AI and the Future of Econometrics: A Working Researcher's Perspective

I run regressions on 400,000-observation panels. I spend half my week writing Stata code and the other half debugging it. I’ve used ChatGPT extensively for my research workflow. Here’s what I think about where AI and econometrics are heading.

What AI Already Does Well

• Boilerplate code. Data cleaning, variable construction, reshape operations. ChatGPT handles these at ~85% accuracy. You still need to check the output, but it saves time.
• Documentation lookup. “What are the options for ivregress?” — faster than reading the manual. The answers are usually correct for well-documented commands.
• First drafts. Generating a skeleton .do file for a standard analysis. You’ll rewrite most of it, but the structure helps.
• Explanations. “Explain the Bacon decomposition in simple terms.” LLMs are excellent at this when the concept is well-represented in the training data.

What AI Gets Wrong — And Why It Matters

• Methodological selection. AI doesn’t choose between DiD estimators based on your data structure. It generates whatever its training data suggests is most common. Given that TWFE is in more textbooks than Callaway-Sant’Anna, it defaults to the wrong estimator for staggered designs.
• Diagnostic completeness. AI generates the regression but not the 5 post-estimation commands that determine whether the regression means anything. This is the most dangerous failure mode — the researcher gets a number and trusts it.
• Execution. AI generates code but can’t run it. You have to copy-paste into Stata, fix the errors, and iterate. This breaks the feedback loop that makes AI useful.

Five Predictions for the Next Five Years

1. Execution-first AI will replace chat-first AI for research

The dominant paradigm will shift from “ask AI for code, then run it yourself” to “tell AI what you want, AI runs it and shows you validated results.” The copy-paste workflow will look as primitive as manually computing regression coefficients.

2. Domain-specific beats general-purpose

General-purpose LLMs will continue to improve at code generation. But for statistical computing, domain-specific tools that encode methodological knowledge will outperform them. The reason: the long tail of statistical methods is too specialized for a general model to handle reliably.

3. Reproducibility becomes automated

AI-assisted workflows will generate replication packages as a byproduct. The prompt, the code, the output, and the execution log will be captured automatically. Journals will start requiring AI-generated audit trails alongside traditional replication packages.

4. The methods gap narrows

Currently, adopting a new econometric method (e.g., switching from TWFE to Callaway-Sant’Anna) requires learning new syntax, reading papers, and debugging unfamiliar code. AI lowers this barrier. Researchers will adopt state-of-the-art methods faster because the implementation cost drops to near zero.

5. The value shifts to judgment

If AI can reliably generate code, run diagnostics, and produce tables, the researcher’s comparative advantage shifts entirely to judgment: which question to ask, which identification strategy is credible, how to interpret the results. The mechanical skills (debugging syntax, formatting tables) lose value. The intellectual skills (research design, causal reasoning) gain value.

This is a good thing. The best parts of research — thinking carefully about identification, defending your design to skeptics, interpreting results in light of prior knowledge — are the parts AI can’t do. And shouldn’t try.