Research map / Large language models / AI and society

Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge

ICML 2025 · Large language models › AI and society · approach: Social-science view of AI evaluation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Position paper arguing that evaluating GenAI systems should be treated as a social science measurement challenge, proposing a four-level measurement framework (background concept, systematized concept, measurement instruments, measurements) with validity lenses.

Problem

Measurement tasks in evaluating generative AI systems lack scientific rigor: concepts are abstract and contested, and ML instruments often conflate conceptual and operational choices, making it unclear what is measured and whether it is valid.

Approach

Adapts a social-science measurement theory framework to GenAI evaluation, separating systematization (narrowing a background concept into an explicit systematized concept) from operationalization (building instruments), and using validity lenses to interrogate both conceptual and operational debates.

Key results

• Proposes a four-level framework (background concept, systematized concept, measurement instruments, measurements) linked by systematization, operationalization, application, and interrogation.
• Recommends seven validity lenses (face, content, convergent, discriminant, predictive, hypothesis, consequential) to interrogate validity.
• Illustrates the framework with a running example on stereotyping text, plus appendix examples on math reasoning, memorization, and refusal.

Surprising findings

• Argues that human-LLM inter-annotator agreement reported as judge 'accuracy' implicitly treats human annotations as ground truth, which is problematic.
• Argues that the framework's social-science roots do not mean GenAI should be anthropomorphized or that human psychometric tests transfer directly to GenAI.

Limitations

• Framework is conceptual and not empirically validated on a set of real evaluations.
• Determining what counts as 'sufficiently valid' has no universal answer and is acknowledged as difficult.
• Paper omits detailed discussion of application process and some validity lenses to save space.

Experimental setup

Baselines: StereoSet CrowS-Pairs HELM BIG-bench DecodingTrust TrustLLM

Builds on: measurement theory framework of Adcock and Collier (2001) psychometrics-based benchmark critiques Messick validity framework

Closest papers by meaning

• Position: AI Competitions Provide the Gold Standard for Empirical Rigor in GenAI Evaluation (ICML 2025 · similarity 0.87)
• Fostering the Ecosystem of AI for Social Impact Requires Expanding and Strengthening Evaluation Standards (NeurIPS 2025 · similarity 0.86)
• No Free Delivery Service: Epistemic limits of passive data collection in complex social systems (NeurIPS 2024 · similarity 0.86)
• Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.85)
• Braess’s Paradox of Generative AI (AAAI 2025 · similarity 0.85)
• AGI-Elo: How Far Are We From Mastering A Task? (NeurIPS 2025 · similarity 0.85)