Artificial Intelligence: Microsoft develops a feature that could fix errors in real time

Microsoft has introduced a new Remediation feature tied to the Azure AI Content Safety service that helps resolve hallucination issues Artificial intelligence (AI) in real time, with the ability to identify, flag and rewrite inaccurate AI results.

The technology company continues to work on its commitment to guarantee trustworthy AI, within the framework of its ‘Trustworthy AI’ initiative, which is focused on helping clients use and create secure AI, “that ensures privacy and has the highest levels of protection.”

In this regard, Microsoft has announced new capabilities and functionalities to improve the protection and privacy of AI systems, so that customers can take advantage of them and mitigate the risks that this technology can cause.

Among these new features, the American technology company has presented a new Correction function, with which it is able to identify and resolve hallucinatory behaviors when using AI models. Specifically, this tool can detect and rewrite incorrect content in AI results, automatically.

As detailed in a statement on its website, this Correction function is integrated into the Azure AI Content Safety service, the set of tools focused on offering advanced protection to achieve responsible AI. These tools include functions such as detecting vulnerabilities, blocking malicious messages, and the hallucination identification system.

Thus, this new Correction feature is based on the aforementioned hallucination identification system for large language model (LLM) results, which was introduced in March of this year, using a ‘grounding’ feature designed to identify text-based hallucinations for applications that use documents, such as those that summarize or act as ‘co-pilots’ for questions and answers.

Correction now joins this toolkit in preview, so once enabled it will scan AI results and identify inaccuracies in the information displayed, then rewrite the response to automatically correct the error.

It should be noted that this tool will identify potential errors by comparing the AI ​​results with the original material from the client in question. Also, when it identifies an error, it will also provide information on why it is an incorrect response. All of this is done before the end user using AI applications can see the error.

However, even though it is a proofreading tool, Microsoft has explained that it is not error-free. This is because, as a Microsoft spokesperson explained to TechCrunch, it relies on small and large language models to “align the results with the base documents.” Therefore, if these models also have errors, it is possible that the error will be passed on to the Proofreading function.

“It is important to note that baseline detection does not solve ‘accuracy’, but it helps align generative AI results with baseline documents,” the Microsoft spokesperson explained.

The company also stressed the importance of having correction capabilities for AI systems, given that concerns about hallucinations “have slowed down many GenAI implementations in higher-stakes domains,” as well as prevented broader implementation of Copilots.

By making the Correction feature available, Microsoft is confident that organizations will be able to deploy conversational interfaces to their customers with greater confidence.