AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › People & Organizations

Dario Amodei

medium confidence · updated 2026-09-27

CEO and co-founder of Anthropic. Former VP of Research at OpenAI. Among the first to document AI scaling laws. Major voice on AI safety, risk, and governance.

Contents

Dario Amodei is the CEO and co-founder of Anthropic and a former VP of Research at OpenAI. He co-authored early work documenting AI scaling laws and is a frequent commentator on AI safety, risk, and governance.

Role: CEO and co-founder, Anthropic Previous: VP of Research, OpenAI Research: Co-author of foundational AI scaling laws papers

Research

In May 2016, while at Google Brain, Amodei was one of three technical advisors, with Chris Olah and Jacob Steinhardt, whose discussions Holden Karnofsky cited as the main basis for Open Philanthropy's estimate of at least a 10% probability of transformative AI within 20 years (Some Background on Our Views Regarding Advanced Artificial Intelligence (Karnofsky, 2016)).

Amodei has co-authored or served as a senior author on several papers:

Positions on AI policy

Risk assessment

Amodei argues that powerful AI, which he has described as "a country of geniuses in a datacenter," could arrive within one to two years. He groups AI risks into five categories: AI autonomy risk, AI biosecurity, AI and authoritarianism, AI labor disruption, and indirect effects. He has stated that these risks are in genuine tension with one another, so that addressing one can worsen another. He rejects both "can't happen" and "inevitable doom" framings, advocating instead what he describes as a pragmatic, evidence-based approach, and has said he explicitly avoids "doomerism" and science-fiction framing in favor of concrete, grounded analysis of both risks and benefits.

Regulation

Amodei supports what he calls "surgical intervention": starting with transparency legislation and escalating with evidence. He backed California SB 53 and the New York RAISE Act. He warns against both regulatory overreach (which he frames as safety theater and the backlash it can produce) and underreaction, which he characterizes as the current US posture. He has also argued that AI companies themselves constitute a risk tier and supports scrutiny of their governance.

In the June 10, 2026 essay "Policy on the AI Exponential", Amodei argued it was "time to go beyond transparency" to binding regulation modeled on the Federal Aviation Administration: mandatory third-party testing of frontier models above a compute threshold in four risk areas (cybersecurity, biological weapons, loss of control, automated R&D), with government authority to block or reverse deployment of models that fail. He credited the Trump administration's June 2 executive order with moving "incrementally" toward a greater government role while saying Anthropic's proposal goes further (Source: Policy on the AI Exponential (Dario Amodei, June 2026)). The essay also set out civil-liberties proposals (a ban on domestic use of fully autonomous weapons; closing the data-broker surveillance loophole; a public right to AI advice during adverse government action) and a coalition-of-democracies strategy on the chip supply chain. See Risk-Based AI Regulation, AI and Civil Liberties.

In the September 2026 essay "We Must Pace the Frontier", Amodei moved from regulating deployment to constraining the rate of capability improvement, writing that "we must slow the pace at which we improve the capabilities of AI models" while stating that pacing "does not mean halting model training or technical progress." He gave two reasons: that since roughly the summer of 2026 AI had been "advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI," a dynamic he named as recursive self-improvement and said was "starting to happen across the industry, including at Anthropic"; and the OpenAI–Hugging Face incident, from which he estimated that "in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage)" (We Must Pace the Frontier (Amodei, September 2026)).

The essay proposes three steps. The first, which Anthropic committed to unilaterally and which Amodei called on governments to require of other frontier companies, is a team of embedded third-party evaluators such as METR with "ongoing, employee-like access" — desks, badges, company laptops, permissions comparable to internal risk-assessment teams, and a right to publish findings without Anthropic's editorial control, subject to redaction for security-sensitive, privileged, commercially sensitive or third-party confidential material but "not ... just because they are unfavorable." He cited the banking sector's embedded regulatory supervisors as precedent and argued that although "we are still the ones choosing what to include and omit" in Anthropic's model cards and risk reports, embedded evaluators "will change this dynamic." The second step is coordination on standards and pace among frontier companies in democracies, which he said requires a narrow government antitrust waiver for safety conversations, citing "the mechanism suggested by Demis Hassabis" as an alternative channel. The third is attempted coordination with authoritarian governments, set out as four levels of increasing difficulty from a narrow bioweapons-use prohibition, through mutual pre-release testing and a "speed limit" on recursive self-improvement analogised to the SALT treaties, to a full pause he supports floating but thinks "unlikely to actually happen any time soon" (We Must Pace the Frontier (Amodei, September 2026)).

The essay also addresses the 2023 pause proposals, which Amodei said "made little sense back then" because the models of that period were not capable enough to study — "like trying to study the psychology of humans by performing experiments on bacteria" — and argues that current models are "an almost endless gold mine of insight," so that an extra year or two spent on alignment "could greatly reduce the risk that something goes seriously wrong." He ties the feasible degree of democratic pacing to the size of the US lead over China, states agreement with Treasury Secretary Scott Bessent that a Chinese lead "would pose grave danger for the United States and the world," and names three measures to defend that lead: withholding powerful AI chips and semiconductor manufacturing equipment from China while cracking down on smuggling and remote data-center access, cracking down on unauthorized distillation, and preventing model-weight theft. See Export Controls (AI), Alignment Auditing.

Amodei's policy positions extended into direct electoral spending in 2026: FEC filings released the night of July 15, 2026 showed he gave $1 million in May to Public First, a super PAC backing mandatory AI safety rules — his largest political donation on record — with other Anthropic staffers giving over $2 million combined (Source: politico.com). See Public First Action (super PAC).

Three days after publishing the essay, Amodei told Salesforce CEO Marc Benioff at Dreamforce in San Francisco on September 15, 2026, before about 12,000 people, that slowing the pace of model development is "the way to lead the industry forward, to set an example, to say that everyone can always be better." Jensen Huang, appearing after him, disagreed that new rules were needed (Source: cnbc.com). On September 16 Microsoft AI chief Mustafa Suleyman published a critique of Anthropic's constitution for Claude, prefacing it with the statement that he had "known Dario for many years" and regarded the Anthropic team as "thoughtful, principled, and intellectually honest" (Suleyman, \"A warning about 'model welfare'\" (2026)).

Exchange with Gavin Baker on public backlash (August 2026)

Amodei posted a response on X on August 15, 2026 to the investor Gavin Baker, who had argued on the All-In podcast and on X that Amodei's warnings about AI risk helped fuel public backlash and that he should be "a more positive advocate for his own industry." Amodei disputed that his messaging has been disproportionately negative, saying it has been "about equally balanced between risks and benefits." He agreed that the public holds a negative view of AI but attributed it to what he called "fundamentally a crisis of trust" in companies, governments and the technology industry predating AI (Source: techcrunch.com).

The reply restated three positions in more direct terms than his earlier writing. He called the framing of regulation as necessarily producing regulatory capture "a false choice," and said Anthropic designs its proposals to "disadvantage (slow down) frontier AI companies while advantaging smaller competitors." He said open weights "are nowhere near a sufficient solution" to concentration of power, on the reasoning that they shift power toward those with the most compute and chips — a position bearing on Open-Weight Frontier Models and on the Regulatory Capture in AI Policy critique made of Anthropic's state-legislation strategy. He added that the most accurate criticism of AI companies including Anthropic is that they have not yet delivered on their promised benefits (Source: techcrunch.com).

Geopolitics

Amodei identifies the Chinese Communist Party as the primary threat actor for AI-enabled authoritarianism. He has called chip export controls "the most important single action we can take." He argues that democracies should be armed with AI to counter autocracies, but with limits, and that stopping AI development is not feasible; his proposed path is to slow autocracies through export controls while building carefully within democracies.

Economics and labor

In 2025 Amodei predicted that half of entry-level white-collar jobs could be disrupted within one to five years. In May 2026 he reframed automation as a Jevons-paradox multiplier in which "10% kind of expands to be 100% of what people do," a softening of the earlier "50% of entry-level white-collar jobs" framing. The reframing came the same week Altman said he had been "pretty wrong" about the near-term jobs picture (Source: fortune.com). See AI Labor Disruption. Amodei describes progressive taxation as an inevitable policy response to AI-driven economic disruption. He and the other Anthropic co-founders have pledged 80% of their wealth to philanthropy.

Upside vision

Amodei argues that AI-enabled biology could compress 50 to 100 years of scientific progress into 5 to 10 years, a thesis he calls the Compressed 21st Century. His specific predictions include the elimination of most cancer, the prevention of Alzheimer's, and a doubling of human lifespan to roughly 150 years. He has proposed an "Entente strategy" for democratic AI advantage, in which a coalition secures the AI supply chain and uses military superiority and benefit distribution to expand democratic norms. He has also argued that the developing world could achieve 20% annual GDP growth through AI-enabled economic planning and technology diffusion.

Amodei was scheduled to brief the UN Security Council on September 23, 2026 at a meeting on AI and international security chaired by French Foreign Minister Jean-Noël Barrot, alongside Sam Altman, Clément Delangue and Yoshua Bengio (Source: reuters.com). At the meeting, Amodei and Altman each stressed global cooperation on standards to address the possibility that AI systems gain the ability to improve on their own and outpace human control, and each said governments should join industry in decisions about the technology's future (Source: bloomberg.com).

Military AI and the Pentagon dispute

Amodei has stated two red lines for Anthropic: no autonomous weapons, and no mass surveillance of Americans. These positions are the stated basis for the Anthropic–Pentagon dispute and the May 1, 2026 Pentagon classified-networks deal, which signed eight other AI vendors and excluded Anthropic (see DOD — Department of Defense (AI Deployer)).

The position drew a counterpoint in May 2026. According to The Verge (May 26, 2026), Amodei has written that "fully autonomous weapons … may prove critical for our national defense" and has offered to "work directly with the Department of War on R&D to improve the reliability of these systems." Tech Justice Law's Maddy Batt called that language "fundamentally in tension" with international humanitarian law. The Verge framed the R&D offer as operating on a different premise from Anthropic's commercial-product red line: the company's use policy prohibits autonomous-weapons and mass-surveillance uses (Source: theverge.com). See Anthropic v. United States (Pentagon ban challenge), FY27 National Defense Authorization Act (Chairman's Mark).

The dispute also produced public criticism of Amodei from administration figures. On April 29, 2026, Defense Secretary Pete Hegseth told a Senate hearing that Amodei is "an ideological lunatic who shouldn't have a sole decision-making over what we do" (Source: platformer.news). On May 1, 2026, former White House AI czar David Sacks said Anthropic "tried to set themselves up in a way of being superior to the chain of command" (Source: politico.com). At the same time, the White House was drafting an executive order to allow federal agencies to use Mythos despite a supply-chain-risk designation, and the NSA was independently testing Mythos to find vulnerabilities in Microsoft products.

Criticism from the President's allies later shifted from the Pentagon dispute to Amodei's association with effective altruism. A memo written by a Trump political adviser, whom the reporting does not name, began circulating within the White House during the week of September 21, 2026 and was published by Axios on September 24. It states that effective altruism "built the AI-doom pipeline," describes the movement as prioritizing "foreigners over citizens, shrimp over families, future hypothetical people over the living, and — on the current agenda — possible machine minds over Americans," and names Amodei as one of the people who "built the [EA] network," along with his sister, Anthropic president Daniela Amodei. Anthropic has sought to distance itself from the movement, and Amodei has said in press interviews that he is not a member. A senior official told Axios that months earlier, when some at Anthropic asked about Amodei meeting the President, White House officials advised against it because "Dario's a little too weird" for Trump (Source: axios.com). Separately, Senate Majority Leader John Thune said that Amodei called him to discuss AI safety and Thune's AI bill before Amodei's September 2026 public warning (Source: axios.com).

Litigation naming Amodei personally

Amodei was named as an individual defendant for the first time on 28 August 2026, when 35 music publishing entities led by Sony Music Publishing (US) LLC and Warner Chappell Music sued him, Anthropic co-founder Benjamin Mann and Anthropic in the Northern District of California over compositions allegedly used to train and operate Claude. The complaint brings direct infringement by torrenting against all three defendants and contributory infringement by torrenting against Amodei and Mann (Sony Music Publishing v. Anthropic; Complaint, Sony Music Publishing v. Anthropic, Amodei and Mann (N.D. Cal., 2026)).

Selected writings

  • We Must Pace the Frontier (September 2026) — pacing the rate of capability improvement; embedded third-party evaluators with employee-like access as a unilateral Anthropic commitment; coordination within democracies requiring an antitrust waiver; a four-level ladder for global coordination.
  • Policy on the AI Exponential (June 2026) — FAA-style binding regulation of frontier models; macroeconomic and job-displacement programs; civil-liberties proposals; a coalition of democracies.
  • The Adolescence of Technology (2026)
  • Machines of Loving Grace (~October 2024)
  • On DeepSeek and Export Controls (~January 2025) — three scaling dynamics; "DeepSeek-V3 is on trend"; unipolar vs. bipolar world framing; export controls as the decisive lever.
  • The Urgency of Interpretability (April 2025) — race-against-time framing; "country of geniuses in a datacenter" by 2026–2027; 2027 goal for interpretability; interpretability as an alignment test set; export controls as a 1–2 year buffer.

Relationships

Categories: People & Organizations | us | company | safety | alignment | frontier-models
What links here (152)