2nd/10_Wiki/Topics/AI_and_ML/AI 거버넌스 정책(AI Usage Policy).md

---
id: wiki-2026-0508-ai-거버넌스-정책-ai-usage-policy
title: AI Governance Policy (AI Usage Policy)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI Usage Policy, AI 거버넌스 정책, AI policy framework, EU AI Act, NIST AI RMF, ISO 42001]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: conceptual
tags: [ai-governance, policy, compliance, risk-management, eu-ai-act, nist-rmf, iso-42001, internal-policy]
raw_sources: []
last_reinforced: 2026-05-09
github_commit: pending
inferred_by: Claude Opus 4.7 (manual cleanup 2026-05-09)
tech_stack:
  language: process / policy
  applicable_to: [Compliance, Engineering, HR, Legal]
---

# AI Governance Policy (AI Usage Policy)

## 📌 한 줄 통찰 (The Karpathy Summary)
> **"자율 = 책임"**. 조직 의 AI 도입 의 legal / ethical / security 의 framework. 규제 (EU AI Act) + 자체 policy + technical guardrail. **금지 X, sandbox + 교육 + accountability**.

## 📖 구조화된 지식 (Synthesized Content)

### 핵심 axis
1. **Acceptable Use**: 매 employee 의 AI 도구 사용 의 boundary.
2. **Data / IP Protection**: 매 prompt 의 sensitive data 의 prevention.
3. **Human-in-the-loop**: 매 critical decision 의 human review.
4. **Accountability**: 매 AI-caused harm 의 legal / financial owner.
5. **Transparency**: 매 user 의 AI 사용 의 disclosure.
6. **Bias / Fairness**: 매 group 의 differential treatment 의 audit.
7. **Compliance**: 매 regulation 의 mapping (EU AI Act, GDPR, ...).

### 주요 regulation (2024-2026)
| Regulation | Region | Key |
|---|---|---|
| **EU AI Act** | EU | Risk-based (4 tier). High-risk = strict (2026 enforcement). |
| **NIST AI RMF** | US | Voluntary framework. 4 function: Govern/Map/Measure/Manage. |
| **ISO 42001** | Global | 매 org 의 AI management standard (cert 가능). |
| **US EO 14110** | US | Federal AI guidance. |
| **China AI Reg** | China | Generative AI 의 strict (2023+). |
| **UK AI White Paper** | UK | Pro-innovation, sector-specific. |
| **Korea AI Act** | KR | 2025 enforcement scheduled. |

### EU AI Act 의 risk tier
1. **Unacceptable**: social scoring, manipulation, biometric mass surveillance → ban.
2. **High-risk**: HR, education, law enforcement, critical infra → strict (audit, doc, human oversight).
3. **Limited risk**: chatbot, deepfake → transparency.
4. **Minimal**: 매 spam filter → no requirement.

→ "내 AI use case 의 tier" 의 매 org 의 분류.

### Internal policy 의 structure
1. **Scope & Definitions**: 매 "AI" 의 정의.
2. **Approved tools**: ChatGPT (Enterprise), Claude (Pro), GitHub Copilot, Cursor, internal LLM, ...
3. **Prohibited tools**: free ChatGPT (data leak), unverified plugin, ...
4. **Acceptable use**: brainstorm, draft, code assist OK. Customer data 의 input X.
5. **Prohibited use**: 매 sensitive data, deepfake, automated hire decision (without review).
6. **Data classification**: public, internal, confidential, restricted.
7. **Approval workflow**: 매 new tool 의 IT + legal + security review.
8. **Training requirement**: 매 employee 의 annual AI literacy.
9. **Incident response**: 매 misuse 의 reporting + escalation.
10. **Audit**: 매 quarter / year 의 review.

### Common 항목 detail

#### Data classification
- **Public**: marketing copy → 매 AI tool OK.
- **Internal**: project plan → enterprise AI 만 (data not training).
- **Confidential**: customer data, financial → strict approval만.
- **Restricted**: PHI, PII, source code (proprietary) → 매 cloud AI X.

#### Human-in-the-loop
- **High-risk decision** (hire, fire, loan, medical): 매 AI 의 recommend, human 의 final.
- **Medium-risk** (content publish): 매 review of AI output.
- **Low-risk** (spam classification): automated OK.

#### Audit log
- 매 AI tool call 의 user, timestamp, prompt summary, output summary.
- Sensitive data 의 detection.
- Anomaly (가장 큰 query, off-hours).

→ Compliance 의 evidence.

### 매 industry 의 specific
- **Healthcare** (HIPAA, FDA): 매 medical AI 의 separate.
- **Finance** (SOC 2, FFIEC): bias audit, explainability.
- **Legal**: privilege protection, billing (AI-assisted = client disclosure).
- **Education**: student data (FERPA), academic integrity.
- **Government**: classified info, FOIA implications.

### Sandbox approach
**Bad**: "Ban all AI" → shadow IT + competitive disadvantage.
**Good**: 매 employee 의 controlled experimentation:
- 매 approved tool list.
- 매 use case 의 review 후 OK.
- Internal LLM (privacy 친화).
- Quarterly review of new tools.

### Vendor management
- 매 AI vendor 의 DPA (Data Processing Agreement).
- Training data clause: "내 data 가 train X".
- Sub-processor list.
- Geographic data location.
- Termination + data deletion.
- Liability.

→ 매 procurement team 의 책임.

### IP / 저작권 의 분야
- **AI-generated content 의 ownership**: 매 country 가 다름 (US 가 human authorship 만).
- **Training data 의 license**: copyright lawsuit 진행 중 (NYT vs OpenAI).
- **Code generation**: license 의 contamination (GitHub Copilot lawsuit).
- **매 AI output 의 originality**: 매 user 가 copyright?

→ 매 case 의 legal 전문가.

### Bias / Fairness audit
- 매 sensitive attribute (gender, race, age) 의 differential outcome.
- Statistical parity / equal opportunity / calibration.
- Counterfactual fairness.
- 매 release 전 의 audit.

```python
# Simple bias check
from aif360.metrics import BinaryLabelDatasetMetric

metric = BinaryLabelDatasetMetric(
    dataset, privileged_groups=[{'gender': 1}], unprivileged_groups=[{'gender': 0}]
)
print(metric.disparate_impact())  # < 0.8 = potential bias
```

### Disclosure / labeling
- AI-generated content 의 명시 (EU AI Act).
- Chatbot 의 disclosure.
- Deepfake watermark.
- Customer-facing AI 의 "이거 AI" notice.

### Incident response
1. **Detection**: monitoring alert / user report.
2. **Containment**: tool 의 disable.
3. **Investigation**: 매 misuse 의 root cause.
4. **Remediation**: data deletion, user notification.
5. **Lesson learned**: policy update + training.

## 💻 패턴 (policy implementation)

### Policy template (markdown)
```markdown
# AI Usage Policy v1.0

## Scope
This policy applies to all employees, contractors, and partners using AI tools for company work.

## Definitions
- AI tool: any system using ML / LLM (ChatGPT, Claude, Copilot, ...).
- Sensitive data: customer PII, financial, source code.

## Approved Tools
- ChatGPT Enterprise (data not used for training).
- Claude (Pro / Team).
- GitHub Copilot Business.
- Cursor (with privacy mode).

## Acceptable Use
- Drafting, brainstorming, code assistance.
- Research and summarization.
- Translation.

## Prohibited Use
- Inputting customer PII or financial data.
- Generating fake content for deception.
- Automated decisions affecting employees (hire/fire).

## Data Classification
- See [data classification guide](#).

## Human Oversight
- Critical decisions: human final review.
- Customer-facing content: human approval.

## Reporting
- Misuse: report to ai-policy@company.
- Incidents: privacy@company within 24h.

## Training
- Annual AI literacy training (mandatory).
- New hire onboarding (within first month).

## Review
- This policy reviewed quarterly by AI Council.
- Last updated: 2026-05-09.
```

### DLP (data loss prevention) check
```python
import re

SENSITIVE_PATTERNS = [
    r'\b\d{3}-\d{2}-\d{4}\b',  # SSN
    r'\b4\d{12,15}\b',           # credit card (Visa)
    r'(?i)password\s*[:=]\s*\S+',
    r'(?i)api[_-]?key\s*[:=]\s*\S+',
]

def check_prompt(prompt: str):
    for pattern in SENSITIVE_PATTERNS:
        if re.search(pattern, prompt):
            block_and_alert(prompt, pattern)
            return False
    return True
```

→ Sensitive data 의 prompt 차단.

### Audit log
```ts
async function auditAICall(user: User, tool: string, prompt: string, response: string) {
  await db.aiAuditLog.insert({
    userId: user.id,
    tool,
    promptHash: sha256(prompt),
    promptLength: prompt.length,
    responseHash: sha256(response),
    timestamp: new Date(),
    classification: classifySensitivity(prompt),
  });
}
```

→ 매 call 의 hashed log (prompt 의 raw 가 storage X for privacy).

### Approval workflow
```yaml
# .github/CODEOWNERS or similar
# 매 new AI tool integration 의 review

ai_integrations/* @ai-council @security-team @legal
*.policy.md @ai-council
```

### Monitoring (anomaly)
```sql
-- 매 user 의 unusual AI usage
SELECT user_id, COUNT(*) AS calls, SUM(prompt_length) AS chars
FROM ai_audit_log
WHERE created_at > NOW() - INTERVAL '7 day'
GROUP BY user_id
HAVING COUNT(*) > 1000   -- threshold
ORDER BY chars DESC;
```

### Bias audit (CI)
```python
# Per-release bias check
def audit_bias(model, test_set):
    results_by_group = defaultdict(list)
    for x, y_true, group in test_set:
        y_pred = model.predict(x)
        results_by_group[group].append((y_true, y_pred))
    
    for group, results in results_by_group.items():
        accuracy = compute_accuracy(results)
        false_positive = compute_fpr(results)
        log({'group': group, 'accuracy': accuracy, 'fpr': false_positive})
    
    # Fail if disparity > threshold
    accuracies = [compute_accuracy(r) for r in results_by_group.values()]
    if max(accuracies) - min(accuracies) > 0.05:
        raise BiasViolation()
```

## 🤔 의사결정 기준 (Decision Criteria)

| 상황 | 정책 |
|---|---|
| Low-risk (spam filter) | Minimal policy + audit |
| Medium-risk (content moderation) | Human review + transparency |
| High-risk (HR, medical, finance) | Strict approval + audit + bias check |
| Public-facing AI | EU AI Act compliance + disclosure |
| Internal tool | Data classification + DLP |
| Vendor AI | DPA + sub-processor review |
| New tool 의 introduction | AI council review 의 30 day |

**기본값**: Sandbox + transparency + human-in-the-loop. "Ban all" / "allow all" 가 X.

## ⚠️ 모순 및 업데이트 (Contradictions & Updates)
- **Innovation vs control**: 너무 strict = shadow IT (employee 가 personal account 사용). 너무 loose = data leak.
- **EU AI Act 의 ambiguity**: 매 tier 의 boundary 가 case-by-case.
- **Multi-jurisdiction**: 매 country 의 다른 regulation. 매 employee location 의 issue.
- **Speed of change**: regulation 의 yearly update. Policy 의 quarterly review.
- **Vendor 의 data assurance**: "data not used for training" claim 의 verification 어려움.

## 🔗 지식 연결 (Graph)
- 부모: [[AI-Ethics]] · [[Risk_Management|Risk-Management]]
- 응용: [[NIST-AI-RMF]] · [[ISO-42001]]
- 기술: [[Model-Card]]
- 응용: [[AI Literacy]] · [[AI Accountability]]

## 🤖 LLM 활용 힌트 (How to Use This Knowledge)

**언제 이 지식을 쓰는가:**
- 회사 의 AI policy 의 첫 draft.
- 매 vendor 의 DPA review.
- AI tool 의 approval workflow design.
- Compliance audit 의 prep (EU AI Act, ISO 42001).
- AI incident 의 response.
- Employee training 의 design.

**언제 쓰면 안 되는가:**
- Specific legal advice (lawyer).
- Country-specific regulation 의 implementation (local counsel).
- Technical implementation 의 detail (engineer).
- Crisis 의 immediate response (incident response team).

## ❌ 안티패턴 (Anti-Patterns)
- **"All AI banned"**: shadow IT.
- **"All AI allowed"**: data breach.
- **No data classification**: 매 sensitive 의 leak.
- **No vendor DPA**: liability vacuum.
- **No audit log**: compliance fail.
- **No incident response**: crisis 의 amplify.
- **No regular review**: regulation 의 outdated.
- **One-size-fits-all**: 매 industry / role 의 different need.

## 🧪 검증 상태 (Validation)
- **정보 상태:** verified (concept-level).
- **출처 신뢰도:** B (NIST AI RMF, EU AI Act 공식, ISO 42001 published).
- **검토 이유:** Manual cleanup. Regulation 의 active update. 매 6 month review.

## 🧬 중복 검사 (Duplicate Check)
- **기존 유사 문서:** [[AI-Ethics]] (parent), [[AI Safety]] (related), [[AI Accountability]] (subset).
- **처리 방식:** KEEP (organizational governance focus).
- **처리 이유:** Policy 가 distinct discipline (ethics + compliance + ops).

## 🕓 변경 이력 (Changelog)
| 날짜 | 변경 내용 | 처리 방식 | 신뢰도 |
|------|-----------|-----------|--------|
| 2026-05-08 | P-Reinforce Phase 1 정규화 | UPDATE | A |
| 2026-05-09 | Manual cleanup — code pattern + regulation map + industry specific + 안티패턴 추가 | UPDATE | B |