Microsoft has launched a new feature called ‘Correction’ to tackle the growing issue of AI hallucinations, where AI systems generate false or misleading information. This advanced tool is designed to identify and correct these inaccuracies, enhancing the reliability of AI-generated content. As AI continues to play a significant role in various sectors, ensuring its output is trustworthy has become a priority. With the introduction of ‘Correction,’ Microsoft aims to improve AI integrity and reduce potential errors. This feature represents a critical step forward in addressing one of AI’s key challenges and reinforces the company’s commitment to making AI systems more accurate and dependable for users and businesses. Here’s what we know about Microsoft’s latest AI enhancement.
The tech giant outlined its new feature, designed to combat AI hallucinations – a phenomenon where AI provides inaccurate information in response to a query and is unable to recognise its inaccuracy. Called “Correction,” the feature is being incorporated into Azure AI Content Safety's groundedness detection system. The tool assesses whether a generated response is based on factual information or not. Although it employs various methods to detect hallucinations, the Correction feature operates in a distinct, specialised manner to address these instances. Since it is offered exclusively via Azure, it appears to be geared towards Microsoft's enterprise customers. The company is also exploring additional ways to minimise AI hallucinations. One key aspect of this feature is its ability to explain why a particular part of the text was flagged as containing inaccurate information.
Users who want to use Correction for Retrieval-Augmentation-Generation-based (RAG) Q&A scenarios and document summarisation must have access to Azure's grounding documents. Users can activate the feature once connected. From then on, the functionality will automatically initiate a correction request whenever an inaccurate or ungrounded sentence is generated. You may think of the grounding documents as a set of rules that the AI system has to abide by in order to provide a response. It could be either the query's source content or a larger database.
Also Read: Microsoft Launches AI-backed Copilot Chatbot for Telegram Users: Know How to Use it
After that, the feature will compare the statement to the grounding document and sift out any false information it finds. To make sure the statement is understood correctly, the feature may modify it if the content is consistent with the underlying document. When first configuring the capability, users will also have the choice to activate reasoning. If this is turned on, the AI function will explain why it concluded the data was inaccurate and required rectification.
