One of the sharper arguments in AI policy right now is not about whether the technology is dangerous. It is about who benefits when governments step in to manage the danger. This week Anthropic’s chief executive, Dario Amodei, waded into that argument to reject a claim that has become popular in parts of Silicon Valley: that regulating AI mostly serves to entrench the handful of companies big enough to comply, and that the safer path is to spread AI capability as widely as possible instead.
The challenge came from investor Gavin Baker, who argued that AI’s risks might be better handled by distributing its capabilities broadly rather than concentrating them through regulation. He also suggested that Amodei’s own public warnings about AI could feed regulatory pressure and restrictions on data centres, making it less likely that AI delivers broad benefits. Amodei called the framing a “false choice.” Fair institutional processes, he argued, can constrain corporate power and protect individuals, and rules can be deliberately designed to fall harder on the most advanced companies than on smaller challengers.
To show this is not just rhetoric, he pointed at specifics. Anthropic backed California’s SB53 and, earlier, took a position on SB1047, and he noted that both exempted companies below set revenue or model-training thresholds. He also said Anthropic has supported testing frameworks that impose tougher requirements on frontier models than on less capable ones, an arrangement he believes can actually help challengers, including developers of open-weight models. The underlying claim is that regulation is not one undifferentiated weight pressing down on everyone equally; write the thresholds well, and it presses hardest on the incumbents.
What makes his position genuinely interesting is where it complicates itself. Amodei also maintains that AI is structurally prone to concentrating power, because the economics and the sheer computing requirements of training advanced models push toward a small number of players. Open-weight models help distribute capability, he allows, but they are not sufficient, because access to large amounts of compute and chips stays concentrated among frontier labs and major hardware providers. That is a notable concession: it partly grants Baker’s premise that openness spreads power, while insisting openness alone cannot solve a problem rooted in who controls the physical infrastructure.
He also pushed back on the charge that he is relentlessly gloomy, citing his essay “Machines of Loving Grace” and its case for AI transforming healthcare and biology. The industry’s real problem, he suggested, is a crisis of public trust, which companies should address by delivering tangible benefits rather than by leaning on optimistic marketing. On the mechanics of oversight he stayed consistent with past positions, backing pre-deployment testing for frontier systems and testing of open-weight models as they approach the frontier, and voicing support for Google DeepMind chief Demis Hassabis’s idea of a FINRA-style body to supervise the field. The debate with Baker will not be settled by a single exchange, but it clarifies the real question under all the noise: not whether to act, but how to write rules that check the powerful without simply handing them the field.