Responsible AI and Content Safety
Foundry apps handle user text, images and tool output, so you must filter harmful input and output and prove the system behaves responsibly. Expect scenario questions on principles, filters and cost.
The six principles
| Principle | Typical clue |
|---|---|
| Fairness | Equal treatment, bias across groups |
| Reliability and safety | Works as intended, handles errors and harmful output |
| Privacy and security | Protect personal data, access control |
| Inclusiveness | Usable by people with different abilities |
| Transparency | Tell users AI is used, explain results |
| Accountability | Humans are responsible, governance and review |
Content filters
Azure OpenAI in Foundry Models applies content filters to both prompts and completions. Categories are hate, sexual, violence and self-harm, each with severity levels (safe, low, medium, high). You choose the threshold per category in a content filter configuration attached to a deployment. A filter set to “high” only blocks high severity content, so lower severities still pass.
Azure AI Content Safety (in the Foundry Control Plane) is also a standalone API. It analyses text and images, and for images you can pass a blob URL. Give Content Safety access to the blob with a managed identity and the Storage Blob Data Reader role (least privilege), not a broad key.
Extra protections
- Prompt Shields: user prompt attacks (jailbreak) and document attacks (indirect prompt injection from retrieved or third-party content).
- Groundedness detection: checks that an answer is supported by the source material.
- Protected material detection: flags copyrighted text or code.
- Custom categories and blocklists: for terms specific to your business.
- Annotate and block modes: annotate returns the result for you to act on, block stops the call.
Cost and performance
Pick models by quality, cost and throughput using the model catalog benchmarks and the model router. Serverless (models as a service) means you do not host weights in your subscription. Watch tokens per minute quotas, and use retry with exponential backoff and jitter on HTTP 429.
Exam traps
- Content filtering is not the same as Prompt Shields. Filters judge harm, shields detect manipulation.
- Setting a filter threshold to high is the loosest setting, not the strictest.
- Transparency is about disclosure. Accountability is about human responsibility.
- Do not fix 429 by creating a new thread or retrying instantly. Use backoff.
- A smaller model for everything saves money but degrades hard questions. The answer is routing.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. Users of a public site are told that an AI system processed their video and text. Which responsible AI principle does this support?
Show answer
Answer: B.Transparency means people understand when and how AI is used and how decisions are made.
2. An agent reads third-party web pages and emails. You want embedded instructions in those documents to be treated as untrusted. What do you enable?
Show answer
Answer: B.Prompt Shields detect user prompt attacks (jailbreaks) and document attacks (indirect injection hidden in content).
3. Costs rose sharply. You need to know whether input size, output size or tool calls are the cause. What do you review?
Show answer
Answer: B.Token usage splits prompt and completion tokens, and tool usage metrics show tool-driven cost.
4. Most requests are simple FAQs, but a few need deep reasoning. How do you cut cost and latency without hurting hard questions?
Show answer
Answer: C.A cascade or model router sends easy prompts to a small model and escalates complex ones.