EA Community Polls Reveal AI Alignment Divides
New poll data shows deep splits in AI alignment beliefs among 15 researchers.
The Effective Altruism (EA) community has released its second round of community polls focusing on contentious issues in AI alignment, designed to surface disagreements among researchers and practitioners. Led by contributors including Scott Alexander (ACX), David Manheim (ALTER), and Jeff Sebo (NYU), the poll collected responses from 15 alignment researchers and invites broader community participation via the EA Forum. The initiative is backed by BlueDot Impact and seeks to compare panel responses with public sentiment to identify areas of consensus and divergence within the field.
The poll covers 20 forward-looking statements on AI ethics, safety, and consciousness, including whether benchmarks will become useless due to eval awareness, whether current AI models can suffer, and whether deceptive AI will evade mechanistic interpretability tools. Participants are asked to express their credence in each statement, with results expected to inform CaML’s research priorities and future policy discussions. The organizers encourage nuanced engagement in comments and invite readers to subscribe to a Substack for updates on the comparative report, which will analyze panel vs. community responses and highlight core areas of disagreement in AI alignment.
- 15 alignment researchers, including Scott Alexander and David Manheim, have responded to the second EA community poll on AI alignment controversies.
- The poll includes 20 statements on topics like AI suffering, benchmark reliability, and S-risk vs. x-risk priorities, with results aimed at guiding CaML’s research.
- Community responses will be compared to researcher panel data, with findings published on EA Forum and a new Substack.
Why It Matters
Highlights deep divides in AI safety priorities and could reshape research funding and policy decisions.