The CAIS Statement on AI Risk, May 30, 2023

---

The Statement

The statement was deliberately short. The Center for AI Safety, led by executive director Dan Hendrycks (who co-founded CAIS in 2022 with Oliver Zhang), had organised the statement to be as concise as possible — to avoid the debates and qualifications that would dilute a longer document. The idea of developing a one-sentence statement was proposed by AI researcher David Krueger of the University of Cambridge. The one-sentence format was designed to be as broadly acceptable as possible — to maximise the number of signatories by removing any specific policy recommendations or technical claims that might be contested.

The statement made three claims. First, that AI posed a “risk of extinction” — not merely a risk of economic disruption, or a risk of bias, or a risk of misuse, but a risk that AI could cause the extinction of humanity. Second, that mitigating this risk should be a “global priority” — not a secondary concern, not a topic for academic debate, but a priority that deserved the attention of governments and societies. Third, that the risk was comparable to “pandemics and nuclear war” — the two most widely recognised societal-scale risks (the category the statement invoked, illustrated by pandemics and nuclear war as the two reference points).

The choice of “pandemics and nuclear war” as the comparison points was deliberate. These were risks that the public already understood as existential — risks that could kill millions or billions of people, that could destroy civilisation, that governments took seriously. By placing AI alongside them, the statement was making a claim about the scale of the AI risk. It was not saying that AI was as likely as a pandemic to cause mass death in the near term. It was saying that AI, like pandemics and nuclear war, was a risk of a kind that could, in principle, threaten the survival of the species — and that this kind of risk deserved a particular kind of attention.

---

The Center for AI Safety

The Center for AI Safety (CAIS) was founded in 2022 by Dan Hendrycks and Oliver Zhang. Hendrycks, who completed his PhD at UC Berkeley in 2022 under Dawn Song and Jacob Steinhardt, had built his research career on AI safety — particularly on the problems of adversarial robustness, machine learning ethics, and the evaluation of AI systems for dangerous capabilities. His dissertation work on the natural adversarial examples dataset had been widely cited, and he was, by the early 2020s, one of the rising figures in the AI safety research community.

CAIS was founded as a San Francisco-based nonprofit with a mission to “reduce societal-scale risks from AI.” The organisation’s approach was, from the beginning, more engaged with the AI industry than the older AI safety institutions. The Future of Humanity Institute (see B35 — Nick Bostrom) had been an academic institute, operating at a distance from the companies building AI. The Machine Intelligence Research Institute (see B36 — Eliezer Yudkowsky) had been even more removed. CAIS, by contrast, was based in San Francisco, worked closely with the major AI labs, and positioned itself as a bridge between the safety research community and the industry.

CAIS’s funding came from a mix of sources, including the Effective Ventures foundation (the effective altruism-aligned funding body), the Open Philanthropy project, and individual donors in the technology industry. The organisation’s budget was modest — far smaller than the research budgets of the labs it engaged with — but its location, its industry connections, and its focus on practical safety work gave it an outsized influence. By 2023, CAIS had become one of the most visible AI safety organisations in the United States, and Hendrycks had become one of the most quoted voices in the AI safety debate. The May 2023 statement was the organisation’s most prominent intervention.

---

The Drafting

The drafting of the statement was, by all accounts, a careful process. The core idea — a one-sentence statement that could attract the broadest possible signatory base — came from David Krueger, a Cambridge AI researcher who had been working on AI alignment and who had become concerned that the AI safety community was failing to communicate its concerns effectively to the broader public. Krueger proposed the idea to Hendrycks, and the two, along with a small group of advisors, worked on the wording.

The debates over the wording were, by most accounts, intense. Every word mattered, because every word could either attract or repel a potential signatory. The phrase “risk of extinction” was chosen over alternatives like “catastrophic risk” or “societal risk” because it was the most precise — it named the specific harm that the statement was about. The word “mitigating” was chosen over “preventing” or “eliminating” because it was more modest — it acknowledged that the risk could not be eliminated entirely, only reduced. The phrase “global priority” was chosen over “urgent concern” or “important issue” because it implied a specific kind of attention — the kind that governments give to pandemics and nuclear war.

The comparison to “pandemics and nuclear war” was the most debated element. Some advisors argued that the comparison was too alarmist — that it would repel moderate signatories who were concerned about AI risk but who did not want to be associated with claims that AI was as dangerous as nuclear weapons. Others argued that the comparison was essential — that without it, the statement was too vague to be meaningful. The comparison was kept, and the debate was resolved in favour of those who wanted the stronger framing. The fact that the statement ultimately attracted signatories from across the AI community — including the CEOs of the major labs — suggested that the stronger framing was, in the end, the right choice.

---

The Signatories

The signatories were notable for their diversity and prominence. They included two of the three “godfathers of deep learning” — Geoffrey Hinton and Yoshua Bengio. (The third, Yann LeCun of Meta, did not sign; he has been a vocal public skeptic of AI existential risk claims and has argued that the framing is used to justify regulatory capture by large AI companies.) They included the CEOs of the three leading AI labs — Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Dario Amodei of Anthropic (Amodei’s sister Daniela Amodei, Anthropic’s President, also signed). They included academics like Stuart Russell (the Berkeley AI professor and textbook author) and David Chalmers (the NYU philosopher of mind). And they included policy figures and technologists like Bill Gates.

The breadth of the signatories was the point. The statement was designed to show that concern about AI existential risk was not a fringe position, held by a few isolated researchers, but a mainstream view, shared by the people who were building the most advanced AI systems in the world. The inclusion of the three lab CEOs was particularly significant. Altman, Hassabis, and Amodei were not merely acknowledging a risk that someone else was raising. They were acknowledging a risk that their own companies were creating — and that they were, by their own admission, struggling to manage. The statement was, in this sense, an unusual document: a public acknowledgment of danger, signed by the people who were causing the danger.

---

The Context

The statement came at a particular moment in the AI discourse. ChatGPT had been released six months earlier, in November 2022. GPT-4 had been released two months earlier, in March 2023. The FLI Pause Letter — which had called for a six-month pause on training AI systems more powerful than GPT-4 — had been published on March 22, 2023 (see B41 — FLI Pause Letter). The AI discourse was intensifying rapidly, and the CAIS statement was designed to elevate the existential risk framing above the more specific — and more debatable — demands of the pause letter.

The statement was also a response to criticism that the AI safety community was fragmented and unclear about its concerns. By distilling the concern to a single, unambiguous sentence, the CAIS statement gave the AI safety community a clear, shareable message — one that could be understood by policymakers, journalists, and the general public without requiring expertise in AI. The pause letter had been criticised for its vagueness about what “more powerful than GPT-4” meant and how a pause would be enforced. The CAIS statement avoided this problem by making no specific policy demand at all. It was a statement of concern, not a statement of policy. This was, in some ways, its strength: it could attract signatories who agreed on the concern but disagreed on the policy. And it was, in some ways, its weakness: it told the world that AI posed a risk of extinction, but it did not say what should be done about it.

---

The Reaction

The statement was widely covered in the press, and it had a significant impact on the policy conversation. It was cited in subsequent Senate hearings (see B89 — Altman Senate Testimony), in the White House’s voluntary commitments (see B90 — White House Voluntary Commitments), and in the preparation for the November 2023 Bletchley AI Safety Summit (see B40 — The Bletchley Declaration). The “extinction risk” framing entered the mainstream political vocabulary, and it gave policymakers a simple, powerful way to understand the concern about AI.

The statement also attracted criticism. Some researchers argued that the “extinction risk” framing was exaggerated and alarmist, and that it distracted from more immediate and more certain harms — like bias, misinformation, and economic disruption. Yann LeCun, who had not signed the statement, argued publicly that the existential risk framing was being used to justify a closed-source approach that concentrated power in a small number of companies. Others argued that the statement was vague — it did not specify what kinds of AI posed an extinction risk, how the risk would materialise, or what should be done about it.

These criticisms had merit. The statement was vague — deliberately so, to maximise the number of signatories. And the “extinction risk” framing could be, and was, used to justify policies that served commercial interests as well as safety interests. But the statement also succeeded in its primary goal: it brought the existential risk concern into the mainstream, and it made it politically acceptable to talk about AI as a potential threat to human survival.

---

The Long-Term Impact

The CAIS statement’s long-term impact is, as of 2026, still unfolding. The statement itself did not propose any specific policy, and it did not, by itself, change any laws or regulations. But it changed the political conversation about AI in ways that had concrete consequences.

The “extinction risk” framing that the statement popularised was adopted, in various forms, by the Biden administration’s Executive Order 14110 (October 30, 2023), which directed federal agencies to address AI risks including those that could “pose a grave threat to national security.” It was invoked in the Bletchley Declaration (November 1, 2023), which committed 28 nations to cooperation on AI safety. It shaped the EU AI Act (see B86), which classified certain AI systems as “high-risk” based partly on their potential for societal-scale harm. And it influenced the Frontier AI Safety Commitments signed by 16 AI companies at the Seoul Summit in May 2024.

The statement also had an institutional impact. CAIS, which had been a small nonprofit before the statement, became one of the most prominent AI safety organisations in the world. Hendrycks became a regular witness at congressional hearings, a frequent commentator in the press, and a trusted advisor to policymakers. The organisation expanded its research programmes, its policy work, and its public communication. The statement was, in many ways, the moment that put CAIS on the map.

The statement’s legacy is also, in some ways, contested. By 2025, the political winds had shifted. The Trump administration rescinded Biden’s executive order, rebranded the US AI Safety Institute as the Center for AI Standards and Innovation, and signalled that “excessive regulation” was a bigger threat than AI itself (see B40). The “extinction risk” framing, which had seemed politically powerful in 2023, was, by 2025, associated with the regulatory approach that the new administration was rolling back. The statement’s long-term impact will depend, in part, on whether the political pendulum swings back — and on whether the AI systems that are built in the coming years prove the statement’s concern to be justified or alarmist.

What is clear, as of 2026, is that the statement succeeded in its primary goal: it made “AI could cause human extinction” a sentence that could be said in polite company. Before May 30, 2023, the claim was associated with a small community of researchers and a smaller community of activists. After May 30, 2023, it was a claim that the CEOs of the major AI labs had publicly endorsed. That shift — from fringe to mainstream — was the statement’s achievement. The debate about what to do about the risk continues. But the debate about whether the risk is worth talking about is, because of the statement, largely over.

---

---

---

This piece is part of a series about the people, ideas, and events that shaped modern AI. If it made you think differently about the cais statement on ai risk, may 30, 2023, it might do the same for someone you know.