Sandeep Shilawat: Continuous Red-Teaming for Trustworthy AI

Sandeep Shilawat argues that simply training AI systems to share human values is a failed approach. He believes continuous, adversarial testing with real enforcement is necessary to make autonomous AI systems safe. Shilawat recently published "Trustworthy AI," where he discusses these ideas. He notes that AI has been evolving rapidly, especially since generative AI moved to the cloud. Events like the Ukraine war and Operation Epic Fury provided strong motivation for writing his book. He highlights a key misconception about red-teaming AI systems: it shouldn't be a one-time event. Because AI is always active and can perform unexpected actions, continuous red-teaming is essential. This ongoing testing is crucial for identifying potential issues. The bottom line is that the industry needs to shift from occasional testing to continuous red-teaming to ensure AI safety and reliability.