AI Governance for Video Processing

The Training Data Question

The #1 question from compliance teams: "Is my video used to train your AI?"

BetterVideo's Position: No. Never.

• We use pre-trained, fixed-weight models
• Your video is processed and returned, period
• No training, no fine-tuning, no aggregation
• This is architectural, not policy

Why This Matters

Training on customer video creates problems:

• Privacy: Patterns from your video leak into the model
• Consent: Did subjects consent to model training?
• IP: Is your video now part of their asset?
• Competition: Are competitors benefiting from your data?

Questions to Ask Vendors

• Is customer video used for model training?
• Can you contractually commit to no training?
• Is there a separate pricing tier to avoid training?
• How do you prevent training if processing is outsourced?

Bias and Fairness

Video enhancement AI can exhibit biased performance:

Known Bias Vectors

• Skin tone: Models may perform differently on light vs. dark skin
• Lighting: Underperformance on low-light footage common in surveillance
• Camera type: Training on high-quality cameras may hurt CCTV performance
• Face angle: Profile and partial faces may enhance poorly

Mitigation Approaches

• Diverse training data: Models trained on varied demographics, conditions
• Multi-model ensemble: Multiple models catch each other's weaknesses
• Continuous monitoring: Track performance across segments
• Transparent testing: Publish performance across conditions

BetterVideo's Approach

We test enhancement quality across:

• Fitzpatrick skin type scale
• Lighting conditions (daylight, indoor, low-light)
• Camera types (smartphone, CCTV, dashcam)
• Face angles and occlusions

Explainability

Regulated industries need to explain what the AI does:

What We Can Explain

• Model architecture: Type of neural network, training approach
• Processing operations: Upscaling, sharpening, denoising, face restoration
• Parameters applied: Strength, thresholds, model versions
• Quality metrics: Resolution, sharpness scores before/after

What We Can't Explain (Fully)

• Individual pixel decisions: Why this pixel is this color
• Internal representations: What the model "sees"

This is the nature of neural networks. But for video enhancement, the explanation that matters is: "The AI increased resolution and sharpness without adding or changing content."

Documentation for Regulators

We provide:

• Model cards describing each model
• Processing logs for each job
• Methodology papers for legal defense
• Audit-ready compliance documentation

Human Oversight

AI governance frameworks (EU AI Act, NIST AI RMF) emphasize human oversight:

Where Humans Are in the Loop

• Input selection: Humans decide what to process
• Output review: Humans review enhanced video
• Decision making: AI enhances, humans decide what it means

Where Automation Is Appropriate

• Processing: AI handles the technical enhancement
• Quality checks: Automated validation of output
• Deletion: Automatic per retention policy

The principle: AI augments human judgment, it doesn't replace it. Video enhancement makes video easier to see; humans still interpret what it shows.