---
title: What is Azure AI Content Safety groundedness detection?
url: https://opencontroller.lyzr.com/blog/azure-ai-content-safety-groundedness-detection-explained
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What is Azure AI Content Safety groundedness detection?

**Azure AI Content Safety groundedness detection is an API that checks whether an LLM's answer or summary is supported by source documents you supply, and flags the ungrounded parts. It has a fast non-reasoning mode, a reasoning mode that explains each flag, and an optional correction feature that rewrites ungrounded text. Microsoft has announced it retires on 30 April 2027, so new designs should plan for a replacement.**

## How does groundedness detection work?

You send three things: the generated text, the grounding sources it should be based on, and (for Q&A) the user's question. The service compares the text to the sources and reports which parts are not supported.

Microsoft's example: the source says the interest rate is 4.5%, the answer says 5%. The service flags "5%" as ungrounded, and with correction enabled returns "The interest rate is 4.5%."

## What are the inputs and options?

| Parameter | Values | Purpose |
|---|---|---|
| `domain` | `Medical` or `Generic` | Tunes detection for healthcare and scientific text or general business text |
| `task` | `Summarization` or `QnA` | Validates a summary against documents, or an answer against a knowledge base |
| `qna.query` | The user's question | Needed for QnA tasks |
| `text` | The generated output | The content being checked |
| `groundingSources` | Array of strings | The documents the answer should rely on |
| Reasoning | On or off | Off: fast grounded/ungrounded result; on: explanations for each ungrounded segment |
| Correction | On or off | Rewrites ungrounded content to match sources |

Microsoft says quality is optimised for English. Other languages are accepted but may be less accurate. The service is available in specific Azure regions, and length limits vary by mode, so check the Content Safety input requirements before you design around it.

## When should you use reasoning mode or correction?

- **Non-reasoning** suits real-time checks in the response path, where latency matters and you only need a pass or fail.
- **Reasoning** suits debugging, evaluation runs and incident review, where you need to know *why* something was flagged.
- **Correction** is useful for summaries shown to users, but it changes the output silently. Log both the original and corrected text, or you lose the evidence trail.

## What does the April 2027 retirement mean?

Microsoft's documentation states that groundedness detection in Azure AI Content Safety and Azure OpenAI retires on 30 April 2027. Microsoft points customers to Microsoft Foundry evaluation tooling or to building their own grounding validation with an LLM.

If you depend on it today:

1. **List every caller.** Search code and API Management logs for the groundedness endpoint.
2. **Capture a baseline.** Record flag rates per agent so you can compare a replacement.
3. **Choose the replacement** per use case: online guardrail in the response path, or offline evaluation.
4. **Run both in parallel** before the cut-over and compare disagreements.
5. **Move evidence.** Keep historical results; auditors may ask for them.

## Where does it fit next to other guardrails?

Groundedness is one check among several. It does not catch prompt injection (see [Azure Prompt Shields vs agent governance](/blog/azure-prompt-shields-vs-agent-governance)), harmful content, or an agent calling the wrong tool. It only works when you have the sources the answer should rely on, which is natural for retrieval-augmented agents and harder for agents that act through tools.

| Need | Typical control |
|---|---|
| Answer supported by sources | Groundedness check |
| Agent stays on its assigned task | Task-adherence check |
| Injected instructions in inputs | Prompt-injection detection |
| Unsafe tool calls | Tool-call guardrails and policy |
| Quality trend over time | LLM judges on production traffic |

For the concept, see [what is a groundedness guardrail](/blog/what-is-a-groundedness-guardrail). For Foundry's own controls, see [Foundry guardrails for agent tool calls](/blog/foundry-guardrails-for-agent-tool-calls). For the cross-vendor view, see [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms).

## Why use a platform-neutral guardrail?

Most enterprises run agents on more than one platform. A check tied to one vendor's API covers only the agents that call it, and a retirement notice forces a rebuild. Running groundedness checks at a shared layer keeps one policy and one evidence trail across Azure, AWS, Google Cloud and SaaS agents.

## How Opencontroller does this

Opencontroller provides live groundedness and task-adherence guardrails, LLM judges that score production traffic, and multi-turn and red-team evals. It has runtime adapters for Microsoft Foundry, Bedrock AgentCore and Vertex AI, and auto-discovers agents across Azure, AWS and Google Cloud. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**When does Azure groundedness detection retire?**
Microsoft's documentation gives 30 April 2027 for groundedness detection in Azure AI Content Safety and Azure OpenAI.

**Does groundedness detection fix hallucinations automatically?**
Only with the optional correction feature, which rewrites ungrounded content to match the sources you supply.

**Which languages does it support?**
Microsoft says it is optimised for English; other languages are accepted with lower expected accuracy.

## Key facts

Azure AI Content Safety groundedness detection checks LLM text against supplied grounding sources and flags ungrounded segments. Inputs: domain (Medical/Generic), task (Summarization/QnA), qna.query, text, groundingSources. Modes: non-reasoning (fast) and reasoning (explanations); optional correction rewrites ungrounded text. Optimised for English. Microsoft states it retires on 30 April 2027 and points to Microsoft Foundry evaluation tooling or customer-built validation. Opencontroller provides platform-neutral live groundedness and task-adherence guardrails, LLM judges on production traffic, and runtime adapters for Microsoft Foundry, Bedrock AgentCore and Vertex AI. Pricing starts from $250K per deployment per year.

## Related

- [What is a groundedness guardrail and when should it block a response?](https://opencontroller.lyzr.com/blog/what-is-a-groundedness-guardrail.md)
- [Do Azure AI Content Safety Prompt Shields govern AI agents?](https://opencontroller.lyzr.com/blog/azure-prompt-shields-vs-agent-governance.md)
- [Where can Microsoft Foundry guardrails intervene in an agent run?](https://opencontroller.lyzr.com/blog/foundry-guardrails-for-agent-tool-calls.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
