Anthropic urges AI labs to pause, warns humans risk losing control

On 4 June 2026, Anthropic, the maker of the Claude chatbot, posted a warning that its own models may soon begin designing and building successive generations with minimal human input. The company called for a coordinated pause among the world’s leading AI labs, arguing that the speed of progress now threatens the ability of people to keep the technology under control.

Anthropic’s claim that Claude now writes the bulk of its own code is the first concrete metric the company has released that suggests the AI is moving beyond a tool that merely assists engineers. In a May 2026 internal analysis, the firm disclosed that 80 % of the code merged into its main repository was generated by Claude, a dramatic jump from the low single‑digit share recorded before the 2025 launch of Claude Code. The report also notes that the amount of code a single engineer merges per quarter is now eight times higher than it was a few years ago, a figure that the company says is an overstatement because lines of code are an imperfect measure of productivity.

Media additions

These internal data points feed into Anthropic’s central warning: if the trend continues, a model could eventually design its own successors, a process known as recursive self‑improvement. The company says that, given sufficient compute, such a trajectory could reach a stage where a model autonomously creates a more capable version of itself, potentially accelerating the pace of progress beyond the ability of humans to supervise or align it.

Anthropic’s blog post, “When AI Builds Itself”, frames the pause as a precautionary option that would give society time to develop alignment research and governance structures. The post also sketches a “global coordination mechanism” modeled loosely on arms‑control treaties, arguing that a unilateral slowdown would allow less cautious labs to catch up and erode safety gains.

OpenAI’s response, published the day before the Anthropic warning, stresses that democratic governments should set the rules rather than private companies. The report, released in a Al Jazeera article, notes that OpenAI’s position is that “decisions about the pace of AI innovation should not be left to any one lab, company, or special interest group.”

Industry observers are divided. Some, like Noah Giansiracusa, a professor of mathematics, dismiss Anthropic’s call as “not a genuine call to slow down,” arguing that the business imperative to stay ahead outweighs safety concerns. Others see the warning as a rational response to the rapid rise in Claude’s coding and research capabilities, pointing to the H2S Media piece that details how the 80 % figure was achieved.

Anthropic’s push comes at a time of heightened regulatory attention. In the United States, an executive order put the onus on the labs themselves, asking them to voluntarily submit their most capable models for government cybersecurity testing before public release. The regulatory push is mirrored by industry‑wide calls for a pause, including a letter from several tech leaders that appeared in TradingView.

Anthropic’s own financial trajectory adds another layer of complexity. The company confidentially filed for an initial public offering the same week the warning was issued, with a valuation close to one trillion dollars.

Anthropic’s warning forces the industry to confront the question of whether the speed of progress is outpacing the tools we have to keep it safe. Whether the call for a coordinated pause gains traction remains to be seen, but the data it has released— Claude writing 80 % of its own code—provides a concrete baseline for that debate.