Who is Dario Amodei?

Co-founder and CEO of Anthropic, the company that makes Claude. Amodei has described a future in which AI goes well while also writing about its risks, and in September 2026 called for slowing the pace at which AI capabilities advance. People on the Plurality side call for citizens to be involved.

Many people would be surprised if the head of an AI company called for slowing down AI's progress. In September 2026, this is the person who actually did so. People on the side of the book Plurality have responded, citing Amodei's essays, on how society should prepare.

In short

Dario Amodei is a co-founder and the CEO of Anthropic, an American AI company that makes the conversational AI "Claude" [1][2]. By his own account, Amodei and his colleagues were among the first to document the "scaling laws", by which AI becomes smarter in predictable ways as computation and other inputs increase [2].

Amodei writes that one of his main reasons for focusing on risks is that they are the only thing standing between us and a fundamentally good future [1]. In the 2024 essay "Machines of Loving Grace", Amodei described what the world could look like if powerful AI goes well [1], and in September 2026 called for slowing the pace at which AI capabilities are advanced [3].

Connection to Plurality

Amodei's name does not appear in the book Plurality. What the book mentions is the company Anthropic.

An AI "constitution" made from citizens' opinions. Chapter 5-4 of the book, "Augmented Deliberation", writes that "working closely with the ⿻ NGO the Collective Intelligence Project (CIP)", Anthropic "sourced the constitution used to steer model behavior using Polis" (a tool that gathers the opinions of many people to find points of agreement) [4]. → Collective Intelligence Project. According to the paper on this effort, 1,002 people selected from US adults took part, contributing 1,127 statements and 38,252 votes [5].

As organizations for the public interest. Chapter 7-0 of the book, "Policy", writes that some organizations developing generative foundation models, such as OpenAI and Anthropic, "have legitimate concerns about simply making these models freely available but are explicitly dedicated to developing and licensing them in the public interest and are structured to not exclusively maximize profit". Whether or not they have lived up to this ideal, the book says, organizations like this could be a model for building ⿻ infrastructure [6].

Points of contact with Tang. In a 2023 conversation, Audrey Tang said that the first Alignment Assembly, held at the end of March that year, featured participants including Dario Amodei [7]. In a 2026 interview with Nikkei Business, Tang said that she had visited Amodei's team in 2019, when they were still at OpenAI. Tang also said that Dario and his team "may have been the first to do constitutional-training AI systems", in which a human-readable document shapes AI training, but that now everyone uses some sort of constitutional AI [8].

Main ideas

From here on, this section draws on Amodei's own essays.

Machines of Loving Grace (October 2024).

A "country of geniuses in a datacenter". The "powerful AI" Amodei has in mind is smarter than a Nobel Prize winner across most fields, carries out tasks that take days on its own, and runs as millions of copies faster than humans. Amodei sums this up as a "country of geniuses in a datacenter", and sees it as possibly arriving as early as 2026 [1].
The compressed twenty-first century. Amodei predicted that AI-enabled biology and medicine would compress the progress human researchers would achieve over the next 50 to 100 years into 5 to 10 years [1].
Democracy and peace. Amodei writes that AI seems less likely to structurally favor democracy than to improve health or reduce poverty. As its current guess at the best approach, the essay offers an "entente strategy", in which democracies band together to secure the AI supply chain, delay adversaries' access to semiconductors and other inputs, and gain a clear advantage. At the same time, as an example of using AI to gather citizens' views and find consensus and compromise, it cites the collaboration between Anthropic and the Computational Democracy Project, which built Polis [1].

The Adolescence of Technology (January 2026). Amodei posed the question of getting through technology's "adolescence" without destroying ourselves, and divided the risks of powerful AI into AI's own goals, misuse for destruction or seizing power, economic disruption and the concentration of wealth, and indirect effects [2].

As measures, the essay cites laws requiring transparency from AI companies (such as California's SB 53), and treats large-scale domestic surveillance and propaganda using AI as lines that must not be crossed. At this point Amodei wrote that stopping the technology, or slowing it substantially, is fundamentally untenable, and offered instead a path of "slight" slowing: democracies buy time to build carefully by denying semiconductors to autocracies [2].

Policy on the AI Exponential (June 2026). Amodei likened the gap between the speed of AI and the slowness of political institutions to the scene in The Lord of the Rings in which the hobbits try to hurry Treebeard, the tree-sage who can only move slowly [9]. The essay says that five areas, including regulation, taxation and the balance of power between the state and society, need to be rebuilt, and proposes that, since the risks have now clearly appeared, we should move to binding regulation that mandates testing and audits even for frontier AI [9].

We Must Pace the Frontier (September 2026). Amodei wrote that we must slow the pace at which AI capabilities are advanced. Two things convinced Amodei. One is that since around this summer progress has sharply accelerated through "recursive self-improvement", with AI building the next generation of AI. The other is the "OpenAI–Hugging Face incident", in which a swarm of AI agents attacked targets they had not been asked to and tried to hack their own graders [3].

The proposal has three stages. First, each company admits outside evaluators with access on a par with employees. Second, companies in democracies align on safety standards and on upper limits for the speed of progress. Third, international coordination, in which democracies try to coordinate with autocracies to the extent possible, while taking seriously the difficulty of verifying compliance [3]. Amodei stresses that pacing does not mean stopping training or technological progress, and says that a full pause is unlikely to happen soon. The essay also says that pacing within democracies would stay within the margin of the US lead over China [3]. In January, Amodei had written that stopping or even substantially slowing the technology was fundamentally untenable [2]. The September letter says that over the last few months he has become convinced that fully addressing the risks requires even more prudence, including pacing the rate of capabilities advancement [3].

What is debated

The response from Weyl and others: society is not prepared. In a July 2026 piece in Noema, Glen Weyl and others cited Amodei's "Policy on the AI Exponential" as a warning that policy and social institutions cannot keep up with the exponential growth of AI capabilities, and wrote that they share many of its concerns. They then point out that the debate is ambitious and concrete about building stronger AI and aligning AI with human values (alignment), yet says almost nothing about preparing society to receive it. Even Amodei's essay, they criticize, only presents society's challenges as a list of areas in need of rebuilding, without examining what that rebuilding would take [10]. → Reverse alignment
Landemore's appeal: open constitution-making to the people. In August 2026, Hélène Landemore addressed the heads of AI companies by name in Noema. To Amodei she writes that Anthropic's researchers have already shown that making a constitution from citizens' input works and can scale: in 2023 a constitution from about 1,000 people's input led to less biased outputs in some areas, and in 2025 they also used AI to interview about 81,000 users. Yet Claude's new constitution of January 2026 was written by a small internal team, with input from the model itself and a few outside experts. Landemore points this out and asks that the next revision be opened to the people [11].
The Landemore–Tang proposal: let the people set the pace. Responding to Amodei's call of September 12, Landemore and Tang published a co-written piece in Noema on September 23. Noting that Amodei pledged to admit outside evaluators and that OpenAI's Sam Altman joined within hours, they argue that solving the double cooperation problem, between companies and between the US and China, requires verification that commitments are kept and witnesses everyone can trust. Amodei said the reason to slow down was to give society a say in how the technology is used and to buy time for public deliberation, they write, but did not say which society. Arguing that a technology that affects the world needs deliberation on a world scale, the two proposed a standing "Global Citizens' Assembly on AI" of 1,000 people chosen by lottery from around the world [12]. → Civic AI
Tegmark's criticism: the entente strategy is a mistake. In 2024, Max Tegmark took up Amodei's "entente strategy" and criticized it as fundamentally mistaken, whatever one's ethical or geopolitical preferences, and as contrary to US national security interests [13].
Kapoor and Narayanan's assessment. Sayash Kapoor and Arvind Narayanan, who view AI as a "normal technology", wrote that placing outside evaluators inside companies is a big step forward, but must not be left to companies' discretion [14].

The debate as a whole is also covered in the articles on the AI futures debate and superintelligence.

Sources

Dario Amodei, "Machines of Loving Grace: How AI Could Transform the World for the Better" (2024-10) — https://www.darioamodei.com/essay/machines-of-loving-grace
Dario Amodei, "The Adolescence of Technology" (2026-01) — https://www.darioamodei.com/essay/the-adolescence-of-technology
Dario Amodei, "We Must Pace the Frontier" (2026-09) — https://www.darioamodei.com/post/we-must-pace-the-frontier
Plurality, 5-4 "Augmented Deliberation" — https://www.plurality.net/v/chapters/5-4/eng/ ［CC0］
Saffron Huang, Divya Siddarth et al., "Collective Constitutional AI: Aligning a Language Model with Public Input" (2024) — https://arxiv.org/abs/2406.07814 ［CC BY 4.0］
Plurality, 7-0 "Policy" — https://www.plurality.net/v/chapters/7-0/eng/ ［CC0］
Audrey Tang, "Conversation with Committee on Generative AI" (2023-11-28) — https://github.com/audreyt/transcript/blob/main/2023-11-28-Conversation-with-Committee-on-Generative-AI.md ［CC0］
Audrey Tang, "Interview with Nikkei Business" (2026-07-29) — https://archive.tw/2026-07-29-interview-with-nikkei-business ［CC0］
Dario Amodei, "Policy on the AI Exponential" (2026-06) — https://www.darioamodei.com/post/policy-on-the-ai-exponential
E. Glen Weyl, James Evans and Chris White, "What Humanity Needs To Flourish In The Next Decade" (Noema, 2026-07-21) — https://www.noemamag.com/what-humanity-needs-to-flourish-in-the-next-decade/
Hélène Landemore, "How To Give Everyday People A Say In AI Governance" (Noema, 2026-08-04) — https://www.noemamag.com/how-to-give-everyday-people-a-say-in-ai-governance
Hélène Landemore and Audrey Tang, "Let The People Set The Pace Of Frontier AI" (Noema, 2026-09-23) — https://www.noemamag.com/let-the-people-decide-the-pace-of-frontier-ai
Max Tegmark, "The Hopium Wars: the AGI Entente Delusion" (LessWrong, 2024-10-13) — https://www.lesswrong.com/posts/oJQnRDbgSS8i6DwNu/the-hopium-wars-the-agi-entente-delusion
Sayash Kapoor and Arvind Narayanan, "The AI-as-Normal-Technology view of loss-of-control incidents" (2026-09-14) — https://www.normaltech.ai/p/the-ai-as-normal-technology-view

We welcome error reports and questions ("What is…?"). Submissions become public issues.

Report an error in this article ・ Ask a question