Balancing innovation and safety: Inside NIST’s AI Safety Institute


The institute's director Elizabeth Kelly details efforts to test models and create guidance to encourage responsible artificial intelligence use.

Back in February, the National Institute of Standards and Technology stood up the AI Safety Institute to establish guidelines for evaluating and testing artificial intelligence models for safety and risk.

Led by Elizabeth Kelly, the institute is working on ways to understand the risks and implications of generative AI. That was was created in the wake of the Biden administration’s AI executive order, which Kelly helped write.

She gave an update on their institute’s activities during a panel discussion at the AWS Summit on June 26 in Washington, D.C. Industry partners and potential partners for Amazon Web Services attend the conference, along with government technology practitioners.

“Our mission is to advance the science of AI safety and the same time advance the implementation of and adoption of that science,” she said.

Whether the institute’s work leads to regulations or other guidelines, Kelly said the intent is not to slow the use and adoption of generative AI. The institute merely wants the tech to advance through standards and guidance, she said

“The Biden administration firmly believes that safety breeds trust. Trust breeds adoption, and adoption breeds innovation,” Kelly said.

As she has outlined in previous remarks, Kelly described three pillars for the institute’s work:

  • Testing
  • Developing guidance
  • Developing tools

In the area of testing, Kelly said that the institute is directly testing AI models and particularly the so-called frontier models. They generate new knowledge or information for users.

“This is going to be an entirely new U.S. government capacity to directly test frontier AI models and systems before deployment,” she said.

One initial pilot project is looking at workload suspension capabilities, which is an AI capability designed to pause a task and resume it later without losing progress or context. This is important for long-term task management, adaptability to changing conditions and for safety checks.

Another priority area is balancing AI’s potential for good such as drug discovery with harmful uses, Kelly said.

“So we are also looking at barriers to development of chemical and biological weapons as well as cyber-attacks,” she said.

In the guidance area, the AI Safety Institute is working on a range of topics including how developers can plan for risks and security mitigations and to prevent malicious use.

“We’ll be putting out an overview of tools and techniques to help detect synthetic content and authenticate content,” Kelly said.

Synthetic content includes deep-fakes, AI-generated images and videos, plus artificially generated datasets used for training other AI-models.

“Our goal here is to spur a robust ecosystem for third-party evaluates and in-house evaluates, who rely on our guidance, as well as researchers, so we can all work together for public safety,” she said.

The AI Safety Institute is also working on fundamental technical research.

“We are looking at questions like interoperability of AI models as well as research and development on new and improved risk mitigation,” she said.

The institute realizes that AI is evolving quickly.

“Our work needs to be iterative and adoptive. It can’t be just issue guidance and walk away,” Kelly said. “We want to have a cycle of research, guidance and real-world use that inform each other and continually adopt and iterate.”