Policymakers and Safeguards for AI

By Jeffrey Dastin

NEW YORK (Reuters) – Policymakers aiming to recommend safeguards for artificial intelligence are facing a formidable challenge: science that is still evolving.

AI developers are grappling with how to prevent abuse of novel systems, offering no easy fix for government authorities, according to Elizabeth Kelly, director of the U.S. Artificial Intelligence Safety Institute.

Concerns in Cybersecurity

Speaking at the Reuters NEXT conference in New York, Kelly emphasized that cybersecurity is a major concern. She noted that ways to bypass guard rails established by AI labs, referred to as "jailbreaks," can be surprisingly simple.

> "It is difficult for policymakers to say these are best practices we recommend in terms of safeguards, when we don't actually know which ones work and which ones don't," Kelly stated.

Challenges of Synthetic Content

Technology experts are exploring how to vet and protect AI across various dimensions. One significant area of concern is synthetic content. Tampering with digital watermarks, which indicate AI-generated images, remains a challenge for authorities trying to provide guidance to the industry.

The U.S. AI Safety Institute

The U.S. AI Safety Institute, established under the Biden administration, is tackling these issues through partnerships across academia, industry, and civil society.

Kelly emphasized that AI safety is a "fundamentally bipartisan issue" in response to questions about the future of the institute under a possible Trump administration.

Kelly recently hosted the inaugural gathering of AI safety institutes from around the world, which was held in San Francisco last month.

International Collaboration

When asked about the outcome of these meetings, Kelly reported that the 10 participating countries are working towards interoperable safety tests with the involvement of more technical experts, or "nerds," rather than the typical diplomatic participants.

> "It was very much getting the nerds in the room," she remarked.