Microsoft has introduced a new feature called “correction” aimed at addressing AI inaccuracies. This feature, available to users of Microsoft Azure for their AI systems, allows automatic detection and correction of incorrect content in AI outputs.

The correction feature is currently in preview as part of the Azure AI Studio, which is a suite of safety tools designed to identify vulnerabilities, detect AI-generated “hallucinations,” and block malicious prompts. When enabled, the correction system scans AI output and compares it with a customer’s source material to identify inaccuracies.

Once an inaccuracy is detected, the system highlights it, provides an explanation, and rewrites the incorrect content—all before the user sees the error. Although this feature aims to mitigate the issues commonly produced by AI models, it is not guaranteed to be completely reliable.

Google’s cloud platform, Vertex AI, offers a similar feature that “grounds” AI models by validating outputs against Google Search, a company’s own data, and forthcoming third-party datasets.

In a statement to TechCrunch, a Microsoft spokesperson explained that the “correction” system utilizes both small and large language models to align outputs with grounding documents. However, the spokesperson also noted that while groundedness detection aids in alignment with source documents, it does not necessarily ensure accuracy.

Source link