Dario Amodei's Long Game: How the Anthropic CEO's Safety Stance Has Shifted

Since co-founding Anthropic in 2021, Dario Amodei has rarely stayed in one place on the question of what to do about powerful AI. His public stance has moved, over five years and thousands of pages of personal writing, from principled alarm to conditional optimism to something closer to a pragmatist's argument that building carefully and building fast are the same project. A profile published this week traces that arc from the early days of the lab through the aftermath of Claude Mythos, and the distance covered is worth examining.

The Congressional Period

Amodei's earliest public positioning centered on what could go wrong. In July 2023, he delivered written testimony to the Senate Judiciary Subcommittee on Privacy, Technology, and the Law, making him one of the first frontier lab chief executives to formally address Congress on AI risk. The message was direct: these systems posed genuine dangers, the industry could not adequately self-regulate, and government action was not optional. He called for mandatory pre-deployment evaluations, liability frameworks, and independent auditing with meaningful enforcement.

That position reflected the founding logic of Anthropic. The company was created partly as a response to concerns that AI development was outpacing the ability to manage its consequences, and the 2023 testimony placed Amodei as one of the more credible voices making that case before legislators who were still trying to understand what large language models actually did.

The Upside Turn

October 2024 brought a visible shift in emphasis. In a 14,000-word essay titled "Machines of Loving Grace," published on his personal website, Amodei did not retreat from his earlier safety concerns. He reframed them. Safety, in the new account, was no longer primarily a reason for caution. It was the remaining obstacle to a genuinely transformative upside. AI developed properly could compress decades of scientific progress, accelerate biomedical research that has stalled for generations, and reshape human cognitive capacity in ways that made urgent, careful investment not just prudent but morally necessary.

The change in emphasis was noticeable. The earlier Amodei wrote primarily about what could go wrong. The October 2024 version wrote primarily about what could go right, provided the safety work kept pace. Anthropic had not changed its mission, but the public framing of that mission had shifted toward its optimistic register.

The Adolescence of Technology

January 2026 brought the longest public statement yet: a 20,000-word essay titled "The Adolescence of Technology." The regulatory position refined accordingly. Where the 2023 congressional testimony emphasized the need for more oversight, the 2026 essay argued for smarter oversight — targeted, standards-based, and as light as possible while still binding. The regulatory model should not destroy economic value, Amodei wrote, but it should be mandatory enough that bad actors cannot simply opt out.

The cars-and-aeroplanes framing that has since become his preferred shorthand arrived in April 2026, in the wake of Claude Mythos Preview, the security model that identified thousands of zero-day vulnerabilities while Anthropic chose not to release it publicly. In a Financial Times interview, Amodei said he wants AI regulated "the way you regulate cars and aeroplanes" — mandatory, third-party-assessed, built around safety standards that bind the whole industry. The position sits considerably closer to the EU AI Act's structural approach than to the dominant Silicon Valley preference for self-governance.

"Regulations should be judicious: they should seek to avoid collateral damage, be as simple as possible, and impose the least burden necessary to get the job done." Dario Amodei, "The Adolescence of Technology," January 2026

State Contracts and the Pentagon

One practical measure of how positions have shifted involves government engagement. Early Anthropic maintained careful distance from defense contracts. As Claude's capabilities grew and enterprise adoption broadened, that distance narrowed. Anthropic's legal challenge to its Pentagon "supply chain risk" designation illustrated the tension between messaging and commercial reality. The appeals court proved skeptical in May 2026, and the company ultimately accepted the designation. Amodei's public comments on state contracts have moved from principled avoidance to conditional engagement — government use cases where the safeguards match the risk level.

The trajectory of Amodei's thinking is easier to explain than it might appear. Each shift tracks the capabilities of the models he was building. As Claude became more capable, safety work became more consequential, and the more consequential the safety work, the more confidently Amodei argued that building carefully and building fast were not opposing goals. His earlier arguments about labor market disruption and his later arguments about AI's potential to transform scientific research share the same structure: acknowledge the risk, then argue that the correct response is not slower development but better-governed development.

Whether that logic holds through the next capability jump — past Mythos into whatever model class arrives in 2027 — remains the open question. Amodei has consistently argued that his position is not driven by commercial interest in building. Whether observers believe him may depend on what Anthropic does with the models that follow.