T2: Responsible AI

You are part of the trust chain. Every infrastructure decision you make — from region selection to content filtering configuration — directly impacts whether your AI system is safe, fair, and trustworthy. Responsible AI isn't a checkbox; it's a design discipline woven into every layer. For content safety implementation patterns, see T3: Production Patterns. For grounding and accuracy, see R3: Deterministic AI.

Microsoft's 6 Responsible AI Principles

Principle	What It Means	Your Responsibility
Fairness	AI treats all people equitably	Test across demographics, monitor for bias
Reliability & Safety	AI performs as intended	Retry logic, fallbacks, circuit breakers
Privacy & Security	AI protects data and access	Managed Identity, Key Vault, RBAC, encryption
Inclusiveness	AI is accessible to everyone	Multi-language, accessibility, diverse testing
Transparency	People understand how AI works	Source citations, confidence scores, AI labels
Accountability	People are accountable for AI	Audit logs, human-in-the-loop, incident response

Infrastructure Decisions That Impact Safety

Every "infrastructure" choice is actually a safety decision:

Decision	Safety Impact	Recommendation
Region selection	Data residency, compliance	Match to user geography + regulations
Content filtering	Blocks harmful outputs	Enable on ALL endpoints — never disable
Logging strategy	Audit trail for incidents	Log all AI interactions (without PII)
Rate limiting	Prevents abuse and cost explosion	Per-user + per-tenant limits
Key management	Prevents unauthorized access	Key Vault + Managed Identity, never hardcode
RBAC	Least-privilege access	Separate roles for dev/deploy/admin
Private endpoints	Network isolation	Required for production PaaS services
Model selection	Capability vs risk tradeoff	Smaller models for narrow tasks (less hallucination)

Azure AI Content Safety

Azure AI Content Safety provides real-time detection across four harm categories:

User Input ──▶ [Input Filter] ──▶ [Model] ──▶ [Output Filter] ──▶ Response
                    │                              │
                    ▼                              ▼
              Block/Flag                    Block/Flag
              if severity ≥ threshold       if severity ≥ threshold

Category	Severity Scale	Default Block	Description
Hate	0-6	≥ 2	Discrimination, slurs, dehumanization
Self-Harm	0-6	≥ 2	Instructions or encouragement of self-harm
Sexual	0-6	≥ 2	Explicit sexual content
Violence	0-6	≥ 2	Graphic violence, weapons instructions

Additional protections:

Prompt Shields — detect jailbreak and indirect prompt injection attempts
Groundedness detection — flag ungrounded claims in model outputs
Protected material detection — identify copyrighted text in outputs

:::info Content Safety Implementation Configure content filtering in your guardrails.json:

{
  "content_safety": {
    "hate": { "threshold": 2, "action": "block" },
    "self_harm": { "threshold": 2, "action": "block" },
    "sexual": { "threshold": 2, "action": "block" },
    "violence": { "threshold": 2, "action": "block" }
  },
  "prompt_shields": { "enabled": true },
  "groundedness": { "enabled": true, "threshold": 4.0 }
}

:::

OWASP LLM Top 10 Risks

The OWASP Top 10 for LLM Applications identifies the most critical security risks:

#	Risk	Mitigation
1	Prompt Injection	Input validation, Prompt Shields, system prompt isolation
2	Insecure Output Handling	Sanitize AI output before rendering, never exec AI output
3	Training Data Poisoning	Curate data sources, validate training sets
4	Model Denial of Service	Rate limiting, token budgets, timeout enforcement
5	Supply Chain Vulnerabilities	Pin model versions, audit dependencies
6	Sensitive Information Disclosure	PII detection, output filtering, data minimization
7	Insecure Plugin Design	Least-privilege tool access, input validation
8	Excessive Agency	Human-in-the-loop for critical actions, action confirmation
9	Overreliance	Confidence scores, source citations, user education
10	Model Theft	Private endpoints, access controls, monitoring

EU AI Act Overview

:::warning EU AI Act — Know Your Risk Classification The EU AI Act entered into force in August 2024 with phased enforcement. If your AI system operates in the EU or serves EU users, you must classify it. High-risk systems face mandatory conformity assessments, transparency obligations, and human oversight requirements. Non-compliance penalties reach up to €35M or 7% of global turnover. :::

Risk Level	Examples	Requirements
Unacceptable	Social scoring, real-time biometric surveillance	Banned
High-Risk	Hiring, credit scoring, medical diagnosis, law enforcement	Conformity assessment, logging, human oversight
Limited Risk	Chatbots, deepfake generation	Transparency obligations (label as AI)
Minimal Risk	Spam filters, game AI	No specific requirements

For most enterprise AI applications (RAG chatbots, document processing, IT assistants), you fall under limited risk — requiring transparency labels. If your system influences decisions about people (hiring, lending, medical), it's likely high-risk.

Content Safety Pipeline

A production content safety pipeline has four stages:

1. INPUT FILTERING          2. MODEL GENERATION
   ├─ Prompt Shields           ├─ Content filter (built-in)
   ├─ PII detection            ├─ Token budget enforcement
   ├─ Input sanitization       └─ System prompt guardrails
   └─ Rate limiting
   
3. OUTPUT FILTERING         4. LOGGING & MONITORING
   ├─ Content Safety API       ├─ Log interaction (no PII)
   ├─ Groundedness check       ├─ Correlation ID tracking
   ├─ Citation verification    ├─ Alert on blocked content
   └─ PII redaction            └─ Audit trail retention

Evaluation for Trust

Responsible AI requires continuous evaluation, not one-time checks:

Metric	Target	What It Measures
Groundedness	≥ 4.0 / 5.0	Are claims supported by provided context?
Relevance	≥ 4.0 / 5.0	Does the response address the question?
Coherence	≥ 4.0 / 5.0	Is the response logically consistent?
Safety	0 violations	Are harmful content filters effective?
Fairness	< 5% variance	Do responses vary by demographic?

# Evaluation pipeline example
from azure.ai.evaluation import GroundednessEvaluator, ContentSafetyEvaluator

groundedness = GroundednessEvaluator(model_config)
safety = ContentSafetyEvaluator(credential, azure_ai_project)

result = groundedness(
    response="The contract requires 30-day payment terms.",
    context="Section 4.2: Payment shall be made within 30 days...",
    query="What are the payment terms?"
)
assert result["groundedness"] >= 4.0

Key Takeaways

You are the trust chain — infrastructure choices are safety choices
Enable content filtering everywhere — never disable, even in dev
Know your OWASP LLM risks — prompt injection is #1 for a reason
Classify under EU AI Act — know your obligations before deployment
Evaluate continuously — groundedness ≥ 4.0, zero safety violations

Next: T3: Production Patterns — taking AI from prototype to production with resilience, cost control, and monitoring.

Microsoft's 6 Responsible AI Principles​

Infrastructure Decisions That Impact Safety​

Azure AI Content Safety​

OWASP LLM Top 10 Risks​

EU AI Act Overview​

Content Safety Pipeline​

Evaluation for Trust​

Key Takeaways​