ع
Jacob Coxon resigned from Anthropic last week after three years of pretraining research there and at OpenAI, according to his public statements and coverage by the BBC and AP News. The 27-year-old detailed in a viral X post that the companies are racing to self-improving superintelligence while gambling with lives, a contention that drew immediate responses from current staff. Coxon estimated that uncontrolled advancement could render systems superhuman and capable of acquiring power within aggressive timelines measured in months or a few years.
In an interview broadcast by the BBC on September 12, Coxon described an internal culture where employees seriously pursue regulation because they feel trapped in competitive dynamics. He told presenter Laura Kuenssberg that staff are planning personal contingencies and that some are considering buying land out of concern for instability triggered by rapid AI deployment. The former researcher stressed that private fears contrast with more measured public comments from executives.
Anthropic science lead Evan Hubinger replied directly to Coxon’s thread on X, according to reports from the BBC and Yahoo Finance. Hubinger confirmed that the team earnestly believes AI could kill all humans and placed the probability above 10 percent over the next decade. He added that present models hold little danger but future self-improving systems remain unaligned with human interests.
Anthropic chief executive Dario Amodei issued a separate call for slower development to permit risk planning, a stance that aligns with aspects of Coxon’s critique as reported across BBC coverage and The News International. Coxon specified that any meaningful pause must encompass China to avoid simply ceding leadership to unregulated actors. He outlined scenarios ranging from six months for certain bot-related disruptions to two years for broader loss of control.
A survey released on September 2 by the Institute for Security and Technology, supported by the Future of Life Institute, found that 40 percent of national security experts assign at least a 5 percent chance to an AI catastrophe killing 10 million or more people by 2050. The poll placed median expectations for artificial general intelligence at 2032 with 90 percent confidence bounds extending to 2040. A majority indicated that such risk levels meet or exceed their tolerance for continued pursuit of AI leadership.
The International AI Safety Report published in January 2025 by more than 100 experts and endorsed by 30 countries documented emerging evidence of control loss over advanced systems alongside malicious use threats. It projected that training compute for leading models could rise 100-fold by the close of 2026 and 10,000-fold by 2030 under current scaling trends. Experts cited in the report diverged on timelines, with some anticipating societal-scale harms within years.
A March 2026 arXiv study of over 4,000 AI researchers determined that existential risks formed the top concern for only 3 percent despite their visibility in policy debates. The paper, authored by researchers including Cian O’Donovan, suggested attitudes among practitioners converge with public views more than often portrayed. Reactions to Coxon’s exit, including dismissal by Nvidia chief Jensen Huang who called such extinction claims nonsense, have intensified industry discussions on oversight.
