TLDR: A new comprehensive study reveals significant gaps in how AI’s social impacts, such as bias, privacy, and environmental costs, are evaluated and reported. First-party (developer) reporting is often sparse and declining, while third-party (independent) evaluations offer more rigor in some areas but cannot cover all aspects. Critical information like data provenance and content moderation labor is frequently overlooked due to a lack of incentives, measurement challenges, and strategic deprioritization. The research emphasizes an urgent need for greater transparency, stronger independent evaluation ecosystems, and policy reforms to ensure a more complete understanding of AI’s societal footprint.
As artificial intelligence, particularly generative AI, becomes increasingly integrated into high-stakes systems, the need to understand its societal implications has never been more critical. Governance frameworks are now heavily reliant on evaluations to assess the risks and capabilities of these powerful AI models. While evaluations of general AI capabilities are common, a new comprehensive study reveals a significant disparity in how social impact assessments—covering crucial areas like bias, fairness, privacy, environmental costs, and labor practices—are reported across the AI ecosystem.
The research, titled “WHO EVALUATES AI’S SOCIAL IMPACTS? MAPPING COVERAGE AND GAPS IN FIRST AND THIRD PARTY EVALUATIONS,” conducted the first large-scale analysis of both first-party (model developers) and third-party (independent organizations, academia, non-profits) social impact evaluation reporting. The study meticulously examined 186 first-party release reports and 183 post-release evaluation sources, complementing this quantitative analysis with in-depth interviews with model developers.
A clear division of labor emerged from the findings. First-party reporting by model developers was found to be sparse, often superficial, and has notably declined over time in key areas such as environmental impact and bias. In contrast, third-party evaluators, including academic researchers, non-profits, and independent organizations, provide broader and more rigorous coverage of bias, harmful content, and performance disparities. This suggests a complementary relationship where independent bodies often fill the gaps left by developers.
However, this complementarity has its limitations. The study highlights that certain critical disclosures can only be authoritatively reported by model developers themselves. These include data provenance, content moderation labor practices, financial costs associated with development, and details of training infrastructure. Interviews with developers revealed that these disclosures are frequently deprioritized unless directly tied to product adoption or regulatory compliance. This creates significant blind spots in understanding the full societal impact of AI.
The research also observed a concerning trend: reporting on social impact dimensions has generally decreased over time. Specifically, environmental costs and emissions reporting saw a significant decline after the third quarter of 2023, and similar patterns were noted for evaluations of bias, stereotypes, and representational harms. Developers cited reasons such as the contextual nature of bias, the desire to avoid negative publicity, and the sensitive nature of environmental data as factors contributing to this decline.
One of the most striking findings was the near absence of reporting on data and content moderation labor. This crucial area, which impacts human workers globally, was reported in only a small fraction of first-party reports, and third-party reporting was largely non-existent. Interviewees emphasized the importance of this dimension, noting its significant and often disparate impact on individuals, yet acknowledging it is frequently overlooked due to measurement difficulties and a lack of attention.
Geographical and sectoral patterns also revealed interesting insights. Academia generally leads in first-party social impact reporting, followed by non-profits and industry. However, evaluations tend to concentrate on the most prominent and commercially influential systems, particularly those developed in the US and China. This popularity-driven focus inadvertently creates transparency gaps for low-resource language models, which receive far less scrutiny regarding their social impacts and risks.
Also Read:
- Charting a Sustainable Path for Artificial Intelligence: A Deep Dive into Green AI Practices
- Agentic AI Transforms Supply Chain Sustainability and Efficiency
The study concludes that current evaluation practices leave major gaps in assessing AI’s societal impacts. These gaps stem from a combination of structural difficulties (like the lack of reliable methodologies for privacy or labor reporting) and strategic deprioritization by developers due to reputational or regulatory risks. To address these issues, the paper calls for urgent policy interventions that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations consistently and accessibly. Investment in standardized frameworks, automated tools, and multi-stakeholder coordination are seen as crucial steps forward. For more details, you can read the full research paper here.
