Doom Debates

Let's CALL OUT the AI Doom “Enablers” Joining OpenAI & Anthropic — Dr. Holly Elmore, PauseAI US

1h 56mJun 30, 2026

Key Themes

AI pause advocacyFrontier AI governanceTechnical safety tensionsEffective altruism conflictPublic criticism normsAnthropic influenceAccountability and liability

Summary

Holly Elmore argues AI safety needs political pressure, not insider deference

In this Doom Debates episode, PauseAI US executive director Holly Elmore makes a forceful case that frontier AI risk cannot be solved by technical insiders alone. She argues for democratic oversight, international coordination, and a pause on advanced AI development, while criticizing parts of the rationalist, effective altruist, and AI safety worlds for joining or normalizing companies such as Anthropic and OpenAI. The conversation ranges from AI governance and public advocacy to harsh criticism, community norms, Anthropic’s influence, Open Philanthropy, and the moral tension between warning about AI doom while using or benefiting from AI tools.

The episode’s core divide is governance versus insider-led safety

Elmore argues that frontier AI risk should not be left mainly to technical communities or company insiders. Her case is that democratic oversight, regulation, diplomacy, and an international pause are necessary because the institutions building AI have incentives that conflict with public safety.

Warnings only matter if they change decisions

A recurring theme is Elmore’s frustration with people who analyze AI risk, assign high p(doom), or identify dangerous model behavior but still use, praise, or help deploy frontier AI systems. She argues that safety evaluations and public concern should translate into concrete restraint, regulation, refusal, or pause advocacy.

Public advocacy needs social permission, not just arguments

Elmore presents PauseAI US as trying to normalize ordinary people saying that AI development should be paused. The discussion of the Yudkowsky/Soares book and LessWrong’s muted response highlights how movements can fail to mobilize when people privately agree with a message but do not create visible social support around it.

Sharp criticism can shift debate, but it carries fairness costs

The episode repeatedly tests the boundary between moral criticism and personal attack. Elmore argues that high stakes justify harsher public rhetoric and less deference to insiders, while the host worries that speculative motive-reading can be unfair and weaken otherwise valid criticism.

Institutional ties shape what communities treat as normal

Elmore’s critiques of Anthropic, Open Philanthropy, 80,000 Hours, and effective altruist networks are not only about individual choices. They are about how funding, jobs, status, friendships, and career pipelines can make a community more sympathetic to frontier AI labs and less willing to demand a pause.

The legitimacy question is as important as the technical question

Several segments ask who should decide whether powerful AI systems are built, released, or constrained. Elmore rejects the idea that private labs should control these choices simply because they have technical expertise, while the host explores whether some private-sector competence arguments deserve a more charitable reading.

Future accountability is used as a deterrent frame

Elmore’s “Holly’s Basilisk” thought experiment is presented as a way to make frontier-AI workers consider how future society might judge them if AI danger becomes undeniable. She emphasizes that the goal is to stop dangerous development through legal and international processes, not to encourage vigilante action.

Select any chapter text to Deep Dive with AI

01Holly Elmore on PauseAI, DC Advocacy, and the Split from Technical AI Safety

The episode begins by introducing Holly Elmore and PauseAI US, which advocates an international treaty to pause frontier AI development until it is safe. Elmore explains her move from San Francisco to Washington, DC as part of a shift from protests at AI companies toward constituent organizing and direct engagement with elected officials. She describes a divide between technical AI safety circles and governance-focused advocates, estimates AI catastrophe risk around 50–60%, and argues that even technical alignment work requires democratic oversight, regulation, and external accountability.

PauseAI US is framed as a grassroots organization seeking an international pause on frontier AI development.

Elmore says the movement has shifted from company-facing protests toward elected officials, constituent pressure, and policy engagement.

A major fault line is whether AI danger should be handled primarily by technical insiders or by democratic governance and diplomacy.

Elmore estimates catastrophic AI risk at roughly 50–60%, while cautioning that she does not naturally think in exact probability terms.

She argues that early AI safety circles avoided public outreach for reputational reasons and that this was a mistake.

02LessWrong’s Muted Book Response and the Case for a Government AI Rollback

The conversation turns to Eliezer Yudkowsky and Nate Soares’s book “If Anyone Builds It, Everyone Dies,” which the speakers say received a surprisingly muted response from LessWrong-adjacent communities despite broader mainstream attention. Elmore presents PauseAI US as trying to give ordinary people social support to advocate for a pause. The discussion then moves to government intervention against Anthropic-related model releases, with Elmore arguing that if a company describes a model as dangerous, regulators should be willing to block or roll it back.

Elmore says PauseAI US promoted the book through launch events and continues to see people join after reading it.

The speakers criticize LessWrong’s response as insufficiently supportive of the book’s AI-risk message.

Elmore sees government action around the “Fable 5” controversy as evidence that elected authorities can restrain AI companies.

The host distinguishes current model utility from future superintelligence risk, while Elmore warns against treating early non-catastrophe as proof of safety.

The chapter ends with agreement that society may need to accept disappointing limits on AI deployment when risks are high.

03Dean Ball, OpenAI, and the Ethics of Public Criticism

The host asks Elmore about her public dispute with AI policy figure Dean Ball after Ball joined OpenAI. Elmore defends her interpretation that Ball’s comments suggest openness to private AI institutions usurping democratic authority, while the host offers a more charitable reading focused on private-sector competence. The broader debate becomes whether harsh public criticism and motive speculation are justified when the stakes are existential. Elmore argues that ordinary benefit-of-the-doubt norms can shield dangerous institutions, while the host worries that speculative psychoanalysis can undermine legitimate criticism.

Dean Ball’s move to OpenAI is used as a case study in whether AI safety figures should join frontier labs.

Elmore defends sharp criticism of people she sees as enabling private AI power over elected government.

The host pushes back against overconfident motive-reading and argues for distinguishing moral criticism from personal speculation.

Elmore argues that AI industry power depends partly on outsiders continuing to grant insiders the benefit of the doubt.

The segment highlights a tension between discourse fairness and urgency under perceived catastrophic risk.

04Beefs with Eliezer Yudkowsky and Zvi Mowshowitz over AI Safety Advocacy

Elmore discusses disagreements with Eliezer Yudkowsky and Zvi Mowshowitz over how AI safety advocates should behave in public. She credits Yudkowsky’s intellectual influence but says he objected to her strategy of calling out AI company employees. The discussion of Mowshowitz focuses on whether high-quality AI analysis and high p(doom) are enough if someone also normalizes or enthusiastically uses frontier AI products. Elmore argues that safety analysis must connect to action, regulation, refusal, or behavioral change.

Elmore says Yudkowsky opposed her public criticism of AI company employees, reflecting different norms around advocacy.

She praises Yudkowsky’s influence while criticizing what she sees as reluctance toward populist, normie-facing activism.

The host describes Mowshowitz’s AI roundups as among the best sources for tracking AI developments.

Elmore argues that high p(doom) is not enough if commentators still hype or normalize products like Claude.

She frames sympathetic treatment of Anthropic by safety commentators as a form of informal regulatory capture.

The chapter emphasizes that evaluations and warnings matter only if they inform concrete decisions.

05AI Doom Beefs, Anthropic’s EA Influence, and Holly’s Basilisk

This section examines moral consistency around using, promoting, or investing in AI while warning about AI catastrophe. Elmore argues that such behavior can be corrupting, especially for people who assign high probability to AI doom. The discussion also covers a dispute with Nirit Weiss-Blatt over attempts to associate PauseAI with a violent incident, Anthropic’s influence in effective altruist and rationalist circles, and Elmore’s “Holly’s Basilisk” thought experiment about future legal or moral accountability for people who knowingly help build dangerous AI systems.

The host acknowledges using AI tools and investing in AI stocks such as Google while expecting a possible boom-and-bust trajectory.

Elmore distinguishes limited use from promotion and argues that AI investment and enthusiastic tool adoption can undermine moral clarity.

She rejects attempts to associate PauseAI rhetoric with violence and says PauseAI is stringent against calls to violence.

Anthropic is described as a major center of gravity for rationalists, effective altruists, money, status, and AI legitimacy.

The “Anthropic defense” is framed as the argument that building dangerous AI is justified because competitors or foreign rivals would otherwise do it.

“Holly’s Basilisk” is presented as a thought experiment about future legal accountability, not a call for vigilante action.

06Critiques of Anthropic Figures: Amanda Askell, Daniela Amodei, and Joe Carlsmith

Elmore focuses on prominent people associated with Anthropic, including Amanda Askell, Daniela Amodei, and Joe Carlsmith. She objects to narratives that frame Claude’s constitution and personality work as value-setting for powerful future systems, while the host argues that focusing on current AI personality may distract from deeper alignment issues. Joe Carlsmith’s move from Open Philanthropy to Anthropic becomes a case study in whether acknowledging AI risk while joining a frontier lab is morally serious. The chapter also considers whether employees could use hiring conditions or public commitments to push labs toward stronger safety practices.

Elmore objects to the idea that a small group at Anthropic should define values for powerful AI systems.

The host argues that public focus on Claude’s personality tuning may obscure harder future alignment problems.

Daniela Amodei is criticized as presenting Anthropic in a more conventional business-first mode.

Joe Carlsmith’s decision to join Anthropic is framed by Elmore as especially disappointing because he had articulated reasons for concern beforehand.

The host suggests prospective employees could negotiate conditions or public commitments as leverage.

Elmore argues Anthropic’s revised responsible scaling commitments weaken trust in conditional safety promises.

07Open Philanthropy, Anthropic Ties, and the Case Against Bet-Hedging

The final chapter turns to Elmore’s criticism of Open Philanthropy, now referred to as Coefficient Giving, and related effective altruist organizations. She argues that they have deprioritized AI pause advocacy while funding technical safety work that often routes talent toward Anthropic or OpenAI. The host summarizes her broader critique as opposition to “bet hedging,” where AI safety actors prepare for multiple futures rather than fully committing to stopping dangerous AI development. Elmore closes by arguing for loyalty to the future people want to create, followed by a show wrap-up and donation appeal.

Elmore says Open Philanthropy gave unsatisfying reasons for not supporting AI pause advocacy.

She argues technical safety work often creates talent pipelines into frontier AI companies.

The discussion presents Anthropic as closely tied to parts of the effective altruist funding and career ecosystem.

The host frames the shared critique as opposition to AI safety leaders spreading their bets instead of prioritizing a pause.

Elmore rejects bet hedging and argues people should commit to the future they want rather than positioning for multiple outcomes.

The episode ends with a recap of AI pause advocacy, public criticism, motives, and community conflict.