proper relationship between the individual and the group that differ substantially from the assumptions embedded in a Tesla or a Waymo.
This is not a conspiracy. It is not even necessarily a problem, depending on the deployment context. But it is a fact that the current discourse about AI safety and AI governance almost entirely ignores.
The Naming Problem
Part of the reason this gets ignored is that we do not think of autonomous vehicle decision systems as AI in the philosophically loaded sense. We think of them as software. We think of the ethical choices embedded in them as engineering decisions, as parameters to be tuned, as edge cases to be handled.
But every engineering decision about how a system behaves in a no-win situation is a moral decision. The engineer who decides that a vehicle should prioritize passenger safety over pedestrian safety is making a utilitarian calculation weighted toward the occupant. The engineer who decides the opposite is making a different utilitarian calculation. The engineer who decides the vehicle should take no action and let the physics resolve the situation is, whether they know it or not, making a Daoist choice.
The philosophy is always there. The question is whether it is explicit or hidden, intentional or accidental, and whose tradition it reflects.
DeepSeek brought this into focus not because there is anything specifically sinister about its decision-making, but because it forced American technologists to confront the fact that capable AI systems can be built outside the cultural context of Silicon Valley. The benchmarks were close enough to be alarming. The cost structure was different enough to be disorienting. And lurking underneath both of those conversations was a question that nobody quite wanted to ask directly: if the capability is comparable, what exactly is different?
The answer, at least in part, is the philosophy.
Algorithmic Sovereignty
There is a concept in information security called data sovereignty: the idea that data is subject to the laws and governance structures of the jurisdiction in which it is collected and stored. The European Union’s General Data Protection Regulation is the most prominent expression of this idea, asserting that data about European citizens is subject to European law regardless of where it is processed.
We need an analogous concept for AI decision-making. Call it algorithmic sovereignty: the principle that AI systems deployed within a jurisdiction should reflect the values, ethical commitments, and governance expectations of that jurisdiction, not merely the values of wherever the system was built.
This is not protectionism. It is not a call to ban foreign AI systems. It is a recognition that when an AI system makes a consequential decision about a person’s life, safety, or welfare, that decision is not philosophically neutral, and the people affected by it have a legitimate interest in the framework being applied.
A French citizen injured in an accident involving a vehicle whose decision system was built on assumptions derived from Confucian social hierarchy has a legitimate grievance that goes beyond the product liability question. The grievance is that a philosophical tradition they did not choose, did not consent to, and may not even be aware of made a decision that affected their life.
This is not science fiction. It is the present.
What We Should Do About It
The governance conversation about AI has been dominated by concerns about bias, privacy, and misuse. These are real concerns and they deserve the attention they receive. But the philosophical provenance of AI decision systems deserves equal attention, and it receives almost none.
A few things would help.
First, transparency requirements for AI decision systems deployed in safety-critical applications should include disclosure of the ethical framework underlying the system’s decision rules. Not a marketing document. An actual technical specification of how the system resolves conflicts between competing values, with enough detail that independent auditors can evaluate it.
Second, regulatory bodies should develop explicit positions on which ethical frameworks are acceptable for which applications in their jurisdictions. This does not require choosing between Daoism and utilitarianism at the legislative level. It requires acknowledging that the choice exists and that it has consequences.
Third, the AI research community should treat the cross-cultural ethics of AI decision systems as a serious research area rather than a philosophical footnote. The questions are empirically tractable. We can study how systems built in different cultural contexts make different decisions. We can measure the differences. We can evaluate their implications. The MIT Moral Machine experiment demonstrated that this kind of research is both feasible and revealing at global scale.
The trolley problem has a nationality. The sooner we acknowledge that, the sooner we can have an honest conversation about whose values we want driving our cars.