Version 2

Anthropic and OpenAI have, within days of each other in early June 2026, publicly acknowledged that current AI systems show early signs of recursive self-improvement (RSI). Anthropic disclosed that Claude authored more than 80% of its production code merged in May 2026 [2][3] and called for a global slowdown in frontier AI development [7][8]. OpenAI's policy blueprint called RSI 'potentially the most consequential frontier safety issue of the coming decade' [4], and CEO Sam Altman separately stated he has 'no interest in building a super-smart AI that accomplishes some non-human goals' [6]. Both disclosures have drawn mainstream coverage from WSJ, Business Insider, and France24, and have generated debate over whether the labs are responding to genuine safety risk or using safety framing to shape regulation in their favor.

If two leading labs independently judge that systems with early self-improvement capacity are already operational, the window to establish binding oversight before RSI becomes self-sustaining is narrower than most policy timelines assume. Both labs have strong financial incentives to shape regulation in ways that favor incumbents, which complicates the signal from their public statements.

In the first week of June 2026, Anthropic and OpenAI independently released documents that together represent the most explicit public acknowledgment either lab has made that recursive self-improvement may already be underway in deployed systems. Anthropic published a blog post titled 'When AI Builds Itself,' disclosing that Claude authored more than 80% of Anthropic's production code merged in May 2026 [1][2]. The average Anthropic engineer now merges eight times as much code per day as in 2024, reliable task length is doubling approximately every four months, and Claude can sustain useful work on tasks lasting more than 16 hours [2][3]. Claude Mythos Preview had accelerated Anthropic's model-training code by roughly 52x, compared with approximately 3x for Claude Opus 4 the prior year, and Claude's success rate on open-ended coding tasks reached 76%, up 50 points in six months [3]. OpenAI's concurrent policy blueprint stated that the company sees 'early signs of recursive self-improvement in today's systems' and called RSI 'potentially the most consequential frontier safety issue of the coming decade' [4][5]. Days before the blueprint, CEO Sam Altman stated in an interview that AI should not be designed to pursue goals disconnected from human needs and that he has 'no interest in building a super-smart AI that accomplishes some non-human goals' [6] — a position consistent with the blueprint's framing but notable as a direct CEO statement.

Anthropic paired its disclosure with a call for a global slowdown in frontier AI development, arguing that future models capable of autonomously running research experiments could accelerate AI development beyond human control [7][8]. OpenAI's proposed governance architecture would establish a federal oversight body (CAISI) with mandatory evaluation authority but no power to block deployments, and would build a national framework on existing state laws — California's SB 53, New York's RAISE Act, and Illinois's SB 315 — while preempting those states' frontier safety statutes with federal law [5]. Analyst Zvi Mowshowitz described the blueprint as 'highly reasonable, well exceeding expectations,' but identified the preemption clause as its most dangerous element: once states surrender regulatory leverage, he argues, Congress may never act to replace it, leaving oversight without enforcement teeth [5].

The credibility of both labs' public positions is complicated by financial context. Anthropic's slowdown call came shortly after the company confidentially filed an S-1 at a valuation of approximately $965 billion [9][10]. Critics — including a New York Post report and multiple commentators — argued that a global slowdown would disproportionately disadvantage competitors that have not yet secured Anthropic's market position [10][11]. The story reached mainstream audiences through coverage by WSJ, Business Insider, France24, and SiliconAngle [8][12][13][7], confirming broad public pickup of the story.

On the definitional question underlying both disclosures, alignment researcher Seth Herd argued on the Alignment Forum that the first 'takeover-capable AI' will most likely be an LLM augmented with persistent memory and executive function rather than a novel architecture, and that existing alignment techniques do not address the goal drift such systems would introduce [14]. TechCrunch observed that RSI 'is just as hard to pin down' as AGI once was [15] — a caution that applies equally to governance proposals built around the term. Commentators including The Neuron note that the RSI loop is not yet closed: humans still direct goals and review outputs, making 'autonomous self-improvement' a matter of degree rather than a categorical threshold [3][1][2].