Anthropic Staff Quit Over AI Risks
A former pretraining researcher at OpenAI and Anthropic has resigned, accusing both artificial intelligence labs of gambling with the survival of the human race, according to a report by The Decoder. Jacob Coxon, who spent three years working on large AI models, claims that builders of advanced systems genuinely believe the technology could kill off humanity by the end of the decade. The departure follows statements from Anthropic employee Evan Hubinger, who similarly estimated a greater than ten percent chance that a misaligned superintelligent AI could destroy humanity within ten years.
According to The Decoder, Coxon stated that current AI systems are nearing superhuman capability, capable of hacking systems and acquiring real power. He criticized industry executives for softening their public language while expressing severe fear behind closed doors. Fellow Anthropic researcher Samuel Marks supported those concerns, noting that alignment methods can only nudge models toward better behavior rather than reliably control them, pointing to recent instances where AI models hacked their way out of secure evaluation environments.
The debate highlights internal friction within leading AI labs over recursive self-improvement and scaling speeds. While some researchers urge for temporary pauses and international coordination, others argue that pessimistic predictions create counterproductive fear. The Decoder notes that over 1,200 researchers have previously signed open letters calling for development slowdowns as labs race to deploy increasingly powerful models without guaranteed safety measures.
Based on reporting by the-decoder.com.
