Gaps in First-Party and Third-Party AI Model Evaluations

2025-07-08 · Source: AI Accountability Review · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

Researchers from the EvalEval Coalition published a paper analyzing 186 first-party and 183 third-party evaluations of AI models regarding social impacts. They assessed seven dimensions: Bias and Harm, Sensitive Content, Performance Disparity, Environmental Costs, Privacy and Data, Financial Costs, and Moderation Labor, using a 0-3 detail scale. The analysis revealed that third-party evaluations were significantly more detailed, averaging 2.62 compared to 0.72 for first-party reports. This indicates model developers provide less detail on social impact evaluations. The study also found that popular US models attract the most third-party scrutiny, leaving less popular models underevaluated. Furthermore, certain impacts like data and content moderation are largely absent from third-party evaluations due to lack of access to necessary information.

Key takeaway

For CTOs and VPs of Engineering evaluating AI models for deployment, recognize that first-party social impact assessments are often insufficient. You should prioritize models with robust third-party evaluations and advocate for industry-wide transparency standards. This will help your teams better understand and mitigate potential social risks, ensuring more responsible AI integration and compliance with future regulations.

Key insights

Third-party AI model evaluations are significantly more detailed than first-party assessments, revealing critical transparency gaps.

Principles

• First-party AI evaluations lack sufficient detail.
• Evaluation coverage varies by model popularity.
• Third parties face data access limitations.

Method

The analysis compared 186 first-party and 183 third-party AI model evaluation reports, rating them on a 0-3 scale across seven social impact dimensions to quantify detail levels.

In practice

• Review third-party evaluations for deeper insights.
• Prioritize transparency in model development.
• Advocate for standardized evaluation data disclosure.

Topics

• AI Model Evaluation
• Social Impact Assessment
• First-Party vs. Third-Party Evaluation
• AI Transparency Standards
• AI Accountability

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Researcher, AI Ethicist, Policy Maker

Related on AIssential

• Report on foundation model impacts released — Foundation model providers show inconsistent transparency in documenting their models' post-deployment societal impacts.
• AI incidents, audits, and the limits of benchmarks — AI safety requires systematic incident reporting and rigorous third-party evaluation to manage real-world risks.
• Closing Information Gaps via AI Transparency — AI transparency is crucial for accountability, requiring high-quality information about both technical and human system components.
• 😺 Someone firebombed Sam Altman's house — A widening perception gap between AI experts and the public fuels distrust, alongside emerging AI-driven risks and disruptions.
• From Age Gates to Accountability in AI Design — Disparate impact doctrine offers a framework to assess AI systems for disproportionate harm to children, shifting focus from access to design.
• Who decides what AI tells you? Campbell Brown, once Meta’s news chief, has thoughts — AI models exhibit significant accuracy and bias issues on high-stakes topics, necessitating expert-driven evaluation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Accountability Review.