David Arbour, Eli Ben-Michael, Avi Feller, Apoorva Lal, Lo-Hua Yuan
arXiv 7 Jun 2026 · Econometrics
arXiv:2606.08853 · PDF · DOI · OpenAlex · Extracted main text
Generative AI and large language models can produce realistic predictions of human behavior from rich, unstructured inputs with little to no task-specific training data. Recent work uses these “digital twin” predictions to supplement human responses in surveys and experiments. We study the special case of using AI-generated predictions to reduce variance in randomized experiments. We argue that doing so requires no new estimators and that researchers can simply include AI predictions as covariates in standard regression adjustment, analogous to adjusting for a prognostic score. A benefit of this approach is a “do no harm” property whereby the adjusted estimator reverts to the unadjusted difference in means when predictions are uninformative. Other methods, such as variants of prediction-powered inference, do not have this guarantee. We provide implementation guidance, including how to obtain continuous scores from discrete LLM outputs and how to use LLMs to featurize unstructured inputs as auxiliary covariates. We demonstrate these ideas in simulations and three empirical applications: a survey mega-study, an email marketing A/B test, and a large-scale technology platform experiment. Overall, efficiency gains are real if modest, with greater benefits in studies that contain substantial text and other unstructured data. We also confirm the do no harm property empirically. Given these gains and limited costs, we recommend adjusting for AI-generated predictions as a regular empirical practice.
appendix boundary found by appendix_command · 76% of the source is main text. Read the extracted text to check this.
The works this paper leans on most, across its whole bibliography — not restricted to papers in our corpus. Ranked by composite intensity, which combines how often a work is mentioned, how many sections mention it, and how much of that falls in the main text rather than the appendix.
| Reference | Intensity | Mentions | Sections | Main text | |
|---|---|---|---|---|---|
| 1 | Winston Lin (2013) Agnostic notes on regression adjustments to experimental data: Reexamining Freedman's critique | 1.000 | 7 | 3 | 100% |
| 2 | Naoki Egami, Musashi Hinck, Brandon Stewart, and Hanying Wei (2023) Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large l… | 1.000 | 6 | 5 | 100% |
| 3 | Alex Deng, Luke Hagar, Nathaniel Stevens, Tatiana Xifara, Lo-Hua Yua… (2023) From augmentation to decomposition: A new look at CUPED in 2023 self | 1.000 | 5 | 3 | 100% |
| 4 | Tianyi Peng, George Gui, Melanie Brucks, Daniel J Merlau, Grace Jiar… (2025) Digital twins as funhouse mirrors: Five key distortions | 1.000 | 5 | 3 | 100% |
| 5 | Xinrui Ruan, Xinwei Ma, Yingfei Wang, Waverly Wei, and Jingshen Wang (2025) Can language models boost the power of randomized experiments without statistical bias? | 1.000 | 5 | 3 | 100% |
| 6 | Peng Ding (2024) A first course in causal inference | 0.874 | 6 | 2 | 100% |
| 7 | Anastasios N Angelopoulos, Stephen Bates, Clara Fannjiang, Michael I… (2023) Prediction-powered inference | 0.874 | 5 | 2 | 100% |
| 8 | Piersilvio De Bartolomeis, Javier Abad, Guanbo Wang, Konstantin Donh… (2025) Efficient Randomized Experiments Using Foundation Models, February 2025 | 0.843 | 3 | 3 | 100% |
| 9 | Chris Engh and P. M. Aronow (2025) Using LLMs to directly guess conditional expectations can improve efficiency in causal estimation | 0.843 | 3 | 3 | 100% |
| 10 | Lisa P Argyle, Ethan C Busby, Nancy Fulda, Joshua R Gubler, Christop… (2023) Out of one, many: Using language models to simulate human samples | 0.737 | 3 | 2 | 100% |
Showing the top 10 of 46 scored citations.