The p(doom) Debate: Do Lab Researchers Really Believe in Extinction Risk

On September 21, AI practitioner arohan publicly questioned whether frontier-lab researchers truly believe in extinction risk, arguing that anyone with p(doom)>0 should not voluntarily accelerate lab research, and inferring that most researchers privately hold near-zero risk beliefs. Former Anthropic researcher Ryan Greenblatt responded with firsthand observations, and the debate extended into sharp criticism of OpenAI's institutional motives.

Confirmed

• arohan's core argument: anyone who genuinely believes p(doom) is above zero should not participate in accelerating frontier research, data preparation, and training; he inferred most frontier researchers actually hold p(doom) near zero and called on them to state this publicly.
• His post was partly prompted by a conversation with a non-AI friend who was deeply fearful of AI doom.
• Ryan Greenblatt responded that nearly all capability researchers he knows at OpenAI and Anthropic—including the most senior ones—consider AI takeover risk significant, clarifying that significant means well above zero, directly contradicting arohan's inference.
• Users such as @anpaure added practitioner perspectives on whether those still accelerating training could hold near-zero p(doom).
• In a Polymarket thread, safety researcher BlancheMinerva accused OpenAI of having long and consistently acted as if destroying the world were acceptable for an extra dollar; safety researcher ciphergoth endorsed this view as not exaggerated.

Why it matters

• The dispute exposes a core tension in AI safety governance: the public cannot easily discern researchers' true beliefs about extinction risk, which directly shape trade-offs between capability pace and safety investment.
• arohan demands explicit public statements to reduce information asymmetry; Greenblatt's observations suggest concern about takeover risk is common among capability researchers, leaving the contradiction of worrying while accelerating unresolved, with no official lab response so far.

2026-09-21 ~ 2026-09-21 · 8 related posts

• Researcher Calls Out p(doom) Contradiction: Why Accelerate Labs If Doom Is Nonzero? — _arohan_ · 2026-09-21
• [source] If your p(doom) > 0, why are you accelerating frontier lab research? — _arohan_ · 2026-09-21
• [source] Greenblatt: senior capability researchers at OpenAI and Anthropic see takeover risk as significant — RyanGreenblatt · 2026-09-21
• Why do AI lab researchers accelerating training still claim nonzero p(doom)? — anpaure · 2026-09-21
• [source] AI safety researcher: OpenAI has consistently acted fine destroying the world for a buck — BlancheMinerva · 2026-09-21

3 near-duplicate retellings: _arohan_ · RyanGreenblatt · ciphergoth