2nd/10_Wiki/Topics/AI_and_ML/Intellectual-Property-in-AI.md

---
id: wiki-2026-0508-intellectual-property-in-ai
title: Intellectual Property in AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI IP, copyright, training data, model IP, fair use, NYT v OpenAI]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [legal, ai-ip, copyright, training-data, fair-use, regulation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Legal
  applicable_to: [AI Development, Legal, Policy]
---

# Intellectual Property in AI

## 매 한 줄
> **"매 training data, 매 model output, 매 model itself 의 IP 의 의 의 의 unsettled"**. 매 NYT v OpenAI (2023+), Getty v Stability, GitHub Copilot lawsuits. 매 modern: 매 EU AI Act + 매 US Copyright Office (2023).

## 매 핵심

### 매 issues
- **Training data**: 매 copyrighted material 의 의 fair use?
- **Output**: 매 AI-generated 의 copyrightable?
- **Model**: 매 trade secret vs open-source.
- **Style**: 매 artist style 의 mimic 의 violate?

### 매 famous cases
- **NYT v OpenAI** (2023+): 매 training on articles.
- **Getty v Stability** (2023+): 매 watermarks in output.
- **Andersen v Stability** (artists vs SD).
- **Doe v GitHub** (Copilot, code).
- **Authors Guild v OpenAI** (2023).

### 매 legal stance (current, evolving)
- **US Copyright Office (2023)**: 매 pure AI output 의 X copyright (no human authorship).
- **EU AI Act (2024)**: 매 training data disclosure 의 transparency.
- **Japan**: 매 broad permitted training (2018 amendment).
- **UK**: 매 narrow text-and-data-mining exception.

### 매 응용 risk
1. Training data sourcing.
2. Output deployment.
3. Style mimicking.
4. Model release.
5. Watermark / provenance.

## 💻 패턴

### Training data audit
```python
@dataclass
class DataSource:
    source: str
    license: str
    provenance: str
    can_train: bool

def audit_training_corpus(sources):
    risky = [s for s in sources if not s.can_train or s.license == 'unknown']
    return {'safe': len(sources) - len(risky), 'risky': risky}
```

### License compatibility
```python
COMPATIBLE = {
    'cc0': True, 'cc-by': True, 'mit': True, 'apache-2.0': True,
    'cc-by-nc': 'check_purpose',
    'cc-by-sa': 'derivative_must_share',
    'gpl-3.0': 'derivative_must_open',
    'proprietary': False, 'unknown': False,
}

def can_train(license, purpose='commercial'):
    rule = COMPATIBLE.get(license)
    if rule == 'check_purpose': return purpose != 'commercial'
    return rule
```

### Output attribution / watermark
```python
# 매 C2PA (modern provenance standard)
from c2pa import Signer

def attach_provenance(media_path, model_id, signer_cert):
    Signer(signer_cert).sign(media_path, claims={
        'generator': model_id,
        'training_data_summary': 'public_domain + licensed',
        'timestamp': now(),
    })
```

### Artist style detection (defensive)
```python
def style_similarity(generated, reference_artist_works):
    """매 매 generated style 의 reference artist 의 의 의 close?"""
    gen_features = clip_encode(generated)
    artist_features = [clip_encode(w) for w in reference_artist_works]
    sim = max(cosine(gen_features, f) for f in artist_features)
    return sim  # 매 > 0.9 → flag
```

### Opt-out registry
```python
OPT_OUT = load_registry('https://spawning.ai/opt-out')

def filter_training_data(images):
    return [img for img in images if img.creator not in OPT_OUT]
```

### Memorization detection (training data leakage)
```python
def detect_memorization(model, training_examples, n_test=100):
    """매 매 model 의 의 의 verbatim 의 reproduce 매?"""
    leaks = 0
    for ex in random.sample(training_examples, n_test):
        prompt = ex.text[:100]
        gen = model.generate(prompt, max_tokens=200)
        if longest_common_substring(gen, ex.text) > 50:
            leaks += 1
    return leaks / n_test
```

### Fair use 4-factor analysis
```python
def fair_use_analysis(use_case):
    return {
        'purpose': 'transformative? commercial?',
        'nature': 'creative or factual? published?',
        'amount': 'how much used? heart of work?',
        'effect': 'market harm? substitute?',
    }
# 매 매 case 의 의 의 의 evaluate — 매 lawyer 의 needed
```

### EU AI Act compliance (training data summary)
```python
def eu_training_data_disclosure(corpus):
    return {
        'general_purpose_ai': True,
        'training_data_summary': summarize_corpus(corpus),
        'compute_used': estimate_compute(corpus),
        'systemic_risk': flops_above_threshold(),
    }
```

### Model release license
```yaml
# 매 매 trade-off
licenses:
  - name: Llama Community License
    type: permissive_with_exceptions
    commercial: yes (with conditions)
    
  - name: Apache 2.0
    type: permissive
    commercial: yes
    
  - name: AGPL-3.0
    type: copyleft
    commercial: yes (must share derivatives)
    
  - name: CC-BY-NC
    type: non_commercial
    commercial: no
```

### Output cleansing (preserve user IP)
```python
def output_clean_for_user_ip(generated, user_input):
    """매 generated 의 의 user input 의 verbatim 매 가능."""
    if generated_contains_user_input(generated, user_input):
        # 매 user retains rights to their part
        return mark_user_section(generated, user_input)
    return generated
```

### LLM legal-compliance prompt
```python
LEGAL_SYSTEM = """You generate legal-aware output.

When asked about IP-sensitive content:
1. Note that AI-generated work may not be copyrightable in some jurisdictions.
2. Cite training data limitations when relevant.
3. Flag if a request seems to ask for verbatim copyrighted material.
4. Recommend lawyer consultation for legal decisions."""
```

### Code verbatim check (Copilot-style)
```python
def code_verbatim_check(generated_code, public_repos):
    """매 매 매 long verbatim 의 detect → user 의 warn."""
    matches = []
    for repo in public_repos:
        for file in repo.files:
            common = longest_common_substring(generated_code, file.content)
            if len(common) > 100:
                matches.append({'repo': repo.name, 'license': repo.license, 'lines': common})
    return matches
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Build model | License audit + opt-out respect |
| Deploy output | Watermark + provenance |
| Style mimicking | Detection + flag |
| EU market | AI Act disclosure |
| Open-source | Apache / Llama license |
| User-generated | Preserve user rights |

**기본값**: 매 license-clean training (audit + opt-out) + 매 watermark output (C2PA) + 매 EU disclosure + 매 lawyer consult for edge cases.

## 🔗 Graph
- 부모: [[Ethics & AI]]
- 변형: [[Model-IP]]
- 응용: [[EU-AI-Act]] · [[GDPR]] · [[C2PA]]
- Adjacent: [[Generative-AI]] · [[Copyright]]

## 🤖 LLM 활용
**언제**: 매 commercial AI deploy. 매 dataset construction.
**언제 X**: 매 academic research only (limited).

## ❌ 안티패턴
- **Train on anything**: 매 lawsuits.
- **No watermark**: 매 misuse / impersonation.
- **Ignore opt-out**: 매 brand risk.
- **No EU AI Act prep**: 매 fines.
- **Skip lawyer**: 매 specific case decisions.

## 🧪 검증 / 중복
- Verified (US Copyright Office 2023, EU AI Act 2024, court filings).
- 신뢰도 B+.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — IP issues + 매 audit / watermark / fair use / disclosure code |