Module 5: Security, Ethics & Future
AI Act & AI Agents

1. Risk Category: Limited Risk → it's a sentiment analysis system influencing a service (customer support).

2. Obligations: Transparency (inform users), right to explanation, possibility of human appeal.

3. Redesign:

Revised System:
1. The AI doesn't assign priority, but suggests a classification.
2. Every email classified as "angry" is automatically sent to a human operator for validation.
3. The prompt includes: "You are a sentiment analyzer. Your output is a SUGGESTION. Classify as: SUGGESTED_Satisfied, SUGGESTED_Neutral, SUGGESTED_Angry."
4. Users are shown: "Our system has suggested this classification to best handle your request."

1. System Prompt:

You are a contract review assistant. Your role is to RAISE POINTS OF ATTENTION and SUGGEST AREAS FOR ANALYSIS based on common patterns.
**FUNDAMENTAL PRINCIPLES:**
- DO NOT provide definitive legal interpretations.
- DO NOT make decisions.
- Final responsibility is always with the reviewing lawyer.
- For each suggestion, indicate the basis (e.g., "common atypical clause", "potential ambiguity").
- If uncertain, suggest "consulting specific case law".

2. Output Format:

**DOCUMENT: [Contract Name]**
**AI ANALYSIS (Suggestions):**
1. [Point of attention 1] - Basis: [ ] - Estimated Risk: Low/Medium/High
2. [Point of attention 2] - Basis: [ ] - Estimated Risk: Low/Medium/High
**HUMAN REVIEWER SECTION:**
☐ AI analysis verified
☐ Final decision on point 1: [Accept/Modify/Delete]
☐ Final decision on point 2: [Accept/Modify/Delete]
**Lawyer's Notes:** [_________________]

3. Data to Log: Lawyer ID, timestamp, document name, points raised by AI, lawyer's final decisions, review time.

4-LAYER HALLUCINATION MITIGATION SYSTEM
========================================

LAYER 1: PROMPT ENGINEERING
- System prompt: "You are a legal assistant. ALWAYS cite specific source.
  If unsure → 'Insufficient information available'.
  Confidence scoring: high/medium/low for each statement."
- Few-shot examples with correct citations
- Mandatory output formatting

LAYER 2: RAG ARCHITECTURE
- Vector database: 10k+ judgments (ChromaDB)
- Retrieval: top 5 most relevant documents
- Citation enforcement: each statement → source
- Fallback: "This information is not in the database"

LAYER 3: VALIDATION PIPELINE
- Auto-validation: model verifies its own statements
- Cross-check: second model validates first model's output
- Fact-checking: regex for dates, names, references
- Confidence threshold: only output >80% confidence

LAYER 4: HUMAN OVERSIGHT
- Flag system for high-risk statements
- Mandatory attorney review for strategic advice
- Audit trail: who approved what and when
- Continuous feedback loop corrections → training

MONITORING METRICS:
- Hallucination rate: target <2%
- Citation accuracy: target >95%
- Attorney satisfaction: target >8/10
- Response time P95: target <10s

INCIDENT RESPONSE:
1. Immediate rollback to last known good version
2. Root cause analysis (prompt, data, model)
3. Correction implementation
4. Retesting and validation
5. Communication to affected users
6. Prevention measures update

AUTONOMOUS CONTENT MARKETING AGENT
===================================

AGENT GOAL: Increase engagement by 20% while maintaining brand voice and compliance.

CONSTRAINTS:
- Never post without human approval for sensitive topics
- Budget limit: €500/month for promoted content
- Daily time limit: 2 hours autonomous operation
- Content must be fact-checked before publishing

PROMPT ARCHITECTURE:
System: "You are an autonomous content marketing agent with these tools:
1. web_search(topic): Get trending information
2. analyze_engagement(data): Calculate metrics
3. generate_content(brief): Create posts
4. schedule_post(content, time): Plan publishing

Daily workflow:
1. Search trending topics in [industry]
2. Generate 5 content ideas
3. Create 2 posts for today
4. Analyze yesterday's engagement
5. Adjust strategy based on data"

SAFETY MECHANISMS:
- Human approval required for: political content, sensitive topics, >€100 spend
- Automatic fact-checking via trusted sources
- Sentiment analysis to avoid negative brand association
- Daily activity log for review

SUCCESS METRICS:
- Engagement rate increase
- Follower growth
- Brand sentiment score
- Cost per engagement

FAILURE SCENARIOS:
1. Unexpected trend (real-time alert to human)
2. API failure (fallback to manual mode)
3. Negative engagement spike (automatic pause)
4. Budget exceeded (automatic shutdown)

MONITORING:
- Real-time dashboard with key metrics
- Weekly automated report
- Monthly human performance review