Microsoft claims its new AI tool fixes hallucinations on the fly

AI hallucinations are an absolute nightmare for enterprises. Catching a bot making things up before a customer notices is already hard enough, but fixing the response on the fly is another story entirely.

Microsoft thinks it has cracked that problem. On September 24, the company launched a new feature called Correction in preview for Azure AI Studio, built directly into its safety toolchain.

Instead of just flagging a hallucination after it happens, Correction intercepts the answer, spots the inaccuracy, and rewrites the content before the end user ever sees it.

Fixing the mistake in real time

Azure AI Studio already handles vulnerability scanning, prompt injection defense, and hallucination detection. Correction takes that setup one step further by comparing outputs directly against a customer's grounding documents.

When a model starts straying from the source material, the system highlights the error, logs an explanation of why it is wrong, and rewrites the offending section.

A Microsoft spokesperson told TechCrunch that the system relies on a combination of small and large language models to align outputs with the provided source files.

However, Microsoft was careful to manage expectations. The company pointed out that groundedness detection does not solve for objective accuracy, but rather helps align generative AI responses with the specific documents provided.

That also means the correction layer itself is not immune to errors, especially since Google's competing Vertex AI platform offers a similar grounding approach that checks outputs against Google Search and enterprise datasets.

Why it matters

We think this is a very clever move that tackles the single biggest roadblock to enterprise AI adoption: trust.

For developers building customer-facing bots on Azure, this feature is worth testing immediately. Building reliable guardrails and re-prompting logic manually takes serious engineering time, so having a native rewrite tool embedded into your cloud stack removes massive friction.

Still, there are obvious trade-offs that teams need to watch closely. Running a response through secondary evaluation models and triggering an automatic rewrite inevitably adds latency and burns additional tokens.

There is also the philosophical problem of trusting an AI to fix another AI. If the validator model misinterprets the source text, it could introduce new errors or strip away nuance.

The real takeaway here is where the cloud war is moving. The battle among providers like Microsoft and Google is shifting away from raw benchmark supremacy toward building the most foolproof safety net for enterprise data.