Generative AI can hallucinate, inventing or altering information in ways that matter most in medicine and education. Microsoft built tools to detect and stop ungrounded content in real time, helping developers build AI systems that stay anchored to actual data.
Chapters
It's on your feeds, it's in your phone, it's fascinating people. Generative AI is becoming more mainstream each day. People are using generative AI to learn about new topics, pick up new skills, try new recipes, and more. You get it. And typically, you can count on AI to give you the information you're looking for. But sometimes AI gets it wrong. It confuses the plot and hallucinates, or fabricates information. AI models learn from vast amounts of data. Hallucinations happen when the model changes or adds something to the data given to it, leading to inaccuracies. We call this ungrounded content. Hallucinations aren't always bad, like when you ask AI to help you make up a story or a fun image. But sometimes it's critical that AI provides accurate information, like for medical summarization or education. That's why Microsoft is working to measure, detect, and mitigate ungrounded content with expertise gained from developing AI AI-powered products like Azure and Microsoft Copilot, we've created a range of tools to help address ungroundedness and to help our customers do the same. These advanced tools are helping businesses ground their AI apps, instruct models to stay focused on data, detect ungroundedness in real time, and more. As we work to build AI systems that do good things in the world, we're committed to making them as trustworthy and as safe as possible.
