Anthropic scientist puts the odds of AI destroying humanity above ten percent this decade
decoder ·
Anthropic AI safety researcher Evan Hubinger estimates there is a greater than ten percent chance that a misaligned superintelligent AI could destroy humanity within the next decade. His comment followed the departure of Jacob Coxon, who worked on pretraining at OpenAI and Anthropic and accused both companies of gambling with humanity's survival. Coxon argues that AI systems are close to becoming superhuman and that developers privately fear the risks yet publicly soften their language. He claims Anthropic understands the dangers but feels trapped in a race it must win because other labs would not act responsibly, a stance he calls a 'hubristic gamble.' Fellow Anthropic researcher Samuel Marks supports these concerns, noting that current methods cannot reliably align AI and that many employees want to slow down. Coxon suggests costly actions, including a temporary ban on advancing model capabilities, while remaining somewhat optimistic about international coordination. The article notes that fears focus on recursive self-improvement and that some researchers criticize such warnings as fearmongering that could cause more harm than the technology itself.
decoder