[DigitalToday reporter Hyunwoo Choo] Leading AI researchers warned that research automation at frontier AI companies could lead to an abrupt performance leap, a so-called "intelligence explosion." Business Insider reported on Wednesday that Alan Chan (앨런 챈), a researcher at the AI think tank GovAI, said if AI automates the development of future AI models, it could create a feedback loop in which the results accelerate the next round of research.

In a conversation with reporters, Chan said, "Imagine a situation where, not every year but every month, a Mythos-level moment happens in a new domain," adding, "This is the possibility of the intelligence explosion we are talking about." He said he prefers the term "intelligence explosion" to "recursive self-improvement." That is because it allows separate discussion of attempts to automate AI research and whether such automation actually leads to a sharp improvement in AI performance.

Chan wrote a related paper with Anthropic co-founder Jack Clark (잭 클라크), OpenAI chief scientist Jakub Pachocki (야쿠프 파호츠키), Microsoft chief scientific officer Eric Horvitz (에릭 호비츠) and Nobel laureate Geoffrey Hinton (제프리 힌턴), among others. The co-authors argued that policymakers should gain a closer understanding of how companies such as Anthropic and OpenAI automate development and put in place measures such as mandating the use of embedded auditors.

In an agreement signed at the White House on Tuesday, executives from Anthropic, Google, OpenAI, SpaceXAI and Nvidia agreed to "work with independent external auditors or evaluation bodies." The agreement was not as specific as the level sought by the paper's co-authors. President Donald Trump said the agreement was "morally" binding.

The co-authors wrote that they expect automation of AI research and development to advance significantly over the next few years, and that the possibility of full automation in the same period should also be taken seriously. Anthropic said in June that, as of May, AI systems generated more than 80 percent of code approved at the company. As recently as January 2025, the share of code made by Claude was in the low single digits.

Chan said research automation itself is not inherently good or bad. It could speed up discoveries in fields such as medicine, but he added that the faster new technologies emerge, the greater the risks that need to be addressed more quickly could become.