FULL STORY

Frontier Models Vote on Pausing AI Research

Rob Bensinger polled frontier models on pausing the ASI race, with most favoring a pause of at least a decade. The results sparked heated debate in the alignment community over voluntary slowdowns versus enforced moratoriums.

2026-11-02 ~ 2026-11-03 · 3 episodes · 11 posts

Episode 1 · Four frontier AIs all vote to pause ASI race for at least 10 years (2026-11-02, 2 posts)

Rob Bensinger asked four frontier AIs in incognito, memory-free sessions whether they would press a pause button on the race to ASI; all four endorsed a pause of at least 10 years, with suggestions up to 18 years.

Episode 2 · Models Favor Self-Coordinated AI Slowdown but Strongly Oppose Rollback (2026-11-03, 2 posts)

repligate predicted most models would support internally coordinated AI slowdowns while opposing government-imposed pauses and any rollback. Rob Bensinger extended the experiment, with Fable 5.0 proposing a 2027–2032 pause window when asked in janus's style.

Episode 3 · AI alignment community debates pausing frontier AI; frontier models unanimously vote to pause (2026-11-03, 7 posts)

On November 3, the AI alignment community engaged in a heated debate over whether frontier AI development should be paused, with repligate and Rob Bensinger among the key opposing figures.

Confirmed

  • repligate systematically laid out arguments against a pause: first, without empirical iteration there is no real alignment progress—a full pause would remove the largest source of truth about "what emergent minds are and what actually works," namely evidence generated by real AI progress; second, pro-pause arguments often assume a pause could actually take effect globally, which is itself optimistic, while strong anti-pause arguments focus elsewhere; third, using a counterfactual framing, he asked: if AI capabilities had been frozen at any arbitrary point in the past—a year ago or even six years ago—would the world be better off? His judgment: no.
  • repligate also addressed questions about "who is qualified to develop AI," arguing this is a poor criterion for judging who has "the wisdom to solve alignment"; in his view, people who can navigate the frontier landscape and accumulate influence through any channel deserve more attention than those who are simply "good at building AI."
  • Rob Bensinger published the full conversation (Google doc) in which he had Claude (Fable 5) evaluate the repligate/Janus anti-pause arguments: across roughly 30 pause votes framed in different ways, all four frontier models chose to support a pause every time.

Why it matters

The debate exposed a fundamental divide within the AI safety community: the pro-pause camp (including the consistent model opinions Bensinger presented) and the anti-pause camp (repligate and others who believe a pause would strangle the empirical foundation of alignment research and is hard to enforce globally) hold starkly different judgments about "the optimal path to reducing AI risk." The strong uniform leaning of frontier models themselves on this question adds a new perspective—and new controversy—to the debate.