Microsoft first launched its “groundedness” detection characteristic in March. To make use of it, a genAI utility should connect with grounding paperwork, that are utilized in doc summarization and RAG-based Q&A situations, Microsoft stated. Since then, it stated, clients have been asking what they’ll do as soon as inaccurate info is detected, in addition to blocking it.
“This highlights a big problem within the quickly evolving generative AI panorama, the place conventional content material filters usually fall quick in addressing the distinctive dangers posed by generative AI hallucinations,” Microsoft Senior Product Advertising and marketing Supervisor Katelyn Rothney wrote in a weblog put up.
Constructing on the corporate’s current groundedness detection, the Correction instrument permits Azure AI Content material Security to each determine and proper hallucinations in real-time — earlier than customers of genAI functions encounter them. It really works by first flagging the ungrounded content material. Then the Azure Security system initiates a rewriting course of in real-time to revise the incorrect parts guarantee alignment with related knowledge sources.
