More than a third of the world’s scientists doing research in Artificial intelligence they believe that the decisions made by AIs could trigger, within the next century, a catastrophe equal to or even more serious than an all-out nuclear war.
That is the terrifying conclusion of a survey carried out by Julian Michael and colleagues at the New York University Data Science Center, collecting the views of 327 researchers who recently co-authored scientific papers on AI and natural language processing. 36% of them, in fact, think that it is possible for AI to trigger a global catastrophe similar to that of a nuclear war. The survey was recently posted on the arXiv preprint server.
As explained in New Scientist magazine, advances in recent years have enabled much greater use of AI models with impressive capabilities, including writing computer code and creating images from simple text prompts.
Various military commanders in the United States have already expressed their reluctance about the possibility of arming intelligent drones with nuclear weapons. But countries like Russia are already working on it, and are currently developing a ‘drone torpedo‘ with autonomous capabilities capable of launching, by itself, a large-scale nuclear attack.
The concern reflected in the survey was much higher among women. 46% of the researchers surveyed, in effect, agreed with the premise.
According to Michael, it is quite possible that the survey even underestimated the number of researchers who believe that AI poses serious risks. Many of them, indeed, would have agreed if the danger posed were less extreme than that of all-out nuclear war. For example, says Michael, 57% of respondents viewed the development of large AI models as “significant steps towards general AI.” That is, to an AI with intellectual capacities equal to those of humans. And another 73% agreed that the automation of work due to the use of AI could lead to social changes with the same impact as the industrial revolution.
All in all, and given that significant advances in the development of AI are expected in the coming years, the fact that ‘only’ 36% of those surveyed see a catastrophic risk as possible could be considered encouraging, since the remaining 64% don’t think the same way.