Times Union photo illustration / Getty Images.

When people ask me why I lose sleep over artificial intelligence, I don’t talk about killer robots. My fear is more prosaic: that we will hand over so many decisions to opaque algorithms that we end up no longer controlling our future.

This “gradual disempowerment” is the default path ahead, if we go on treating AI as less risky than a taco cart.

Advertisement

Article continues below this ad

The “taco cart” comparison is not a joke: In New York you need a license, a food safety course, and a Department of Health inspection before you can sell a plate of tacos on the sidewalk. Yet any company with enough money and talent can train a powerful AI model capable of drafting legislation, writing malware or optimizing content for addictiveness — without even writing a safety plan.

The regulatory asymmetry would be comical if it were not so dangerous.

When I tell people I teach and research AI safety, they’re often glad to hear that someone’s working on that. The truth is, as even the CEOs of AI labs admit, we are still a long way from understanding how AI systems work and how to make them safe.

To their credit, the heads of the leading AI companies acknowledge that their technology could create risks to public safety, including future

© Times Union