Since leaving Google, the former CEO has invested in a number of AI startups, and he’s cautioned that any AI regulation should strike a balance to ensure it doesn’t stifle innovation.
Schmidt acknowledged that the development of AI poses dangers, but he said the biggest threats haven’t arrived yet. If and when those threats do materialize, Schmidt seems to think the world will deal with it.
“By the way, do you know what we’re going to do when computers have free will?” Schmidt said at the conference on Wednesday. “We’re going to unplug them.”
“Let’s see who unplugs who,” replied Yoav Shoham, cofounder and co-CEO of AI21 Labs, who also spoke at the event with Schmidt.
Yes, the thought of racing to unplug AI systems once they’ve gained free will — and catching that in time if it were to happen — isn’t exactly the most comforting thought experiment. But, Schmidt said researchers have conducted detailed assessments of the dangers of AI, and “the answer is: you can see the danger coming.”
It’s worth noting that the former Google CEO has invested in efforts to combat AI risks. Schmidt partnered with OpenAI to launch a $10 million grant program to support technical research with the company’s Superalignment team, which was dedicated to managing risks associated with AI. Despite the team’s disbandment last week, OpenAI will continue to move forward with the grant program, a spokesperson told Business Insider.
At the moment, Schmidt said the current form of AI isn’t that dangerous — except for disinformation, which is “out of control” and poses a “real issue for democracies.”
Disinformation has become an issue in the last couple of years since AI has emerged. Recent research using Meta and OpenAI systems indicated that a range of AI systems have learned to systematically induce “false beliefs in others to accomplish some outcome other than the truth.”
Deepfakes have also become a problem, with AI-generated porn of public figures and impersonations of political leaders. People have reported AI-generated calls faking messages from President Joe Biden. In 2022, fraudsters pleaded guilty to charges of using targeted robocalls to dissuade voters from using mail-in ballots.
Schmidt said that the real dangers of large language models are cyber and biological attacks, which aren’t yet here. But “they’re coming in three to five years,” he predicted.
Schmidt did not immediately respond to BI’s request for comment ahead of publication.
