Two leading left-wing lawmakers are proposing to ban self-improving artificial intelligence models that can escape human controls and to start a federal department to monitor extra AI development.
Sen. Bernard Sanders, Vermont independent, and Rep. Greg Casar, Texas Democrat, underscored the urgency of Congress stopping “superintelligence” as they unveiled their legislation on Wednesday.
“If the leaders of the major AI companies acknowledge that they are losing control of this extremely dangerous technology, it is irresponsible, in my view, for society to allow them to move forward and make these products even more advanced,” Mr. Sanders said.
He said the concern extends beyond the possibility that advanced AI could lead to human extinction and includes threats to financial, defense, energy and healthcare systems “and virtually every computer that is connected to the internet.”
Mr. Casar, who chairs the Congressional Progressive Caucus, said the legislation he and Mr. Sanders are proposing would “stop the race towards disaster.”
Their bill would ban what the industry calls “recursive self-improvement,” the ability for AI models to improve their own functions without human input.
“This is not a radical notion,” Mr. Casar said. “All it does is take us back a couple years in time when humans, rather than AI, were in charge of AI research.”
The bill would also ban AI models that can design nuclear or biological weapons.
To allow more AI development to continue, the legislation would create a Cabinet-level federal agency called the Department of Artificial Intelligence to provide oversight of emerging technology before public deployment and enforcement of safety guardrails once new AI models are released.
“You’re dealing with a technology that not a whole lot of people fully understand,” Mr. Sanders said. “And what we need to do is bring together the most advanced investigators and scientists in the world who do understand this, who can give us guidance as to how we go forward.”
Mr. Casar said the department would evaluate AI models before, during and after their training periods to ensure they are safe for deployment and conduct regular inspections of the models after their release.
If the department discovers any signs that humans are losing control over the AI or other “dangerous characteristics,” it “would have the power to put an immediate pause on a system and sequester it from the internet,” he said.
The congressman said he has had conversations with leaders of Anthropic and OpenAI and pushed back on their suggestions that U.S. companies cannot pause development of potentially dangerous models without an international agreement that would force China to do the same.
“China trains their models off of American models, so stopping dangerous AI development in America will slow down dangerous advancements in China,” Mr. Casar said. “And a pause and guardrails in America will send a signal to China that we’re serious about an agreement.”
The bill from Mr. Casar and Mr. Sanders goes further in trying to slow AI development than most lawmakers have proposed, but they argue that is what is needed to keep the world safe.
“It’s, to my mind, insane not to act boldly,” Mr. Sanders said. “We want to be safe, not sorry.”
Contact the author
Please read our comment policy before commenting.