California AI regulation bill heads to must-pass hearing

SACRAMENTO, Calif. (CN) — Artificial intelligence hasn’t yet reached a point where it could help someone create a virus that causes a pandemic.

But scientists and others at a Wednesday demonstration before California lawmakers are thinking about that day and planning for it.

“We want to close that vulnerability,” said Kevin Esvelt, an associate professor at MIT Media Lab and co-founder of Secure Bio. He added that the risk is small, “but small does not mean zero.”

Following the demonstration against the potential threat AI poses, Esvelt spoke at a press conference about Senate Bill 1047, known as the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act.

The legislation — written by Democratic state Senator Scott Wiener, of San Francisco — would make requirements for the developer of an AI model that has a certain level of computing power. Those requirements must be met before the model could be trained or released. It also would create a state entity for oversight and guidance of the models that fall under the law.

Esvelt said the existence of AI forces a question of defense versus the development of the technology. His argument is technology shouldn’t be created before people have a defense for it.

For Esvelt, that means avoiding the creation of AI that helps a bad actor create a pandemic. He also fears the possible public backlash against his field if someone abused AI to make a virus that spread worldwide.

“The state of biosecurity is relatively terrible,” Esvelt said.   

Wiener said a narrative has formed around his legislation, with the argument that someone who isn’t knowledgeable about AI has no right to an opinion about AI safety.

“My response to that is this is a societal issue,” he said.  

Saying opponents of the bill have called it “sci-fi,” Wiener questioned why they would oppose something they believe is fiction.

It’s because they know the threat is real, he added.

“We can all visualize these harms,” Wiener said. “Shutting down the grid. Shutting down the banking system.”

Current AI models aren’t yet good enough to pose that level of risk, Esvelt said. They continue to make errors that would send a bad actor down the wrong path.

But as they improve over time, the risk becomes real. At the current rate of AI advancement, Esvelt said the world might reach that point in two to five years.

Esvelt said a traditional bioweapon is anthrax, which could be spread over a city. A virus that causes a pandemic would be much easier to deploy.

“Causing a pandemic, you need to infect a handful of people,” he added.

An AI that can advise a bad actor in how to create a virus isn’t a fear everyone in Esvelt’s field has. However, many people do think it’s plausible.

That uncertainty is one reason for Wiener’s bill, said Ari Kagan, an advisor to Economic Security California Action, a sponsor of the legislation, who added the bill will impose testing of AI to determine if it has that capability someone could misuse.

The forecast wasn’t completely bleak. Esvelt said AI could potentially lead to developing technology that, for example, sterilizes someone’s breath before it enters another person’s lungs.

The question is, how to enable AI to bring that technology to fruition and not allow it to help someone cause harm on a massive scale.

Senate Bill 1047 is scheduled to be heard Thursday in the Assembly Appropriations Committee. It must pass this committee to reach the Assembly floor for a vote.

Bills that fail to pass either chamber’s appropriations committee on Thursday will die.

Subscribe to Closing Arguments

Sign up for new weekly newsletter Closing Arguments to get the latest about ongoing trials, major litigation and hot cases and rulings in courthouses around the U.S. and the world.

Email *

Submit