Clark, one of Anthropic’s co-founders, said methods for completely disabling an AI system should be addressed through public policy in case AI software becomes extremely dangerous.
Clark said many companies, including Anthropic, have various methods for shutting down their systems. He said the requirements and oversight mechanisms for an “emergency shutdown button” should be part of the broader policy debate on AI.
Proposal for rules governing an emergency shutdown mechanism
Responding to the BBC’s question about whether companies should be required to have an emergency shutdown button and whether the mechanism should be controlled by a third party, Clark said this was an issue that the public might want to understand and that policymakers might want to regulate in the future.
Clark said specific statistics would not be useful in assessing the possibility that AI could destroy all of humanity, and argued that leaving the AI sector completely unchecked was a bad idea.
“We are gambling by taking enormous risks,” Clark said. He stressed that the sector’s course needed to change.
Former researcher accuses companies
Jacob Coxon, who worked as a researcher at OpenAI and Anthropic over the past three years, argued when resigning from Anthropic that the two companies had not acted responsibly enough on AI.
Coxon claimed that the companies were engaged in an unchecked race to develop AI systems capable of improving themselves. He alleged that these systems could soon evolve into superhuman entities capable of hacking everything.
Coxon also claimed that some industry employees developing AI models seriously believed these systems could kill humanity within 10 years.
