Artificial intelligence can turn photos and consumer language into useful economic data, but a new econometrics method argues that researchers must also account for what the algorithm notices accidentally. A new statistical framework called two-step extraction is designed to remove that hidden bias before machine-generated features are used to estimate business effects.
The research comes from Nikhil Malik of Carnegie Mellon University’s Tepper School of Business and Max Wei of the University of Southern California’s Marshall School of Business. Their paper, Estimation Bias from Machine-Learned Features in Econometric Models, has been accepted at Management Science; a working-paper version is available through SSRN.
The problem begins when a machine-learning system is asked to extract one apparently measurable feature from unstructured material. A vision model may be estimating the presence or quality of a product, while also relying on its background, lighting, packaging, or surrounding objects. A language model may score sentiment while quietly using vocabulary, writing style, or other contextual signals. Those peripheral clues can influence the extracted value and then distort the economic relationship researchers are trying to measure.
Two-step extraction addresses that problem by adding a second machine-learning model. Instead of treating the first model’s output as clean data, the second model learns how peripheral information affects the original feature extraction. The researchers then integrate both stages with the downstream econometric model, rather than simply stacking an AI prediction on top of a conventional regression. The approach is intended to work with different feature-extraction systems and even external APIs, without requiring access to proprietary model code.
The reported results show why that distinction matters. In simulations and a large analysis of Amazon product reviews, conventional methods overstated the relationship between language sentiment and review helpfulness by as much as 228 percent. Existing corrections that failed to account for peripheral features still produced distortions of up to 41 percent, while the new method’s estimates came within 3 percent of the baseline values.
The work arrives as economists, marketers, and data scientists increasingly use pretrained computer-vision and language systems to replace manual coding of images and text. That convenience can make large datasets practical, but it also turns model behavior into part of the measurement process. The broader econometrics literature has already warned that machine-generated variables cannot automatically be treated like ordinary observed data; newer methods such as double or debiased machine learning address related inference problems involving estimated nuisance functions.
For companies, the immediate implication is less about choosing a particular AI model than auditing how its outputs enter an analysis. Two-step extraction offers a way to preserve the scale of automated feature generation while reducing the risk that an unnoticed background cue becomes an apparently meaningful business finding. The next test will be whether the framework performs as reliably across real-world datasets, model updates, and proprietary tools beyond the study’s simulations and review-analysis application.
