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:

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

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

Challenges

Detection Services

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