How soon do AI researchers expect artificial general intelligence?

The largest survey of its kind — 2,778 researchers published at top AI venues — yields two patterns worth holding. First, timelines compressed sharply in a single year: the aggregate estimate for unaided machines outperforming humans at every task reached 10% by 2027 and 50% by 2047, the latter 13 years earlier than the same team's survey just 14 months prior, with 21 of 32 shorter-term milestones now expected sooner (autonomous payment-site construction, indistinguishable AI music, autonomous LLM fine-tuning all ≥50% likely by 2028). Yet full automation of all human occupations was forecast far later (50% by 2116), exposing a gap between "outperforms at tasks" and "automates all jobs."

Second, and more striking: a majority of participants placed at least 5% credence on AI causing human extinction or similarly severe permanent disempowerment, consistent across four differently-phrased questions, with 38–51% putting ≥10% on extinction-level outcomes. The biggest area of consensus, the authors note, is on how wide-open the possibilities are — expert disagreement is itself the finding.

This is a discourse anchor for AI-futures writing: it quantifies expert opinion as fast-moving and deeply uncertain, not converged. It grounds the risk register that Where do frontier AI models actually pose the greatest risk today? measures empirically — expert credence on catastrophe is high even where measured capability thresholds are not yet crossed — and it tempers timeline-confident narratives with the surveyed spread.