2nd/10_Wiki/Topics/AI_and_ML/Authenticity.md

---
id: wiki-2026-0508-authenticity
title: Authenticity
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [진정성, AI authenticity, content provenance, C2PA, deepfake detection]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: applied
tags: [authenticity, ethics, branding, ai-disclosure, c2pa, watermark, provenance, deepfake]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: ethics / cryptography
  applicable_to: [Brand Strategy, Content Provenance, AI Disclosure]
---

# Authenticity

## 📌 한 줄 통찰
> **"매 진짜 의 힘"**. 매 internal value + 매 external action 의 일치. 매 deepfake 시대 의 가장 큰 differentiator. 매 AI 의 "I'm an AI" 의 honesty 의 trust 의 maximum.

## 📖 핵심

### 매 layer
1. **Internal consistency**: 매 self 의 honesty.
2. **Relational transparency**: 매 mask X.
3. **Moral courage**: 매 cost 가 있어도 매 belief.
4. **Vulnerability**: 매 weakness 의 share (Brené Brown).
5. **Provenance**: 매 origin 의 verify 가능.

### 매 modern context
1. **Brand authenticity**: 매 marketing 의 가장 큰 lever (Edelman Trust Barometer).
2. **Influencer**: 매 BeReal, 매 unfiltered.
3. **AI content**: 매 disclosure 의 default.
4. **Deepfake era**: 매 provenance 의 의무.
5. **Whistleblower / journalism**: 매 source verify.

### 매 AI 시대 의 challenge
- **Generative content**: 매 image / voice / video 의 indistinguishable.
- **Personalized deepfake**: 매 target 의 specific.
- **Voice cloning**: 매 3 second 로 OK.
- **Synthetic media**: 매 election interference.
- **Bot persona**: 매 Twitter / Reddit 의 manipulation.

### 매 verification standard

#### C2PA (Coalition for Content Provenance)
- 매 cryptographic signature 의 manifest.
- 매 camera → edit → publish 의 chain.
- 매 Adobe / Microsoft / Sony 의 backing.

#### IPTC Photo Metadata
- 매 EXIF 의 extension.
- 매 capture / edit history.

#### Watermark (visible / invisible)
- 매 SynthID (Google).
- 매 statistical watermark in LLM output.

#### Blockchain provenance
- 매 NFT 의 origin.
- 매 immutable timestamp.

### 매 detection
- **Deepfake detection**: 매 ML 기반 (FaceForensics++).
- **Voice deepfake**: 매 spectral analysis.
- **AI-text detection**: 매 GPTZero, 매 Originality.ai (매 false positive 많음).
- **Reverse image search**: 매 source 의 trace.

### 매 ethical 권장
- **AI 사용 의 disclose**.
- **Synthetic content 의 watermark**.
- **Source 의 verify**.
- **Persona 의 honest** (no false biography).
- **Vulnerability OK**.

## 💻 패턴 (응용 — provenance + disclosure)

### C2PA manifest (구조)
```json
{
  "claim_generator": "Adobe Photoshop 25.0",
  "format": "image/jpeg",
  "instance_id": "xmp:iid:abc123",
  "claim": {
    "title": "My Photo",
    "format": "image/jpeg",
    "assertions": [
      { "label": "c2pa.actions",
        "data": { "actions": [{ "action": "c2pa.created" },
                              { "action": "c2pa.edited",
                                "parameters": { "name": "color-correct" } }] } },
      { "label": "c2pa.training-mining",
        "data": { "entries": { "c2pa.ai_generative_training": { "use": "notAllowed" } } } }
    ]
  },
  "signature": "..."
}
```

### AI disclosure (UI)
```tsx
<Message author="AI Assistant">
  <DisclosureBadge>AI-generated</DisclosureBadge>
  {content}
</Message>
```

→ 매 message-level explicit.

### LLM watermark (Aaronson scheme)
```python
# 매 generation 의 token 선택 의 cryptographic hash 의 bias
def watermark_logits(logits, prev_token, key):
    h = hash(prev_token + key)
    bias = derive_bias(h, vocab_size)  # 매 small bias
    return logits + bias

# Detection
def detect_watermark(text, key):
    score = sum(check(token, prev, key) for prev, token in pairs(text))
    return score > THRESHOLD
```

### Content authentication (verify chain)
```python
def verify_c2pa(image_path):
    manifest = read_c2pa_manifest(image_path)
    if not manifest: return 'unverified'
    
    if not verify_signature(manifest): return 'tampered'
    
    chain = manifest.get('chain', [])
    for step in chain:
        if not verify_step(step): return 'broken_chain'
    
    return f'authentic, {len(chain)} edits tracked'
```

### Persona honesty
```ts
const aiPersona = {
  name: 'Aria',
  identity: 'AI assistant',  // 매 honest
  // ❌ NOT biography: 'Born in Seattle, 25 years old'
  greeting: "Hi! I'm Aria, an AI. How can I help?",
  responseToHumanQuestion: () =>
    "I'm an AI — I don't have personal experiences, but I can help you think through this.",
};
```

## 🤔 결정 기준
| 상황 | 적용 |
|---|---|
| Brand strategy | Vulnerability + consistency + transparency |
| AI agent | Identity disclose + persona honest |
| Generative content | C2PA + watermark + disclosure |
| Journalism | Source verify + provenance |
| Marketing | Genuine story > polished |
| Influencer | Behind-the-scenes + flaws OK |

**기본값**: 매 disclose + provenance + vulnerability.

## 🔗 Graph
- 응용: [[C2PA]] · [[Content-Provenance]]
- Adjacent: [[Deepfake]] · [[Anthropomorphism]] · [[EU-AI-Act]]

## 🤖 LLM 활용
**언제**: 매 brand / agent persona design. 매 content provenance system. 매 AI disclosure policy.
**언제 X**: 매 fake "vulnerability" 의 manipulation. 매 manufactured 'authentic' marketing.

## ❌ 안티패턴
- **Performative vulnerability**: 매 fake 의 share.
- **Fake AI persona biography**: 매 deception.
- **No disclosure**: 매 trust 의 long-term destroy.
- **Watermark 의 옵션** (audit 없이): 매 disable 의 OK.
- **C2PA 의 partial**: 매 missing step 의 invalidate.
- **"매 perfect = 매 authentic"**: 매 polished 의 manufactured 의 의심.

## 🧪 검증 / 중복
- Verified (C2PA spec, Edelman Trust Barometer, Brown).
- 신뢰도 B.
- Related: [[C2PA]] · [[Anthropomorphism]] · [[AI-Disclosure]] · [[Deepfake]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — layers + C2PA + watermark + AI disclosure |