UK ARIA adviser warns AI safety ‘window is closing’ within five years

David Dalrymple’s warning that the AI safety “window is closing” is notable because it comes from inside ARIA, a UK state-backed advanced research agency explicitly tasked with high‑risk, high‑reward projects. He is essentially arguing that capability growth curves are now outpacing both scientific understanding and control mechanisms, and that we may hit economically transformative thresholds—AI doing most valuable tasks better than humans—within five years.([cxodigitalpulse.com](https://www.cxodigitalpulse.com/ai-safety-window-is-closing-as-systems-advance-faster-than-controls-warns-aria-expert/))

For the race to AGI, this framing re-centers the debate from abstract “existential risk” to something much more operational: not whether systems will become extremely capable, but whether we can build tools to make them predictably reliable and steerable in time. Dalrymple’s point that a single day of fully automated R&D by late 2026 could kick off further acceleration is particularly important; recursive improvement doesn’t require self‑modifying superintelligences, just tightly coupled loops between AI scientists and AI‑accelerated science.

His comments also underscore how misaligned incentives are. Economic and geopolitical pressure is pushing labs to scale faster, while safety science—mechanistic interpretability, robust alignment benchmarks, formal verification—remains under‑resourced. If ARIA’s own experts are this blunt in public, it’s a signal that governments cannot keep treating frontier AI development and AI safety as separate policy tracks.