Nature commentary argues AGI has effectively already arrived

Four University of California, San Diego academics — philosopher Eddy Keming Chen, machine-learning researcher Mikhail Belkin, linguist Leon Bergen and data scientist David Danks — published a Comment in Nature arguing that, by reasonable standards, current large language models already constitute artificial general intelligence.

Their case rested on separating what intelligence requires from what humans happen to have as a byproduct of evolution. AGI, they argued, does not require omniscience, embodiment or superhuman ability; no individual human meets that bar either. Instead they proposed judging systems by breadth across domains and depth within them, comparable to ordinary human cognition, and pointed to LLM performance on behavioural tests — including variants of the Turing test — as evidence the threshold had already been crossed.

The argument drew on the authors’ different disciplinary vantage points, combining philosophical analysis of what a definition of intelligence should demand with empirical claims about model capability. It was explicitly framed against stricter definitions of AGI common in both AI-safety and AI-boosting circles, which the authors characterised as setting the bar high enough to exclude most humans as well as current models.

The piece drew immediate pushback, including from cognitive scientist Gary Marcus, who argued in response that the case overstated what current systems could reliably do and that redefining AGI downward risked obscuring rather than resolving the underlying disagreement about capability. The exchange illustrated a recurring pattern in the field: as models cleared benchmarks once treated as proxies for general intelligence, disputes shifted from what models could do to what the word “AGI” should be taken to mean.