AI companies are building transformative technology while telling the world that someone else should really be in charge of deciding whether it happens. That contradiction sits at the heart of a recent public dialogue between Jack Clark, head of policy at Anthropic, and Samuel Kimbriel of the Aspen Institute, which exposed a tension that will define AI governance for years to come: the choice to build powerful AI was made in private by a handful of labs and investors, but the choice about how to regulate it is being offered to the public.

Why Does the Oppenheimer Comparison Matter?

The conversation between Clark and Kimbriel kept circling back to a moment from 1954, when physicist J. Robert Oppenheimer testified before the Atomic Energy Commission about his support for the hydrogen bomb. Oppenheimer hadn't changed his mind about the morality of thermonuclear weapons. Rather, he had encountered a new bomb design that was, in his words, "technically sweet." That elegance was enough. As he explained to the commission: "When you see something that is technically sweet, you go ahead and do it, and you argue about what to do about it only after you have had your technical success".

That formulation captures something uncomfortable about how transformative technologies actually get built. The moral accounting comes later. The technical achievement comes first. And by the time the public is asked to weigh in, the technology already exists, reshaping the world around it.

Clark acknowledged this dynamic directly. We regulate cars, toothbrushes, and nuclear weapons, he pointed out. But in each case, someone built the thing first. "Social media ran an uncontrolled experiment on the world," he said. "We all now think and talk a bit differently because of social media. That was a choice. We can choose things to be different." Yet the choice to build social media wasn't put to a public vote either.

Clark

What Question Did the Audience Actually Ask?

Near the end of the dialogue, a young woman in the audience posed the question that Clark's framing had been skirting: "Every frontier lab now admits the technology carries enormous risk, even existential risk. So my question is, what gives you, Anthropic, and the rest of the frontier labs the right to continue building something that could destroy everybody, when none of us can actually opt out of it?"

Clark did not dismiss the question. But he also did not answer it directly. Instead, he reframed it. The choice shouldn't rest with companies like Anthropic, he said. Rather, "outside compliance, regulatory, testing and verification systems" should decide when each lab was allowed to advance further. He noted that governments were already moving faster than expected, with the US and UK building testing agencies whose tools sometimes exceeded the companies' own capabilities.

It was a graceful answer, but it contained an implicit concession: Asked what gives his company the right to build something that could destroy everybody, the head of policy at a leading AI lab did not claim that right exists. He simply described a system where someone else would eventually take control. Meanwhile, Anthropic and its peers continue building at the frontier, as fast as science and computing power allow.

How Are Top AI Researchers Voting With Their Feet?

The tension between building and governing is playing out in real time across the AI industry. Just days apart in late June 2026, two of Google DeepMind's most celebrated researchers announced departures that sent shockwaves through the sector.

Noam Shazeer, who helped build Google's LaMDA chatbot system in 2021 and later cofounded the viral startup Character.ai, announced he was joining OpenAI. Days later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for creating AlphaFold (an AI system that solved a 50-year challenge in predicting protein structures), announced he was joining Anthropic.

The departures raised a question that industry watchers are asking openly: Is Google DeepMind slipping in the AI race? The lab's latest models, Gemini 3.5 Flash and Gemini 3.1 Pro, often rank outside the top five on various AI benchmark leaderboards, trailing models from Anthropic and OpenAI, as well as Chinese labs like Zhipu AI and MiniMax. Meanwhile, its pace of model development appears to be lagging. Google announced Gemini 3.5 Pro in May with a June release target, meaning roughly four months between major frontier releases. By contrast, Anthropic released two significant Claude Opus updates in that same period and debuted an entirely new class of models called Mythos, which leads in long-range autonomous task completion, particularly in coding and cybersecurity domains.

For researchers like Jumper, the move appears driven by scientific opportunity rather than compensation. Anthropic's CEO Dario Amodei recently told Bloomberg that the company intends to expand its work in biology, and Jumper's hiring is clearly part of that plan. The signal is clear: the labs perceived as moving fastest and taking the biggest scientific bets are winning the talent war.

Why Does Google DeepMind's Culture Matter to the AI Risk Debate?

Conversations with current and former Google DeepMind employees reveal a pattern that Shazeer himself had criticized years earlier: the lab is burdened by its size, with a culture described as bureaucratic and risk-averse. Google, with billions of users and fiduciary duties to public shareholders, cannot afford the same kinds of bets that venture-funded startups like OpenAI and Anthropic can take.

This structural difference matters enormously to the AI risk conversation. The labs that are moving fastest, taking the biggest risks, and attracting the world's top talent are precisely the ones that are also most willing to acknowledge existential risks and call for stronger regulation. Anthropic's leadership has been notably candid about the dangers. CEO Dario Amodei published a blog post calling for government authority to legally block or reverse the deployment of frontier AI models that fail safety tests on threats like cyberhacking and bioweapons.

Yet those same labs are also the ones building the technology at maximum speed, while waiting for regulatory systems that do not yet fully exist. The regulation of AI, Clark noted, remains mostly the stuff of blog posts.

How to Understand the Core Tension in AI Governance

• The Build-First Problem: Transformative technologies like AI, social media, and nuclear weapons are typically built first by private actors, then regulated after they already exist and have reshaped society. The public never votes on whether the technology gets created, only on what to do about a world that already contains it.
• The Talent-Driven Acceleration: The labs moving fastest and taking the biggest risks are attracting the world's top researchers, creating a competitive dynamic where caution is punished. Google DeepMind's more conservative approach is losing talent to Anthropic and OpenAI, which are perceived as more scientifically ambitious.
• The Regulation Lag: Even as AI companies acknowledge existential risks and call for stronger government oversight, they continue building at maximum speed. The regulatory systems that would actually enforce safety standards are still being designed, while the technology they're meant to govern is already loose in the world.

Clark's answer to the young woman's question ultimately revealed the deepest issue: the people building AI are not claiming the right to decide whether it gets built. They are simply building it anyway, while hoping that someone else will eventually take control. That someone, they suggest, should be governments. But governments are moving slowly, and AI labs are moving fast.

The question of who gave AI companies the right to build the future remains unanswered. But the future is being built regardless, by researchers who are compelled by the technical beauty of the work, by the competitive pressure to move faster than rivals, and by the genuine belief that AI will solve problems that nothing else can. Whether that turns out to be wisdom or recklessness may depend on whether the regulatory systems Clark describes actually materialize before the technology reaches the point where regulation becomes impossible.