AI as a Tool for Economic Research

A capability-based map of AI in economics

The most useful classification is not a list of supervised learning, deep learning, or large language models. It is the task AI performs in the research pipeline: generating behavior, converting unstructured information into variables, assisting causal estimation, or forecasting complex dynamics.

A single model can serve several tasks, but the validation target changes with the task. Behavioral similarity, measurement validity, causal credibility, and predictive accuracy are not interchangeable.

• Behavior generation: compare means, distributions, individual predictions, and cross-task stability
• Information representation: validate labels, context sensitivity, and external validity
• Causal estimation: state the identification assumptions and report orthogonalization, cross-fitting, and robustness
• Dynamic forecasting: use rolling out-of-sample tests and evaluate structural breaks and tail risk

AI and Experimental Economics: Behavior Generation

LLMs can generate choices conditional on experimental instructions and agent profiles. This makes them useful for hypothesis exploration, material checks, and sensitivity analysis. Yet an LLM does not face real income, risk, or opportunity costs; its output is model behavior before it is evidence about human behavior.

Validation should progress from treatment-effect direction and group means to full distributions, individual prediction, cross-task consistency, and robustness to prompts and model versions.

AI and Economic Text Analysis: Information Representation

Dictionary and topic models are transparent but struggle with negation, polysemy, and domain context. Supervised learning and embeddings improve semantic representation but depend on labels and training samples. Pretrained models and LLMs add contextual extraction while introducing prompt and version sensitivity.

The output is an economic measurement variable. Its meaning should be established with domain annotations, human review, and validation against external outcomes.

AI for Causal Policy Evaluation: Causal Estimation

Double machine learning uses flexible models for high-dimensional nuisance functions and combines orthogonal scores with cross-fitting to reduce regularization and overfitting bias. Causal forests help characterize conditional average treatment effects.

These methods improve estimation under an identification strategy; they do not create identification. Unobserved confounding, reverse causality, and policy-induced behavioral change remain design problems.

AI and Economic Time-Series Forecasting: Dynamic Forecasting

Deep time-series models and foundation models can represent nonlinearity, long dependence, and cross-dataset transfer. Economic evaluation must preserve temporal order through rolling or expanding-window out-of-sample tests.

Average error is insufficient for macroeconomic and financial risk. Calibration, quantile performance, extreme scenarios, data revisions, and stability after institutional change also matter.

Combining the four capabilities

A policy study may use text models to construct variables, causal methods to estimate effects, and simulations calibrated with those estimates. These stages can be connected, but each retains its own error source and validation requirement.