The Most Common AI Mistakes Economistss Make

When you ask ChatGPT or Claude to suggest variables for a model, the AI returns statistically sensible lists without knowing your theory of mechanism. You then run the model and report coefficients as if they represent economic relationships you understand.

The fix

Write down the causal chain you believe exists before you speak to any AI tool, then ask the AI to critique that chain rather than generate one.

Stata AI and Bloomberg terminals generate confidence bands around predictions using historical variance. Economistss then present these as if they capture structural breaks, policy shifts, or regime changes that historical data cannot predict.

The fix

Always separate statistical precision from genuine uncertainty by asking what economic events the model cannot see coming, then widen your confidence interval manually to reflect those blind spots.

Perplexity and Claude summarise papers quickly, but you may cite their paraphrase of a finding without checking the original paper's actual result, sample size, or limitations. Policy recommendations then rest on summaries of summaries.

The fix

Read the abstract and methodology section of any paper before you use it to support a policy claim, regardless of how an AI tool described it.

When you ask Claude to build a small open economy model, the AI draws on decades of textbook examples with specific assumptions about capital mobility and wage rigidity. You may not realise these assumptions are there until your results contradict your country's actual behaviour.

The fix

After AI generates a model specification, ask it to state every assumption it made, then cross-check each one against your institutional knowledge of the economy you study.

Bloomberg AI and similar tools identify historical correlations between variables very fast. You then present a correlation as economic insight without testing whether the relationship holds in different time periods or whether it reflects simultaneity bias.

The fix

When AI shows you a strong correlation, stop and ask whether a shock to one variable would actually cause movement in the other, or whether both respond to a third cause.

You ask ChatGPT to help you assess your forecasting model, and it shows you how well the model fits historical data. You then present this fit as evidence the model works, when you have actually measured overfitting.

The fix

Always hold out a test period before you touch your model, then evaluate only on that held-out data, and document this split in your methodology.

Standard errors from AI-assisted estimation assume your model is correct. They do not account for omitted variables, functional form errors, or structural breaks. You report tight confidence intervals around parameters that may be deeply wrong.

The fix

Report standard errors from your model, then separately list three ways your model might be fundamentally wrong, and note that true parameter uncertainty is larger than the statistics show.

When you ask an AI tool to build a cross-country dataset, it may select countries based on data availability rather than your research question. You then run regressions on a biased sample and claim findings apply to economies very different from those in your data.

The fix

Before analysis, write down which economies your question is actually about, and justify why the countries you selected or excluded belong in or out of that group.

AI tools generate multi-step-ahead forecasts using iterated predictions, where each forecast feeds into the next. Errors compound, but the AI output looks smooth. You report a five-year forecast without disclosing that uncertainty grows exponentially as you move forward.

The fix

For any forecast more than one or two steps ahead, explicitly show how forecast error bands widen, and state the horizon at which your forecast becomes too uncertain to inform policy.

You start with a testable hypothesis about what policy works. The AI-assisted model finds a different effect. You then reverse your conclusion to match the model rather than investigating why the model disagreed with your prior reasoning.

The fix

When your model contradicts your economic intuition, treat it as a signal to audit the model, not as reason to abandon your intuition.

Claude or ChatGPT generates clear prose explaining your regression results, including interpretations of coefficients. You paste this language into your report without mentally translating it back to the economic meaning of the numbers.

The fix

Always write your own one-sentence interpretation of every coefficient before you read what the AI wrote, then check whether the AI interpretation matches yours.

You ask an AI tool to run your model with different specifications or samples. The tool reports that findings are robust. You then skip the work of thinking through which alternative specifications would actually test your core assumption.

The fix

Before you run sensitivity analysis, list the three changes in assumptions that would most undermine your main conclusion, and test exactly those.

Perplexity or Claude suggest you include additional variables to improve R-squared or reduce residuals. You add them and the model fits better, but you cannot explain why those variables matter economically.

The fix

Only add a variable to your model if you can explain why economic theory says it should be there, regardless of what the AI or the fit statistics say.