Jul 20, 2023 - Technology

White House can't say what AI is safe

Axios' Ryan Heath, former Sen. Heidi Heitkamp, Google president Kent Walker, White House science policy head Arati Prabhakar and Aspen Security Forum director Anja Manuel. Photo: Aspen Security Forum

National security officials and contractors at the elite Aspen Security Forum are just as worried and excited about AI as the rest of the world.

Why it matters: Officials from across government are struggling to keep pace with the development of AI — with the White House admitting it has no way to know if a given AI product is safe.

What they're saying: "We should be clear that we actually don't have tools and methods today to know when something is safe and effective," Arati Prabhakar, director of the White House Office of Science and Technology Policy, told the forum.

Driving the news: Prabhakar gave an inside view of a new program to educate senators about AI announced Monday.

  • Sen. Chuck Schumer (D-N.Y.) last week gave the discussion program a dry run, and “over half the Senate attended," per Prabhakar. "It's hosted in a bipartisan way, and the quality of the conversation and questions were very good," she said.
  • Former Sen. Heidi Heitkamp said Schumer’s plan was an effort to avoid jurisdictional fights between Senate committees by involving all of them.

What they're saying: Prabhakar rejected the idea of a single federal government regulator of AI, saying “it's not a workable model."

  • Both Prabhakar and Kent Walker, Google's president of global affairs, urged patience around development of AI regulation.
  • "This is not a six-month thing," Prabhakar said, arguing that a federal privacy law would be a "fantastic step" towards constraining AI providers.
  • "We need the best regulations, not the first regulations," Walker said.

Open source AI models are a source of hot debate in the White House and Silicon Valley, but the panelists steered clear of directly criticizing the safety policies around Meta's new Llama-2 model.

  • "The (AI) horse has to be safe and effective before it's out of the barn, irrespective of whether you're putting it out as an open source or as a proprietary model,” said Prabhakar.

Of note: When Axios asked the Aspen audience of 400 to put their hand up if they agreed with the idea that the U.S. had "done enough to make AI understandable and accessible and representative,” no one put their hand up.

Go deeper