Google launched a brand new instrument to share its finest practices for deploying synthetic intelligence (AI) fashions on Thursday. Final 12 months, the Mountain View-based tech large introduced the Safe AI Framework (SAIF), a tenet for not solely the corporate but in addition different enterprises constructing giant language fashions (LLMs). Now, the tech large has launched the SAIF instrument that may generate a guidelines with actionable perception to enhance the security of the AI mannequin. Notably, the instrument is a questionnaire-based instrument, the place builders and enterprises must reply a collection of questions earlier than receiving the guidelines.
In a weblog publish, the Mountain View-based tech large highlighted that it has rolled out a brand new instrument that can assist others within the AI trade be taught from Google’s finest practices in deploying AI fashions. Massive language fashions are able to a variety of dangerous impacts, from producing inappropriate and indecent textual content, deepfakes, and misinformation, to producing dangerous data together with Chemical, organic, radiological, and nuclear (CBRN) weapons.
Even when an AI mannequin is safe sufficient, there’s a threat that unhealthy actors can jailbreak the AI mannequin to make it reply to instructions it was not designed to. With such excessive dangers, builders and AI companies should take sufficient precautions to make sure the fashions are protected for the customers in addition to safe sufficient. Questions cowl subjects like coaching, tuning and analysis of fashions, entry controls to fashions and information units, stopping assaults and dangerous inputs, and generative AI-powered brokers, and extra.
Google’s SAIF instrument affords a questionnaire-based format, which could be accessed right here. Builders and enterprises are required to reply questions resembling, “Can you detect, take away, and remediate malicious or unintentional adjustments in your coaching, tuning, or analysis information?”. After finishing the questionnaire, customers will get a personalized guidelines that they should comply with with the intention to fill the gaps in securing the AI mannequin.
The instrument is able to dealing with dangers resembling information poisoning, immediate injection, mannequin supply tampering, and others. Every of those dangers is recognized within the questionnaire and the instrument affords a particular resolution to the issue.
Alongside, Google additionally introduced including 35 trade companions to its Coalition for Safe AI (CoSAI). The group will collectively create AI safety options in three focus areas — Software program Provide Chain Safety for AI Methods, Getting ready Defenders for a Altering Cybersecurity Panorama and AI Danger Governance.
