This submit is the primary in a Constructing AI Responsibly collection, which explores high considerations with deploying AI and the way Microsoft is addressing them with its Accountable AI practices and instruments.
Technically, hallucinations are “ungrounded” content material, which suggests a mannequin has modified the information it’s been given or added extra data not contained in it.
There are occasions when hallucinations are helpful, like when customers need AI to create a science fiction story or present unconventional concepts on every thing from structure to coding. However many organizations constructing AI assistants want them to ship dependable, grounded data in eventualities like medical summarization and schooling, the place accuracy is vital.
That’s why Microsoft has created a complete array of instruments to assist deal with ungroundedness based mostly on experience from growing its personal AI merchandise like Microsoft Copilot.
Firm engineers spent months grounding Copilot’s mannequin with Bing search knowledge by means of retrieval augmented era, a method that provides further data to a mannequin with out having to retrain it. Bing’s solutions, index and rating knowledge assist Copilot ship extra correct and related responses, together with citations that enable customers to lookup and confirm data.
“The mannequin is wonderful at reasoning over data, however we don’t assume it ought to be the supply of the reply,” says Chicken. “We predict knowledge ought to be the supply of the reply, so step one for us in fixing the issue was to convey contemporary, high-quality, correct knowledge to the mannequin.”