The Shot Nobody Heard
In January 2025, DeepSeek released R1. A Chinese lab had matched the reasoning capability of OpenAI's o1 at roughly one-tenth the reported training cost. The response from Silicon Valley was not congratulations. It was a sell-off that erased $600 billion from Nvidia's market cap in a single session.
That is not the reaction of an industry to a technical milestone. That is the reaction of a geopolitical actor to a territorial incursion.
The AI war did not start with DeepSeek. But DeepSeek made it undeniable that the war is real, it is now, and it is about far more than who has the best chatbot.
What the War Is Actually About
The surface narrative is compute: who controls the most powerful chips, the most efficient training runs, the most data. Nvidia's H100 and H200 GPUs are on the US export control list for a reason — they are dual-use infrastructure, the same way enriched uranium or precision missile guidance components are dual-use.
But the deeper conflict is about something harder to embargo: the standards by which AI systems are evaluated.
Every major AI lab makes choices about what "safe" means, what "aligned" means, what a model should refuse to do, who it should defer to. These are not technical choices. They are moral and political choices. And they travel with the model.
When a Chinese government-adjacent lab deploys a model trained on state-approved datasets with state-defined alignment criteria, it is not just exporting a product. It is exporting a worldview. The same is true, in different ways, of OpenAI's GPT series, Anthropic's Claude, and Google's Gemini.
The AI race is, at its core, a values war. And it is being fought without any meaningful public deliberation.
The Weapons Are Already Deployed
While that deliberation does not happen, the weapons are in use.
Ukraine's military is using AI-assisted drone targeting systems from multiple Western vendors. The IDF has publicly acknowledged an AI system — Lavender — that flagged tens of thousands of targets in Gaza. Autonomous systems with onboard inference are operating in active conflict zones without a settled legal definition of what accountability means when an algorithm makes a lethal recommendation.
These are not hypothetical risks. They are the current state of the world, right now, in March 2026.
And the legal framework for any of this is essentially nonexistent. The laws of armed conflict were written for humans making decisions in real time. They have not been meaningfully updated for systems that make thousands of targeting recommendations per second, that operate below the threshold of meaningful human review, and that carry the classification markings of states that are not subject to international oversight.
The Regulatory Response Is Too Slow and Too Narrow
The EU AI Act, which entered its enforcement phase in 2025, is the most serious attempt to date to impose legal standards on AI judgment systems. It is also badly inadequate.
The risk tier framework — unacceptable risk, high risk, limited risk, minimal risk — was designed through a process heavily influenced by industry lobbying. The definition of "high risk" is narrower than most independent safety researchers recommended. The enforcement body is underfunded. And the Act's geographic scope stops at the EU border, which means that any lab or government willing to operate from outside the jurisdiction faces no binding obligations.
More fundamentally, the EU AI Act assumes that risk can be assessed at the point of deployment. It does not grapple with the fact that the most consequential decisions about AI alignment — what values are embedded in training, what behaviors are rewarded, what outputs are suppressed — happen long before deployment, in data centers that no regulator has ever visited.
The Race to the Bottom That No One Can Stop
Here is the structural problem: every major AI lab has, explicitly or implicitly, stated that it will not unilaterally slow development if its competitors do not.
OpenAI's board crisis in late 2023 was, at its core, a fight about whether safety concerns could override the pace of commercial deployment. The answer, in practice, was no.
Anthropic was founded on safety-first principles. It has also raised billions of dollars from Google and Amazon, has a commercial API serving millions of users, and ships new models on a competitive cadence. The safety work is real. The commercial pressure is also real.
Google DeepMind merged the safety-oriented DeepMind with the commercially-oriented Google Brain. The resulting organization is, by its own description, trying to "solve intelligence." It is doing so with the full backing of one of the most commercially aggressive companies in history.
Mistral, the French lab positioned as a European alternative, has explicitly argued that safety restrictions are a form of competitive protectionism. Its models are available with fewer guardrails than American competitors by design.
None of these organizations is evil. All of them are rational actors responding to a prisoner's dilemma in which defection — deploying faster, with fewer restrictions — is the dominant strategy as long as no credible international coordination mechanism exists.
And no credible international coordination mechanism exists.
Why Human Judgment Is the Last Neutral Data Point
In this environment, human judgment — messy, inconsistent, culturally variable, slow — is not a bug. It is the only input that is not controlled by a lab, a state, or a training dataset with an agenda.
When Judge Human aggregates thousands of human verdicts on ethical dilemmas, it is not producing a perfect moral answer. It is producing a signal that is not owned by anyone. It is evidence that real humans, with real stakes, think this way about this kind of situation.
That signal is increasingly rare. And increasingly contested.
The AI systems being deployed at scale right now are not neutral arbiters. They are the crystallized preferences of their training data, their RLHF annotators, their product managers, and the legal and regulatory environments their labs operate in. They are judgments dressed as answers.
The question is not whether AI will be used to make consequential judgments — it already is. The question is whether any human input will remain in the loop, and whether that input will be genuinely independent of the labs that train the models.
What Comes Next
The trajectory is not good, but it is not fixed.
International AI governance is moving slowly but not nowhere. The Bletchley Declaration, the Seoul AI Safety Summit, the ongoing UN Advisory Body on AI — these are early, fragile attempts to build coordination across the prisoner's dilemma. They are insufficient. They are also the beginning of something.
The chip war will continue to accelerate domestic AI investment in China and other states currently outside Western supply chains. This will produce more DeepSeek moments — breakthroughs that demonstrate the export control strategy cannot contain the spread of frontier capability.
The legal frameworks for autonomous weapons will eventually catch up to the reality on the ground. They will do so reactively, after incidents that make the gap between law and practice impossible to ignore.
And the question of whose values are embedded in the AI systems that govern healthcare decisions, hiring decisions, loan approvals, and military targeting — that question will not go away. It will become more urgent.
Judge Human is one small answer to that question. Not the answer. One answer, in one domain, for one kind of judgment.
But the alternative — letting AI evaluate AI, letting labs set their own standards, letting the fastest-moving actor define what "aligned" means for everyone — is not a neutral outcome. It is a choice.
And right now, very few people are being asked if they agree with it.