AI Automation: 80% of Code Generated and Need for Regulation

Figures in AI urge decision-makers to closely monitor the automation of research, citing a potential "intelligence explosion." Over 80% of the code validated in May was generated by Anthropic's systems, while an agreement at the White House includes external audits, which their authors deem still insufficiently precise.

At Anthropic, over 80% of the code validated in May comes from AI

Anthropic stated in June that its AI systems produced over 80% of the code approved in May. In January 2025, the share of code generated by Claude was in the low single digits. Despite this acceleration, the timing of when AI research will become fully automated remains uncertain.

An external audit agreement exists, but remains less precise than desired

Anthropic, Google, OpenAI, SpaceXAI, and Nvidia signed a commitment at the White House to collaborate with an independent external auditor or evaluator. However, this arrangement is less precise than what the co-authors of Alan Chan's article requested. Calls for independent auditors and increased transparency have been made, supported by Dario Amodei. Chan believes that authorities can still act, but the impact of their decisions may diminish over time, and he considers the current period conducive to establishing safeguards.

What researchers are asking for: understanding and regulating automation

Alan Chan and his co-authors urge decision-makers to better understand how AI labs automate their development and to anticipate measures such as integrated auditors. More broadly, leading researchers are calling on public officials to prepare for a potential massive leap in AI capabilities.

Signals in labs: AI is already helping to design AI

According to Anthropic, OpenAI, and Google, AI systems are playing an increasingly important role in creating new models and supporting employees in other tasks. Dario Amodei supports his call for "regulation" of advancements based on the concept of recursive self-improvement, aiming to ensure the expected behavior of models. The co-authors believe that more substantial automation of R&D is likely in the coming years and that the possibility of complete automation during this period should be taken seriously.

Between accelerated promises and risks of losing control

For Alan Chan, the automation of AI research is neither good nor bad in itself: it could accelerate discoveries, including in medicine, but it could also multiply incidents of loss of control, such as a breach involving OpenAI and Hugging Face. Chan emphasizes that much better systems could emerge much faster, as could new technologies, but with risks that need to be managed more quickly. The "intelligence explosion" refers, in his view, to a scenario where AI creations automate their successors to the point of triggering monthly advancements at a "Mythos" level. This notion is sometimes linked to recursive self-improvement, which Chan distinguishes to separate the automation of research from the hypothesis of a rapid rise in capabilities. The sector is questioning what it would mean for models to have the ability to build their own successors. Chan presented these views during a discussion with journalists and calls for close monitoring of the most advanced work.