Generative Augmented Inference

Large language models enable inexpensive AI-generated annotations, but using them reliably for causal inference remains challenging. Naively pooling AI and human data induces bias, while existing methods such as Prediction-Powered Inference (PPI; Angelopoulos et al., 2023) treat AI outputs as proxies of true labels -- an assumption often violated for generative model outputs in practice. We propose Generative Augmented Inference (GAI), a framework that treats AI outputs as general, potentially high-dimensional informative features for learning human labels rather than as surrogates. GAI flexibly models this relationship using nonparametric methods, enabling consistent estimation and valid inference from combined human and AI data. We establish asymptotic normality and show that, under random labeling, GAI strictly improves asymptotic efficiency over human-data-only estimation whenever AI outputs are informative for true labels. Empirical studies on real-world datasets demonstrate that GAI significantly reduces estimation error and improves confidence interval quality across diverse generative data sources relative to human-only and PPI-based estimation.

Human labels are often expensive to collect, while AI systems such as large language models can produce large amounts of cheap auxiliary data. However, AI-generated answers are not always reliable: they can be biased, only weakly accurate, or expressed in forms that do not directly match human labels. This creates a challenge for researchers who want to use AI-generated data without letting AI mistakes distort their scientific conclusions. We propose Generative Augmented Inference (GAI), a statistical framework that uses AI outputs as helpful clues rather than as replacements for human labels. Instead of assuming that AI answers are correct or nearly correct, our method learns how AI-generated information relates to human responses and then corrects for remaining errors. This allows researchers to combine a small amount of human-labeled data with abundant AI-generated data while still producing valid estimates and uncertainty measures. Across real-world studies, our method improves estimation accuracy and confidence interval quality compared with human-only and existing AI-augmented approaches. This can substantially reduce data-collection costs in fields such as market research, medicine, and social science.