Inference with AI-Generated Covariates

Empirical researchers increasingly use large language models (LLMs) to extract structured features, such as sentiment scores, classifications, and expectations, from unstructured data and treat these generated features as observed covariates in downstream estimation. This practice can invalidate inference when systematic, input-dependent errors in generated features, such as hallucination and look-ahead bias, distort the downstream moment conditions. Even after correction, generated features remain noisy proxies whose error profiles differ across models and prompts. We introduce AI-Powered Inference (AI-PI), a method-of-moments framework for valid and efficient inference that combines three components: a moment-specific bias correction based on a small human-labeled calibration set; adaptive weights that optimally combine multiple model-prompt pairs; and an optimal calibration-set design that concentrates costly human labels where the generated features are least reliable. We establish consistency and asymptotic normality of the AI-PI estimator, allowing for data-adaptive labeling designs, cross-fitted LLM-pipeline tuning, and overidentified GMM. Simulations confirm substantial gains over naive LLM regressions and over debiasing without optimal weighting or labeling design. In an application to news-based sentiment and stock returns, AI-PI produces stable conclusions where naive analyses vary substantially across LLM and prompt choices, with a confidence interval roughly half as long as using the human-labeled data alone.