Dario Amodei: How His AI Safety Position Evolved, 2021-2026

In October 2024, Dario Amodei published a 14,000-word essay arguing that AI could compress decades of cancer and mental-health research into five to ten years. Three years earlier, he had left OpenAI with 14 colleagues over what he called a difference in vision on how seriously to treat AI risk. Both positions are his, and together they trace one of the more significant shifts in how a frontier AI executive publicly frames the tension between capability and safety.

The 2021 Premise: Safety as a Founding Condition

Amodei left OpenAI in December 2020, with 14 researchers following over the next several months, including his sister Daniela, who had been VP of Safety and Policy. The founding pitch to investors was deliberately dual: Anthropic would build the most powerful AI systems while simultaneously figuring out how to make them safe. The company raised $124 million in its initial round on that premise, per Inc.

The safety commitment took technical form in 2022 with Anthropic's Constitutional AI paper, a method for training models to be helpful, harmless, and honest using AI-generated feedback against a written set of principles, rather than relying exclusively on human preference labelling at scale. It was an early attempt to build safety reasoning into the training loop itself, rather than applying filters after model training completed.

In July 2023, Amodei delivered written testimony to the Senate Judiciary Subcommittee on Privacy, Technology, and the Law, making him one of the first frontier lab CEOs to formally address Congress on AI risk. His position at that point was cautionary and constructive: AI poses serious risks requiring government engagement, and industry cannot self-regulate its way to adequate safety.

The 2024 Pivot: Optimism at Scale

"Machines of Loving Grace," published on Amodei's personal website in October 2024, marked a visible shift in register. The 14,000-word essay did not walk back his safety concerns. Instead, it argued that safety was the only remaining obstacle to a genuinely transformative upside. "Most people are underestimating how radical the upside of AI could be," he wrote, claiming that AI could compress decades of biological research into five to ten years, potentially resulting in reliable prevention and treatment of most cancers, improved mental health outcomes including near-elimination of PTSD and clinical depression, and meaningful extension of healthy human lifespan.

The essay covered five thematic areas: biology and health, mental health and neuroscience, economic development, governance and democracy, and AI development itself. On the first three Amodei was markedly optimistic. On governance he was more cautious, noting that "no such automatic direction exists for democracy and peace" and that AI could "just as easily strengthen authoritarianism through surveillance, propaganda, and opinion manipulation." The optimism was not unconditional; it was contingent on responsible development and some form of international cooperation.

That same month, Bloomberg reported that Amodei believed AI could outsmart most humans as soon as 2026, a claim that would have read as a warning three years prior but now sat within his broader optimist framing. His counterpart at Google DeepMind, Demis Hassabis, was drawing similar timelines through AlphaFold and drug-discovery partnerships.

2025-2026: Smarter Rules, Not Fewer Deals

In November 2025, Fortune reported that Amodei said he was "deeply uncomfortable" with a small group of technology executives determining the future of AI without broader government involvement. He also predicted that "AI could displace half of all entry-level white collar jobs in the next one to five years," a claim that sat in tension with his optimist essay on economic development but reflected the same underlying belief: the transition would be fast and the stakes were high enough to require state intervention.

In June 2025, Amodei published an op-ed criticising a Senate proposal to impose a ten-year moratorium on state-level AI regulation as part of a broader Republican tax bill. He called it "far too blunt an instrument," per The Hill. His preferred alternative was narrower: require frontier AI developers to publicly disclose their testing methodologies and risk mitigation strategies, codifying practices that Anthropic, OpenAI, and Google DeepMind already follow voluntarily. The Senate ultimately rejected the moratorium by a vote of 99 to 1.

By early 2026 the picture had grown more complex. At Bloomberg House in Davos in January, Amodei said he felt "maybe a bit more negative about the public awareness and the actions of wider society" on AI risk, per Bloomberg. In February, Anthropic formally dropped the conditional pause commitment it had once maintained, under which the company had pledged to halt development if its own AI became too dangerous. In March, Bloomberg and the Financial Times reported that Amodei had reopened talks with the Pentagon that had previously collapsed when Anthropic demanded assurances against using its AI for mass surveillance of US citizens or in autonomous weapons systems. The new discussions proceeded on terms Anthropic defined.

What It Means

The 2021 Amodei and the 2026 Amodei are not in contradiction, but the distance is real. He left a company because he disagreed with how it weighted safety against competitive pressure; he now runs one that has dropped its pause commitment and is negotiating military contracts, while simultaneously calling for mandatory transparency rules and warning that society is under-responding to AI's pace. The position has moved from "regulate more" to "regulate smarter," and from principled avoidance of state contracts to conditional engagement with them. Whether that is a pragmatic evolution or a rationalisation depends on which of his public statements you weight most heavily; what the record shows is that his positions have consistently followed Anthropic's institutional interests and competitive position, even when framed in the language of safety.

Sources

• darioamodei.com: Machines of Loving Grace (Oct. 2024)
• Bloomberg: Anthropic CEO Thinks AI May Outsmart Most Humans As Soon As 2026 (Oct. 2024)
• Bloomberg: Anthropic's Amodei on AI: Power and Risk, Davos (Jan. 2026)
• Bloomberg: Anthropic Reopens Talks with Pentagon (Mar. 2026)
• Fortune: Amodei Is 'Deeply Uncomfortable' with Tech Leaders Determining AI's Future (Nov. 2025)
• The Hill: Anthropic CEO Calls GOP AI Moratorium 'Far Too Blunt an Instrument' (Jun. 2025)
• Senate Judiciary Committee: Written Testimony of Dario Amodei (Jul. 2023)
• Inc.: Amodei Says He Left OpenAI Over a 'Difference in Vision'

Editorial standards: every claim is sourced. Tips: [email protected]