Disclosure and Evaluation as Fairness Interventions for General-Purpose AI

Fairness is context-specific, so general-purpose AI systems need process-based rather than outcome-based fairness approaches. Since the same AI model serves multiple contexts with different fairness requirements, the paper argues we should focus on ensuring stakeholders follow fair processes rather than trying to enforce uniform fairness outcomes across all uses.

AI providers should invest in research to understand how their development decisions affect fairness downstream. The paper emphasizes that "the current state of AI understanding offers limited insight into how upstream factors translate into downstream fairness impacts," making it critical for providers to study these connections rather than assuming their models will be fair in all deployment contexts.

Transparency about model deployment is essential for external accountability. The paper recommends that system providers disclose "whom they are serving their models to, or at the very least, reveal sufficient information for external researchers to conduct such research"—enabling independent evaluation of fairness impacts that providers themselves may not anticipate.

System deployers bear responsibility for evaluating fairness in their specific contexts. Because deployers are "closer to real-world contexts and can leverage their proximity to end users," they should "conduct rigorous evaluations across different levels of fairness" and responsibly disclose information about how they personalize systems for different users.

Information-gathering and disclosure should be prioritized over attempting to enforce fairness outcomes. Rather than regulatory mandates about what "fair" results look like, the paper advocates for systematic processes that ensure stakeholders intentionally gather and share information needed to identify and address fairness harms.