If artificial intelligence has become the tech Wild West, some new safety features from Microsoft’s Azure are meant to rein it in. Microsoft shared a series of tools on Thursday that it says will help its customers’ AI models prevent hallucinations — or the tendency of chatbots to make things up. The features are in its Azure AI, a cloud-based service that provides support and technologies to developers and organizations.
One such feature is known as groundedness detection, which is designed to identify text-based hallucinations.
Microsoft says the new feature will find and flag “ungrounded material,” or content that doesn’t seem to be anchored in facts or common sense, in chatbot responses to help improve their quality.
Read more: The biggest AI chatbot blunders (so far)
Microsoft launched its own chatbot, Copilot, in February 2023. It also has an extensive partnership with OpenAI that includes Azure OpenAI Service, which gives developers the ability to build their own AI applications through direct access to OpenAI models backed by Azure. Azure AI’s customers include consulting firm KPMG, telecomm giant AT&T, and Reddit.
Among the other tools rolled out on Thursday are prompt shields, which block attacks on generative AI models, like prompt injections or malicious prompts from external documents that steer models away from training and safety guardrails.
“We know that customers don’t all have deep expertise in prompt injection attacks or hateful content, so the evaluation system generates the prompts needed to simulate these types of attacks,” Sarah Bird, Microsoft’s chief product officer of responsible AI, told The Verge, adding that customers will then see a score and outcomes based on a model’s performance in these simulations.
Azure AI will also roll out two other monitoring and safety features soon, Microsoft said.
These types of issues, though seemingly benign, have resulted in some awkward (and some highly problematic) gaffes on the part of AI-powered text and image generators. Google’s Gemini AI sparked controversy in February after generating historically inaccurate images like racially diverse Nazis. OpenAI’s ChatGPT recently completely off the rails last month with gibberish and hallucinations that left users scratching their heads.