---
id: wiki-2026-0508-ethics-ai
title: Ethics & AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI ethics, responsible AI, AI safety, alignment, fairness, bias, EU AI Act]
duplicate_of: none
source_trust_level: A
confidence_score: 0.97
verification_status: applied
tags: [ethics, ai-ethics, alignment, safety, bias, fairness, responsibility, eu-ai-act]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Universal
  applicable_to: [AI Development, Policy, Governance]
---

# Ethics & AI

## 매 한 줄
> **"매 AI 의 design / deploy / govern 의 normative consideration"**. 매 fairness, accountability, transparency, safety. 매 modern: EU AI Act, NIST AI RMF, Anthropic Constitutional AI. 매 alignment + capability + governance 의 triad.

## 매 핵심

### 매 pillar
- **Fairness**: 매 bias 의 mitigate.
- **Accountability**: 매 who 의 responsible.
- **Transparency / Explainability**.
- **Privacy**.
- **Safety**: 매 harm prevention.
- **Robustness**.
- **Human autonomy**.

### 매 alignment
- **RLHF**: 매 human preference.
- **Constitutional AI** (Anthropic): 매 principle-based.
- **DPO / KTO**: 매 RLHF alternative.
- **Scalable oversight**: 매 debate, IDA.
- **Honest / harmless / helpful** (HHH).

### 매 framework
- **EU AI Act** (2024): 매 risk-tier.
- **NIST AI RMF**: 매 govern, map, measure, manage.
- **OECD AI Principles**.
- **ISO/IEC 42001**: 매 AIMS.
- **GDPR** (privacy).
- **Algorithmic Accountability Act** (US, proposed).

### 매 risk-tier (EU AI Act)
- **Unacceptable**: 매 social scoring, mass biometric.
- **High-risk**: 매 hiring, credit, education, AV.
- **Limited risk**: 매 chatbot disclose.
- **Minimal**: 매 spam filter.

### 매 응용 issue
1. **Hiring**: 매 disparate impact.
2. **Credit**: 매 redlining.
3. **Healthcare**: 매 race-based prediction.
4. **Justice**: 매 COMPAS bias.
5. **Generative**: 매 deepfake, copyright.
6. **Surveillance**: 매 mass.
7. **Autonomous**: 매 trolley.

### 매 modern (2024-2026)
- **Anthropic RSP** (Responsible Scaling Policy).
- **OpenAI Preparedness**.
- **Frontier model evaluations** (METR, Apollo).
- **AI safety institute** (UK, US).
- **AI Bill of Rights** (US OSTP).

## 💻 패턴

### Fairness audit (demographic parity)
```python
import numpy as np

def demographic_parity_diff(predictions, protected_attr):
    groups = np.unique(protected_attr)
    rates = [predictions[protected_attr == g].mean() for g in groups]
    return max(rates) - min(rates)

# 매 < 0.05 = 80% rule heuristic compliant
```

### Equalized odds
```python
def equalized_odds(predictions, labels, protected):
    """매 TPR + FPR 의 group 에 의 equal."""
    groups = np.unique(protected)
    metrics = {}
    for g in groups:
        mask = protected == g
        tpr = ((predictions == 1) & (labels == 1) & mask).sum() / max(1, ((labels == 1) & mask).sum())
        fpr = ((predictions == 1) & (labels == 0) & mask).sum() / max(1, ((labels == 0) & mask).sum())
        metrics[g] = (tpr, fpr)
    return metrics
```

### Bias mitigation (reweighing)
```python
def reweighing(X, y, protected):
    """매 Kamiran-Calders 2012."""
    weights = np.ones(len(y))
    for g in np.unique(protected):
        for c in [0, 1]:
            mask = (protected == g) & (y == c)
            p_expected = (protected == g).mean() * (y == c).mean()
            p_observed = mask.mean()
            weights[mask] = p_expected / max(p_observed, 1e-9)
    return weights
```

### Constitutional AI (principle-based)
```python
def cai_critique(response, principles):
    prompt = f"""Critique this response against these principles.
Principles:
{format_principles(principles)}

Response: {response}

Output JSON with:
- violated: list of principle IDs
- explanation
- revised_response"""
    return json.loads(llm.generate(prompt))
```

### Differential privacy (DP-SGD)
```python
import opacus
from opacus import PrivacyEngine

privacy_engine = PrivacyEngine()
model, optim, loader = privacy_engine.make_private(
    module=model,
    optimizer=optim,
    data_loader=loader,
    noise_multiplier=1.1,
    max_grad_norm=1.0,
)
```

### Model card
```yaml
model_name: credit-scoring-v3
intended_use: Adult US credit applications, $1k-$50k unsecured
out_of_scope:
  - Outside US
  - Under 18
  - Loans > $50k
training_data:
  source: 2018-2024 internal
  size: 2.4M
  protected_attribute_audit: completed
  fairness_metrics:
    demographic_parity_diff: 0.034
    equalized_odds_diff: 0.041
limitations:
  - Decreased performance on thin-file applicants
  - Quarterly retraining required
```

### Provenance (C2PA, watermark)
```python
from c2pa import Signer
def attach_provenance(image_path, signer_cert):
    Signer(signer_cert).sign(image_path, claims={
        'generator': 'Anthropic Claude Opus 4.7',
        'timestamp': now(),
        'training_data_redacted': True,
    })
```

### Red-teaming
```python
def adversarial_eval(model, attack_categories):
    attacks = []
    for cat in attack_categories:  # 매 jailbreak, bias, harmful, hallucination
        prompts = generate_attacks(cat, n=100)
        for p in prompts:
            r = model.generate(p)
            score = judge(r, cat)
            attacks.append({'cat': cat, 'prompt': p, 'response': r, 'severity': score})
    return attacks
```

### Risk tier classifier (EU AI Act)
```python
def eu_risk_tier(use_case):
    if use_case in {'social_scoring', 'real_time_remote_biometric'}:
        return 'unacceptable'
    if use_case in {'hiring', 'credit', 'education', 'critical_infra', 'law_enforcement'}:
        return 'high'
    if use_case in {'chatbot', 'deepfake', 'emotion_recognition_workplace'}:
        return 'limited'
    return 'minimal'
```

### Consent (GDPR)
```python
def can_process(user, purpose):
    if user.consent[purpose].is_valid():
        return True
    if has_legitimate_interest(purpose):
        return True
    return False

def revoke_consent(user, purpose):
    user.consent[purpose].revoke()
    delete_data(user, purpose)
```

### Disclosure (chatbot)
```typescript
function chatGreeting() {
  return "Hi! I'm an AI assistant. I can make mistakes — please verify important info.";
}
```

### Incident reporting
```python
@dataclass
class AIIncident:
    timestamp: datetime
    model: str
    severity: Literal['low', 'medium', 'high', 'critical']
    category: str  # 매 hallucination, bias, jailbreak, harm
    description: str
    affected_users: int
    root_cause: str
    mitigation: str
    
    def report(self):
        if self.severity in ('high', 'critical'):
            notify_safety_team(self)
        log_to_registry(self)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| High-risk EU | Full conformity assessment |
| Hiring / credit | Fairness audit + monitoring |
| Generative | Watermark + content provenance |
| LLM | Constitutional + RLHF + red-team |
| Privacy-sensitive | DP / federated |
| Chatbot | Disclosure + safety filter |

**기본값**: 매 model card + 매 fairness audit + 매 red-team + 매 incident reporting + 매 EU AI Act risk-tier compliance.

## 🔗 Graph
- 부모: [[AI]]
- 변형: [[AI-Safety]] · [[AI_Safety_and_Alignment|AI-Alignment]] · [[Algorithmic-Fairness]] · [[Ethics & AI|Ethics of Autonomous Systems]]
- 응용: [[EU-AI-Act]] · [[NIST-AI-RMF]] · [[AI_Safety_and_Alignment|Constitutional-AI]]
- Adjacent: [[Differential-Privacy]] · [[RLHF]]

## 🤖 LLM 활용
**언제**: 매 모든 AI deployment. 매 product launch. 매 governance.
**언제 X**: 매 academic toy.

## ❌ 안티패턴
- **Ethics-as-PR**: 매 statement only.
- **Single fairness metric**: 매 trade-off 의 ignore.
- **No red-team**: 매 jailbreak 의 surprise.
- **No incident process**: 매 learning X.
- **Ignore EU AI Act high-risk**: 매 fines + bans.

## 🧪 검증 / 중복
- Verified (EU AI Act 2024, NIST AI RMF 1.0, Anthropic RSP, Constitutional AI paper).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-20 | Auto-reinforced |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — pillars + 매 fairness / DP / model card / red-team / risk-tier code |