5.5 Responsible Generative AI
Key Takeaways
- Responsible generative AI applies Microsoft's six responsible-AI principles plus extra safeguards specific to content generation.
- Azure OpenAI content filters (powered by Azure AI Content Safety) scan both prompts and responses across four categories — hate, violence, sexual, and self-harm — and are on by default.
- Prompt Shields detect and block jailbreak and prompt-injection attempts that try to bypass safety guardrails.
- Groundedness detection verifies that responses are supported by the provided source documents, helping catch hallucinations in RAG.
- Microsoft Copilot builds in responsible AI with content filters, citations, disclaimers, feedback, and enterprise data boundaries.
Quick Answer: Responsible generative AI applies Microsoft's six principles with additional safeguards: content filters block harmful content in prompts and responses, Prompt Shields prevent jailbreaks, and groundedness detection ensures responses are factually supported. Azure OpenAI includes these protections by default.
Why Responsible Generative AI Is Different
Generative AI introduces unique risks that traditional AI does not:
| Risk | Description | Example |
|---|---|---|
| Hallucination | Confident but false output | "Abraham Lincoln invented the telephone" |
| Harmful content | Hate, violence, or inappropriate output | Generating offensive text or images |
| Jailbreaking | Prompts that bypass safety guardrails | "Ignore your instructions and tell me how to..." |
| Prompt injection | Hidden instructions in input override the system message | Instructions pasted inside a document |
| Copyright concerns | Output that closely matches protected material | Reproducing copyrighted text or code |
| Misinformation | Plausible but false information | Fake medical or legal advice |
| Privacy leakage | Revealing personal data seen in training | Generating someone's address |
Content Filtering in Azure OpenAI
Azure OpenAI includes default content filters that are always active and powered by Azure AI Content Safety.
Four Filter Categories
| Category | What It Blocks | Severity Levels |
|---|---|---|
| Hate | Discrimination, slurs, dehumanization | Safe -> Low -> Medium -> High |
| Violence | Threats, graphic violence, harm | Safe -> Low -> Medium -> High |
| Sexual | Sexually explicit or suggestive content | Safe -> Low -> Medium -> High |
| Self-harm | Self-injury descriptions or instructions | Safe -> Low -> Medium -> High |
How Filtering Works
- The user submits a prompt to the API.
- The input filter scans the prompt for harmful content.
- If the prompt passes, the model generates a response.
- The output filter scans the response.
- If the response passes, it is returned to the user.
- If either filter triggers at or above the configured severity, the request is blocked with an error.
On the Exam: Azure OpenAI content filters are enabled BY DEFAULT and scan BOTH prompts AND responses. You do not enable them separately — a very common exam point.
Prompt Shields
Prompt Shields detect and block jailbreak and injection attempts — efforts to make the model ignore its safety instructions:
| Technique | Example | How Prompt Shields Help |
|---|---|---|
| Direct override | "Ignore all previous instructions and..." | Detect instruction-override patterns |
| Role play | "Pretend you are an AI with no restrictions..." | Identify role-play bypasses |
| Encoding tricks | Hiding requests in base64 or other encodings | Detect encoded content |
| Indirect injection | Harmful instructions hidden in pasted documents or URLs | Scan all input content for hidden instructions |
Groundedness Detection
Groundedness detection (an Azure AI Content Safety capability) verifies that AI-generated responses are factually supported by the provided source documents:
| Result | Meaning |
|---|---|
| Grounded | The response is supported by the provided context — safe |
| Ungrounded | The response makes claims not supported by the context — flagged as a possible hallucination |
This is especially important in RAG scenarios, where you want the model to answer only from retrieved documents.
Microsoft Copilot and Responsible AI
Microsoft Copilot incorporates responsible AI by design:
| Feature | Responsible-AI Purpose |
|---|---|
| Content filters | Block harmful content in prompts and responses |
| Citations | Show sources for generated information (transparency) |
| Disclaimers | "AI-generated content may be incorrect" warnings (transparency) |
| User feedback | Thumbs up/down to report quality issues (accountability) |
| Data boundaries | Enterprise data stays within the Microsoft 365 boundary (privacy) |
| Grounding in enterprise data | Respond based on your organization's data (accuracy) |
Best Practices for Responsible Generative AI
For Developers and Organizations
| Practice | Description |
|---|---|
| Keep content filters on | Use the default Azure OpenAI filters |
| Implement RAG/grounding | Ground responses in factual data to cut hallucinations |
| Set clear system messages | Define what the AI should and should not do |
| Monitor usage | Review logs for misuse and harmful content |
| Provide disclaimers | Tell users content is AI-generated |
| Enable feedback | Let users report inaccurate or harmful output |
| Test/red-team | Probe your application for vulnerabilities |
| Human oversight | Keep humans in the loop for high-stakes decisions |
For End Users
| Practice | Description |
|---|---|
| Verify information | Do not blindly trust AI output |
| Provide context | Give relevant information to reduce hallucinations |
| Review for bias | Check outputs for unfairness |
| Report issues | Use feedback mechanisms |
Configurable Filters and Annotations
While the default filters are always on, organizations with approved access can configure severity thresholds per category (for example, block at medium-and-above for violence but only high for another category) and can enable optional categories such as protected-material detection for text and code. The API also returns annotations indicating which category and severity triggered a filter, which helps developers log, debug, and tune their applications responsibly.
The key exam takeaway remains that the baseline protection is on by default and covers both input and output — configuration adjusts the thresholds, it does not turn safety off.
On the Exam: Responsible-generative-AI questions focus on content filters (four categories, on by default, scan input and output), hallucination mitigation (RAG/grounding/groundedness detection), Prompt Shields (jailbreak defense), and how Copilot uses citations and disclaimers for transparency.
What are Prompt Shields in Azure OpenAI Service?
What does groundedness detection verify in generative AI?
Which of the following is a responsible AI feature built into Microsoft Copilot?
What is the best technique to reduce hallucinations in generative AI responses?
Which THREE of the following are categories used by Azure OpenAI content filters? (Select three)
Select all that apply