Link: Microsoft claims its AI safety tool not only finds errors but also fixes them

Microsoft is launching a new feature called "correction" to address AI inaccuracies in their Azure AI systems. This feature automatically detects and corrects erroneous AI outputs before they are visible to the user.

The "correction" function is part of Azure AI Studio, which includes tools for detecting AI vulnerabilities and blocking harmful prompts. It identifies and revises inaccuracies by comparing AI-generated content with a customer’s original material.

Although this tool aims to clean up AI-generated mistakes, Microsoft acknowledges its limitations. The correction system employs both small and large language models for content alignment but is not foolproof.

Google’s Vertex AI offers a similar feature called "grounding," which cross-references AI outputs with Google Search and proprietary datasets. This helps in verifying the data AI models generate.

Microsoft emphasized that while their system improves content relevance by aligning with documented sources, it doesn't entirely ensure accuracy. Groundedness detection helps, but doesn't completely solve for accuracy.

The ongoing enhancements in AI safety features highlight the tech industry’s commitment to improving reliability in AI communications. These developments suggest a future where AI errors are less frequent and less disruptive. #

--

Yoooo, this is a quick note on a link that made me go, WTF? Find all past links here.