The threat of analytic flexibility in using large language models to simulate human data

A new study reveals that using large language models to generate synthetic datasets ("silicon samples") produces highly variable results depending on configuration choices, with correlation outcomes ranging from r=.23 to r=.84 on the same task. This demonstrates that analytic flexibility in LLM-based data generation poses a significant threat to research validity and reproducibility in social science applications.

The emergence of silicon samples—synthetic datasets created by prompting large language models to simulate human respondents—promises to accelerate social science research by eliminating recruitment constraints and costs. However, this study exposes a critical vulnerability in the methodology: the numerous discretionary choices researchers make during configuration can dramatically alter whether synthetic data actually matches human behavior patterns.

The research identifies a fundamental tension in LLM-based synthetic data generation. Across 252 different configuration combinations varying model selection, sampling parameters, prompt formats, and demographic information, no single approach proved consistently superior across validation metrics. Configurations excelling at recovering participant rankings simultaneously failed to match response distributions, illustrating that apparent fidelity is partly an artifact of which validation criterion researchers prioritize.

This matters substantially for the credibility of AI-assisted research workflows. When the same underlying task yields correlation coefficients spanning from r=.23 to r=.84 depending on configuration choices, researchers face implicit incentives to select configurations that produce desired conclusions—a form of analytic flexibility that undermines statistical validity. The problem intensifies because many configuration decisions appear methodologically defensible without prior guidance.

For AI developers and research institutions, the findings suggest that silicon samples require substantially more standardization and pre-registration before serving as reliable substitutes for human data collection. The study's recommendation for greater transparency around configuration choices and sensitivity analyses establishes a floor for responsible silicon-sample research, though adoption remains voluntary. Organizations relying on LLM-generated data for decision-making should implement validation protocols comparing synthetic outputs across multiple independent configurations.

• →LLM-generated synthetic datasets show extreme sensitivity to configuration choices, with correlation validity ranging from r=.23 to r=.84 on identical tasks
• →No single configuration approach performed consistently well across multiple validation criteria, forcing researchers to select which metrics to prioritize
• →Analytic flexibility in silicon-sample generation creates undisclosed degrees of freedom enabling researcher bias in selecting favorable outcomes
• →Current silicon-sample methodologies lack standardization guidelines and require substantial validation before replacing human-collected research data
• →The study calls for pre-registration and sensitivity analyses to mitigate configuration-driven variation in synthetic data fidelity