Google announces new red teaming system for gen AI

As Google positions its upgraded generative AI as instructor, assistant, and advice guru, the corporate can be making an attempt to show its fashions into a nasty actor’s worst enemy.

“It is clear that AI is already serving to folks,” mentioned James Manyika, Google’s senior vice chairman of analysis, know-how, and society, to the gang on the firm’s Google I/O 2024 convention. “But, as with all rising know-how, there are nonetheless dangers, and new questions will come up as AI advances and its makes use of evolve.”

Manyika then introduced the corporate’s newest evolution of crimson teaming, an {industry} customary testing course of to search out vulnerabilities in generative AI. Google’s new “AI-assisted crimson teaming” trains a number of AI brokers to compete with one another to search out potential threats. These educated fashions can then extra precisely pinpoint what Google calls “adversarial prompting” and restrict problematic outputs.

The method is the corporate’s new plan for constructing a extra accountable, humanlike AI, however its additionally being bought as a solution to tackle rising considerations about cyber safety and misinformation.

The brand new security measures incorporate suggestions from a crew of specialists throughout tech, academia, and civil society, Google defined, in addition to its seven ideas of AI improvement: Being socially helpful, avoiding bias, constructing and testing for security, human accountability, privateness design, upholding scientific excellence, and public accessibility. Via these new testing efforts, and industry-wide commitments, Google’s trying to placing product the place its phrases are.