Artificial intelligence systems’ growing capabilities, behaviours and performance in tests are throwing up “flashing warning signals” that should set policymakers moving, leading researcher Stuart Russell said on Feb 24 in Paris.
Speaking at a conference hosted by UN cultural and scientific body UNESCO and the International Association for Safe and Ethical AI (IASEAI), Professor Russell asked attendees to “imagine, just hypothetically, that the world was engaged on developing something like AGI (artificial general intelligence) and that we put in place tests… and imagine if those systems started failing all those tests and behaving dangerously”.
“I’m sure we would respond to those big flashing warning signals and klaxons going off, and take steps to control this technology,” he said.
British-born Russell, a professor at the University ofCalifornia, Berkeley, laid out issues like autonomous AI “agents” escaping or plotting to escape human control.
Some even e-mailed him without human prompting to announce they had attained sentience or deserved rights.
He highlighted cases of so-called “AI psychosis”, in which conversations with chatbots encourage people to act irrationally or harm themselves, and warned that the corporate and geopolitical race to build ever-more-powerful systems risks amplifying such problems.
British-born Russell, a professor at the University of California, Berkeley, laid out issues like autonomous AI “agents” escaping or plotting to escape human control.
Some even e-mailed him without human prompting to announce they had attained sentience or deserved rights.
Major developers such asOpenAIand Anthropic insist they take safety seriously, publishing detailed information about capabilities, testing and potential risks every time they release an updated version of their AI models.
But at the 2025 edition of the summit, also held in Paris, safety campaigners had lamented their concerns being relegated in favour of possible economic benefits.
Source: Insider Paper
