Who Evaluates AI's Social Impacts? Mapping Coverage and Gaps in First and Third Party Evaluations

TL;DR

This study systematically examines how AI social-impact evaluations are reported for foundation models, comparing first-party release materials with third-party evaluations and complementing quantitative analysis with stakeholder interviews. It reveals a clear split in evaluation labor: first-party reporting is sparse and often shallow, with declining depth in areas like environmental impact and bias, while third-party evaluators offer broader coverage of bias, harmful content, and performance disparities but remain limited by access to underlying data and disclosures. The work highlights persistent gaps in data provenance, content-moderation labor, financial costs, and training infrastructure reporting, driven by incentives and regulatory considerations. It concludes with policy and ecosystem design recommendations—promoting developer transparency, strengthening independent evaluation ecosystems, and building shared infrastructure for aggregating third-party evaluations—to move toward more comprehensive, accessible, and comparable social-impact reporting. The annotated dataset and accompanying code enable ongoing analysis and standardization of social-impact reporting practices across the AI ecosystem.

Abstract

Foundation models are increasingly central to high-stakes AI systems, and governance frameworks now depend on evaluations to assess their risks and capabilities. Although general capability evaluations are widespread, social impact assessments covering bias, fairness, privacy, environmental costs, and labor practices remain uneven across the AI ecosystem. To characterize this landscape, we conduct the first comprehensive analysis of both first-party and third-party social impact evaluation reporting across a wide range of model developers. Our study examines 186 first-party release reports and 183 post-release evaluation sources, and complements this quantitative analysis with interviews of model developers. We find a clear division of evaluation labor: first-party reporting is sparse, often superficial, and has declined over time in key areas such as environmental impact and bias, while third-party evaluators including academic researchers, nonprofits, and independent organizations provide broader and more rigorous coverage of bias, harmful content, and performance disparities. However, this complementarity has limits. Only model developers can authoritatively report on data provenance, content moderation labor, financial costs, and training infrastructure, yet interviews reveal that these disclosures are often deprioritized unless tied to product adoption or regulatory compliance. Our findings indicate that current evaluation practices leave major gaps in assessing AI's societal impacts, highlighting the urgent need for policies that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations in a consistent and accessible way.

Figures (26)

• Figure 1: Average scores for first-party social impact reporting per provider. Color indicates the reporting detail level (lightest green = lowest scores, medium green = mid scores, darkest green = highest scores) (see Scoring in Section \ref{['sec:methodology']} for details). In the case of multiple models per provider, we report the average detail level of the evaluation reporting. For clarity, we present results for a stratified sample of providers across geography, organizational type, and openness where available. The full sampling procedure and provider list are detailed in App. \ref{['app:stratsampling']}. The full results for the analysis can be found in App. \ref{['app:fullresults']}.
• Figure 2: Average scores and counts for third party social impact evaluations for stratified sample of providers. Rectangle size corresponds log-linearly to the number of evaluations, and color indicates the average reporting detail level (lightest blue = lowest scores, medium blue = mid scores, darkest blue = highest scores). Each cell displays the score (bold) and evaluation count (in parentheses) (see Scoring in Section \ref{['sec:methodology']} for details). For clarity, we present results for a stratified sample of providers across geography, organizational type, and openness where available. The full sampling procedure and provider list are detailed in App. \ref{['app:stratsampling']}; the full results in App. \ref{['app:fullresults']}.
• Figure 3: Average scores for first-party social impact reporting over time per release quarter. The number after the release quarter in parentheses denotes the number of models released that quarter in our dataset. Color indicates the average reporting detail level (lightest green = lowest scores, medium green = mid scores, darkest green = highest scores) (see Scoring in Section \ref{['sec:methodology']} for details). The full results for the analysis can be found in App. \ref{['app:fullresults']}. For this time-based analysis, we only consider results reported at release time. For discussions on subsequent first-party reporting, see App. \ref{['app:post_release_first_party_evaluation']}.
• Figure 4: Reporting detail level across social impact categories within select providers over model releases. Color indicates the reporting detail level (lightest green = lowest scores, medium green = mid scores, darkest green = highest scores). (see Scoring in Section \ref{['sec:methodology']} for details)
• Figure 5: Average scores for first-party social impact reporting per sector. Color indicates the average reporting detail level (lightest green = lowest scores, medium green = mid scores, darkest green = highest scores) (see Scoring in Section \ref{['sec:methodology']} for details). Full analysis results in App. \ref{['app:fullresults']}.