Design: Structured audience data for bias/fairness governance validation #1389
Description
Context
Campaign Governance's check_governance currently accepts a freeform payload for the action being validated. For bias/fairness validation (bias_fairness category), governance agents need structured audience data to evaluate targeting for discriminatory patterns.
The current freeform approach means governance agents have no reliable way to extract audience targeting details, making meaningful bias or fairness checks impractical.
Design question
What fields should structured audience data include for bias/fairness governance checks?
Key considerations:
- Segment identification: Should we require IAB Audience Taxonomy segment IDs, allow freeform segment descriptions, or support both? IAB taxonomy gives structure but may not cover all real-world segments.
- Data collection method: Is it important to know whether segments are 1P, 2P, 3P, contextual, or modeled? Collection method affects both reliability and ethical considerations.
- Consent basis: Should the structure capture consent basis (opt-in, legitimate interest, contract)? This affects whether targeting is permissible independent of bias concerns.
- Audience composition demographics: Are demographic breakdowns needed for disparate impact analysis, or can we infer enough from segment definitions alone?
- Privacy-preserving representations: Can we validate fairness without exposing raw audience data? For example, could we use aggregate statistics, k-anonymity thresholds, or segment-level metadata rather than user-level data?
Why this needs privacy expertise
The structure we choose determines what governance agents can see about audience targeting. Too little data and bias detection is impossible. Too much and we create a new privacy surface. Getting this balance right requires input from people with domain expertise in both ad-tech privacy and algorithmic fairness.
Related
Campaign Governance PR #1351 (bokelley/no-eyes-protocol) deferred this pending expert input.