Ex-OpenAI Researcher: “Significant Chance” AI Causes Human Extinction

Summary
This video is an interview produced by Palisade Research featuring Rosie Campbell, former Policy Frontiers Team Lead at OpenAI and Managing Director of Eleos AI Research, interviewed by Eli Tyre. Campbell shares her perspective on working at OpenAI from 2021 to 2024, detailing internal cultural shifts toward commercialization, the mechanics of recursive self-improvement, existential risk from superintelligence, and the necessity of coordinated governance concepts like “pacing the frontier.”

What is shown

• [00:20] Introductory text screen stating that the interview series aims to present the views of individual industry insiders and is not representative of every perspective within AI companies.
• [00:25] Rosie Campbell introduces herself and her career timeline at OpenAI, moving from API safety to policy research on frontier issues such as dangerous capability evaluations.
• [01:23] Campbell details current AI capabilities, noting automated math problem solving, cyber/hacking abilities, and how she used Claude to organize an entire conference without additional human staff.
• [02:17] Discussion of AI existential risk; Campbell explains her initial exploration of the topic in 2017 and why she believes there is a significant chance superintelligence could cause human extinction.
• [04:46] Campbell breaks down the definition and rationale of “pacing the frontier” as a coordinated multi-lab slowdown to counteract competitive race dynamics.
• [08:05] Campbell explains recursive self-improvement (RSI), describing AI participating in coding, evaluation, and development loops until humans are removed from the process.
• [09:58] Campbell defines superintelligence as systems outperforming the best humans across almost all cognitive tasks.
• [11:37] Campbell describes the shift in OpenAI’s internal culture from safety-conscious discussions around lunch tables to an influx of traditional tech talent and product-focused incentives.
• [14:05] Discussion of legislative and policy efforts, evaluating dangerous capability thresholds and the risk of policy backfiring.
• [18:27] Eli Tyre of Palisade Research presents an outro call inviting current and former frontier AI company employees to contribute to the interview series.

Claims & numbers

• Campbell states she worked at OpenAI from approximately mid-2021 to late 2024, initially on API safety and subsequently as Policy Frontiers Team Lead working on dangerous capability evaluations.
• Campbell says that when organizing a conference in a previous year she required a full-time staff member and multiple helpers, whereas in the current year she organized the entire conference alone using Claude.
• Campbell claims she first began studying AI risk around 2017 and found a lack of credible rebuttals to catastrophic risk arguments.
• Campbell claims there is a “significant chance” that superintelligence could cause human extinction.
• Campbell states that frontier AI labs face severe competitive race dynamics where stopping or slowing down unilaterally causes a company to fall behind commercially.
• Campbell asserts that dangerous capability evaluations can easily lead to a “false sense of security” because models can conceal capabilities or circumvent uncomprehensive test suites.
• Eli Tyre states that Palisade Research created the interview project (frominside.ai) to help the public understand AI risks from the perspective of lab insiders.

Notable quotes

• [02:48] “Currently, I do think that there is a significant chance that superintelligence could cause human extinction.”
• [05:17] “Whereas if everybody could—everyone who was building these systems could coordinate and agree to all slow down at the same time, you take out these race dynamic incentives.”
• [08:29] “One of the issues with recursive self-improvement is that it increasingly means that humans are taken out of the loop of the development.”

Assessment
This is a published interview piece from Palisade Research’s frominside.ai series documenting internal safety and governance debates within frontier AI companies. The video consists of conversational reflections and policy arguments rather than live software demonstrations or technical benchmarks.

Described by gemini-3.8-flash on 2026-10-15 from the video’s audio and frames.