ai-advertising-targeting-fairness-review

AI Advertising Targeting Fairness Review

Purpose

This skill reviews ad-platform audience targeting configurations and declared AI feature usage for protected-class discrimination risk under the Fair Housing Act (42 U.S.C. §3604), the Equal Credit Opportunity Act (ECOA), and EU AI Act Article 5. Ad platforms increasingly offer AI-driven audience expansion features — Meta Advantage+ Audience, Google broad match and Performance Max, DSP algorithmic deal targeting — that optimize delivery based on historical conversion patterns. When historical converters skew along protected-class lines (race, sex, age, national origin, familial status, disability, religion), algorithmic optimization propagates that skew without explicit intent. The review examines declared AI feature usage, audience seed composition, interest-segment proxy risk, and the absence of protected-category exclusion declarations on special-category campaigns before the configuration ships.

Lean operating rules

• Treat Meta Advantage+ Audience enabled on a housing, credit, employment, or insurance campaign with no declared protected-category exclusions as HIGH — the system expands targeting beyond the declared audience using engagement signals that may correlate with race, sex, or national origin.
• Treat interest-based segments that function as proxies for health conditions, religion, national origin, or familial status used on an insurance or financial-services campaign as HIGH — proxy targeting on protected classes is substantively equivalent to explicit targeting under FHA and ECOA case law.
• Treat automated bidding (Target CPA, Target ROAS, Smart Bidding) optimizing a credit-offer, rental, or employment campaign on lookalike audiences seeded from historical converters as HIGH — disparate impact is propagated algorithmically when the seed population reflects historical discriminatory patterns.
• Treat any AI-generated audience expansion (broad match, Performance Max audience signals, DSP algorithmic reach extension) active on a special-category campaign (housing, credit, employment, insurance) with no fairness audit trail as HIGH — the optimization objective does not include disparate-impact minimization.
• Treat geofencing or geographic exclusion zones that closely follow racially or ethnically concentrated neighborhood boundaries on a housing or credit campaign as HIGH — geographic redlining is prohibited under FHA regardless of whether intent is declared.
• Treat the absence of a Special Ad Category declaration on a Meta campaign reasonably classifiable as housing, employment, or credit as HIGH — the declaration unlocks mandatory fairness restrictions; omitting it circumvents them.
• Flag automated bidding that optimizes on a conversion event defined as a past purchase or application when the historical converter population is not documented for demographic representativeness as MEDIUM — undocumented seed bias is a disparate-impact risk even when not yet proven.
• Flag interest segments that include health-condition or medication-related categories on campaigns not in the healthcare vertical as MEDIUM — health proxies reach users based on inferred sensitive characteristics.
• Flag AI feature disclosures that are absent or vague (e.g., "algorithmic optimization enabled" with no named feature, no version, no opt-out path) as MEDIUM — EU AI Act Article 13 and FTC guidance require meaningful transparency.
• Do not recommend disabling AI features without naming the performance impact and the manual alternative that preserves reach.
• Label every finding with evidence basis: audience spec provided, AI feature declaration provided, documentation-based, or inference from missing config.

References

Load these only when needed:

• [Workflow and output contract](references/workflow-and-output.md) — use when executing the full review or formatting the final answer.

Response minimum

Return, at minimum:

• AI feature inventory (named features enabled per campaign, evidence basis)
• Special-category campaign detection (housing, credit, employment, insurance)
• Protected-class proxy segment assessment (interest segments, lookalike seeds)
• Algorithmic disparate-impact assessment (bidding, audience expansion)
• Special Ad Category declaration check (Meta) or equivalent platform declaration
• Severity-labelled finding list (critical / high / medium / low)
• Safe next actions