AI, Power, and Trust: Dario Amodei Defends His Balancing Act

Dario Amodei argues for targeted regulation of frontier models and acknowledges that the AI industry still has to deliver on its promises in healthcare.

After investor Gavin Baker accused him of supporting a path that would concentrate AI in the hands of a few companies and political leaders, while fueling public distrust through messaging seen as overly pessimistic, Dario Amodei responded on both fronts: regulation and public trust.

The Anthropic chief first rejects the binary choice between regulating the technology, at the risk of strengthening dominant players, and distributing it widely. In his view, impartial institutions can instead limit the power of large companies by subjecting them to common rules. Poorly designed regulation can favor established players, he acknowledges, but that outcome is not inevitable.

His defense draws on several measures supported by Anthropic. California’s SB 53 exempts companies with less than $500 million in annual revenue. The evaluation processes advocated with US authorities would also subject frontier models to stricter testing than less advanced systems. The proposed approach is to regulate leading labs without slowing their challengers to the same extent, including those distributing models as open weights.

For Amodei, however, the concentration of power predates regulation. It stems from the amount of compute, chips, and infrastructure required to develop the most powerful models. Open weights broaden access, but also shift some of that power toward organizations capable of financing and operating these resources. On their own, they would therefore be an insufficient answer.

He supports a framework designed to address cyber, biological, and alignment risks, constrain frontier labs, and preserve room for open weights models. He favors pre-deployment evaluations for the most advanced systems, whether open or closed, and backs Demis Hassabis’s proposal for a FINRA-style organization tasked with testing models before they reach the market.

The second part of his response concerns public perceptions of AI. Dario Amodei disputes the idea that his warnings are responsible for the negative sentiment surrounding the technology. He instead sees a decades-old crisis of trust in companies, governments, and the tech industry. A positive communications campaign would not be enough to repair it, particularly since the promise that AI could cure cancer has, in his view, become a cliché often seen as deceptive.

His response can be summed up in one sentence: “The thing that will work is actually curing cancer.” The Anthropic CEO argues that the fairest criticism of AI companies is their current failure to deliver on their biggest promises. Yet he maintains his projection that AI could help treat most human diseases within five to ten years, a highly ambitious estimate that remains unproven. His sense of urgency is also personal: his father died from hepatitis C several years before the arrival of a treatment that likely could have saved him.

Anthropic says it is rapidly expanding its work in biology and medicine, with early indications expected in the coming months and more substantial results over the longer term. Amodei would rather wait for measurable progress before promoting it publicly, while continuing to discuss the risks associated with the most capable systems.

Fidji Simo, OpenAI’s former CEO of Applications, shares his assessment of AI’s medical promises but points to a material constraint. More capable models will not automatically produce treatments unless biological data, scientific infrastructure, and experimental tools advance at the same pace. She believes cancer could benefit from AI more quickly because of the extensive datasets accumulated across genomics, pathology, imaging, and clinical outcomes. For many chronic diseases, that foundation remains far too fragmented.