watermark
**AI Content Watermarking**
**Why Watermarking?**
Detect AI-generated content for authenticity verification, misinformation prevention, and attribution.
**Text Watermarking**
**Statistical Watermarking**
Subtly bias token selection during generation:
```python
def watermarked_sample(logits, prev_tokens, key):
# Create watermark hash from previous tokens
hash_value = hash(key + prev_tokens)
# Partition vocabulary into green/red lists
green_tokens = get_green_list(hash_value)
# Boost green token probabilities
for token in green_tokens:
logits[token] += delta
return sample(logits)
```
**Detection**
```python
def detect_watermark(text, key, threshold=0.5):
tokens = tokenize(text)
green_count = 0
for i, token in enumerate(tokens):
hash_value = hash(key + tokens[:i])
green_list = get_green_list(hash_value)
if token in green_list:
green_count += 1
z_score = (green_count - expected) / std
return z_score > threshold
```
**Image Watermarking**
| Technique | Approach |
|-----------|----------|
| Visible | Overlay logo/text |
| Invisible | Modify pixel values imperceptibly |
| AI detection | Train classifier on AI images |
| C2PA metadata | Content authenticity standard |
**Challenges**
| Challenge | Consideration |
|-----------|---------------|
| Robustness | Watermarks may be removed |
| Paraphrasing | Text rewrites remove watermark |
| Quality impact | May slightly affect output quality |
| Adversarial | Active attempts to evade |
**Detection Services**
| Service | Content Type |
|---------|--------------|
| GPTZero | Text |
| OpenAI classifier | Text |
| Hive | Images |
| Content Credentials | Images (standard) |
**C2PA Standard**
Industry standard for content authenticity:
```
Image metadata includes:
- Creation tool
- Edit history
- Creator identity
- Generating AI model
```
**Best Practices**
- Combine multiple detection methods
- Train on diverse AI-generated content
- Account for false positives
- Update detectors as models evolve
- Transparency about detection limits