News
Microsoft Debuts AI Feature To 'Correct' Hallucinations
A new feature in Microsoft's Azure AI Content Safety service enables AI systems to correct hallucinations on the fly, resulting in more accurate and useful outputs.
The "correction" capability, now in preview, can be accessed in the Azure AI Content Safety service via the groundedness detection feature, which is also still in preview. Azure AI Content Safety is a set of APIs designed to mitigate potentially harmful or ungrounded AI content. One of these APIs enables groundedness detection -- the ability for an AI to flag hallucinations, or outputs that are not grounded in its training data.
The new correction feature goes a step further, enabling the AI to rewrite outputs that have been flagged as hallucinations.
As Microsoft explained in a blog post this week, correction unlocks a "first-of-its-kind capability" that improves on existing mitigations, like content filtering.
"[F]iltering is not always the mitigation that makes sense and can result in a poor user experience when it outputs text rendered incoherent without redacted content," the blog read. "Correction represents a first-of-its-kind capability to move beyond blocking."
To use the correction capability, app developers must first activate it. When the groundedness feature detects that the app's AI model has produced a hallucination, it triggers the correction capability, prompting the AI to compare the flagged output to its training data. If the AI finds more accurate information in the data, it revises the flagged output. All of this occurs in the background; the end user only sees the corrected output, not the hallucination.
Additionally, the feature will not issue a correction if there's no training data to support one. "If the sentence lacks any content related to the grounding document, it may be filtered out completely," per the blog.
Microsoft sees this feature as a way to lower the barrier to AI adoption, especially in industries like healthcare, where hallucinations can have significant consequences.
More information on the correction capability is available here.