Bài 10: AWS Responsible AI Tools — Clarify, A2I & Guardrails

1. Amazon SageMaker Clarify

SageMaker Clarify helps detect bias in data and models, and provides model explainability — the go-to AWS service for Responsible AI.

1.1. Three Core Capabilities

Capability	When	What it does
Pre-training bias detection	Before training	Detects imbalances in training data across demographic groups
Post-training bias detection	After training	Detects bias in model predictions (e.g., different accuracy across groups)
Explainability (SHAP)	After training	Shows feature contributions to each prediction

1.2. Key Bias Metrics in Clarify

Metric	Pre/Post training	What it measures
Class Imbalance (CI)	Pre-training	Distribution of classes across groups
Difference in Proportions (DPL)	Pre-training	Proportion of positive labels across groups
KL Divergence	Pre-training	Distribution divergence between groups
Disparate Impact (DI)	Post-training	Ratio of positive predictions across groups
Accuracy Difference (AD)	Post-training	Accuracy gap between groups
Treatment Equality (TE)	Post-training	Ratio of FP to FN across groups

1.3. Clarify Workflow

1. Configure Clarify Job
   ├── Specify sensitive attributes (gender, age, race)
   ├── Define facets (groups to compare)
   └── Choose bias metrics to compute

2. Run Pre-training Analysis
   ├── Upload training dataset
   └── Get bias report on data distribution

3. Train Model

4. Run Post-training Analysis
   ├── Compare predictions across groups
   └── Get SHAP values for explainability

5. Monitor with SageMaker Model Monitor
   └── Detect bias drift over time in production

Exam tip: "Which AWS service can detect if a model makes more errors for one racial group vs another?" → SageMaker Clarify (post-training bias detection).

2. Amazon Augmented AI (Amazon A2I)

Amazon A2I cung cấp human-in-the-loop (HITL) workflows cho AI predictions — đặc biệt quan trọng khi model confidence thấp hoặc high-stakes decisions.

2.1. How A2I Works

AI Prediction Flow with A2I:
                                          ┌─────────────┐
User Request → AI Model → Confident?  YES → Return result
                              │
                              NO (below threshold)
                              ↓
                     ┌──────────────────┐
                     │  Create Human    │
                     │  Review Task     │
                     │  (A2I Workflow)   │
                     └────────┬─────────┘
                              ↓
                     ┌──────────────────┐
                     │  Human Reviewer  │  ← AWS Mechanical Turk
                     │  Reviews &       │  ← Private workforce
                     │  Corrects        │  ← Third-party vendor
                     └────────┬─────────┘
                              ↓
                     Return human-verified result

2.2. A2I Components

Component	Purpose
Human review workflow	Defines when and how to trigger human review
Worker task template	UI for human reviewers to make decisions
Workforce	Who does the review (private, Mechanical Turk, vendor)
Activation conditions	Confidence threshold triggers (e.g., < 95%)

2.3. Built-in A2I Integrations

Service	A2I Use Case
Amazon Textract	Review low-confidence document extractions
Amazon Rekognition	Review low-confidence content moderation
Custom ML models	Any SageMaker model can trigger A2I

Exam tip: "A healthcare company needs a human to review AI diagnoses when the model is less than 90% confident" → Amazon A2I with activation condition set to confidence < 90%.

3. Amazon Bedrock Guardrails — Deep Dive

3.1. Guardrail Policies

Policy	How it works	Configuration
Content filters	Block by category + severity	None / Low / Medium / High for each category (Hate, Insults, Sexual, Violence, Misconduct)
Denied topics	Define topics to block	Natural language description + sample phrases
Word filters	Block specific words	Custom word/phrase list + profanity filter toggle
Sensitive info (PII)	Detect PII with action	BLOCK or ANONYMIZE for each PII type (SSN, email, phone, name, address, ...)
Contextual grounding	Check if answer is grounded	Grounding threshold (0-1) + relevance threshold

3.2. How Guardrails Process Requests

User Input
    ↓
[INPUT GUARDRAILS]
    ├── Content filter check
    ├── Denied topic check
    ├── Word filter check
    ├── PII detection → BLOCK or ANONYMIZE
    ↓ (if passes all checks)
Foundation Model generates response
    ↓
[OUTPUT GUARDRAILS]
    ├── Content filter check
    ├── Denied topic check
    ├── Word filter check
    ├── PII detection → BLOCK or ANONYMIZE
    ├── Contextual grounding check
    ↓ (if passes all checks)
Response returned to user

If blocked → Return configured "blocked" message

3.3. Guardrails vs System Prompts

Aspect	System Prompt	Guardrails
Enforcement	Soft — model may ignore	Hard — enforced by the platform
Bypass risk	Can be bypassed via prompt injection	Cannot be bypassed by prompts
PII handling	Model asked to not output PII	Programmatic detection & redaction
Auditability	Limited	Full logging and metrics

Exam tip: "A company needs to GUARANTEE that PII is never in model responses" → Bedrock Guardrails (not system prompts, which can be bypassed).

4. Content Moderation on AWS

Service	Content Type	Use Case
Amazon Rekognition	Images & Video	Detect inappropriate content, faces, text
Amazon Comprehend	Text	Toxicity detection, sentiment analysis
Bedrock Guardrails	FM input/output	Filter harmful content in GenAI apps
Amazon A2I	Any	Human review for edge cases

5. AI Governance

5.1. Governance Framework

Area	What to implement
Policy	Organization-wide AI ethics guidelines
Risk assessment	Evaluate risks before deploying AI systems
Monitoring	Continuous monitoring for bias, performance drift
Audit trail	Log all model decisions for accountability
Human oversight	Human-in-the-loop for high-stakes decisions
Documentation	Model cards, AI Service Cards

5.2. SageMaker ML Governance

• SageMaker Model Cards: Document model details and intended use
• SageMaker Model Dashboard: Centralized view of all model status
• SageMaker Model Monitor: Detect data drift, model quality degradation
• SageMaker Role Manager: Fine-grained access control for ML

6. Practice Questions

Q1: An insurance company wants to ensure their claim approval model treats customers of all ages fairly. Which AWS service should they use to detect age-based bias in the model's predictions?

• A) Amazon Rekognition
• B) Amazon SageMaker Clarify ✓
• C) Amazon Bedrock Guardrails
• D) Amazon Comprehend

Explanation: SageMaker Clarify can run post-training bias analysis comparing model predictions across age groups, using metrics like Disparate Impact and Accuracy Difference.

Q2: A document processing application using Amazon Textract needs human review when the extracted data has low confidence. Which AWS service provides this capability?

• A) Amazon SageMaker Ground Truth
• B) Amazon Augmented AI (A2I) ✓
• C) Amazon Mechanical Turk directly
• D) Amazon Bedrock Agents

Explanation: Amazon A2I has built-in integration with Amazon Textract and can automatically trigger human review workflows when extraction confidence falls below a defined threshold.

Q3: A chatbot must NEVER reveal customer credit card numbers in its responses, even if the data exists in the knowledge base. Which approach provides the STRONGEST guarantee?

• A) Add "never output credit card numbers" to the system prompt
• B) Fine-tune the model to not output PII
• C) Use Amazon Bedrock Guardrails with PII filters set to BLOCK ✓
• D) Remove credit card numbers from the knowledge base

Explanation: Bedrock Guardrails with PII filters provide programmatic detection and blocking of credit card numbers in both input and output — this cannot be bypassed by prompt injection, unlike system prompts.