For Economistss
Cognitive Sovereignty Checklist for Economistss
When you use AI tools like Claude or ChatGPT to build economic models, the outputs can feel authoritative. You risk embedding untested assumptions into your models and presenting forecasts with false precision. Your judgement about economic theory, causal mechanisms, and policy trade-offs must remain your own, not derived from what the AI suggests.
These are suggestions. Take what fits, leave the rest.
0 / 20 applicable
Tap once to check, again to mark N/A, again to reset.
Test Model Assumptions Before You Trust Them
Write out the theoretical mechanism before asking AI to model itbeginner
Specify the causal chain you expect: how variable A moves variable B, and why. Only then ask the AI to help build it. This stops you accepting assumptions the model embedded without your knowledge.
Verify each coefficient makes economic sensebeginner
When AI generates a model output, check whether the sign and magnitude align with economic theory and prior evidence. A negative relationship between unemployment and wage growth should match established labour economics, not just fit your data.
Run the model on a subset of years and check stabilityintermediate
Rebuild the same model using different time periods. If coefficients swing wildly between periods, the relationship is unstable. AI will not flag this unless you ask it to test across subsamples.
Ask the AI to show you the five most questionable assumptions it madeintermediate
Force explicitness. The model made choices about functional form, lag structure, or variable treatment. Make the AI list them. Then decide whether you agree or need to change them.
Compare AI-assisted models to simpler alternatives you built yourselfadvanced
Build a basic model the old way, by hand. Then compare it to the AI version. If they diverge sharply, investigate why. The AI version may be overfitted to noise rather than signal.
Document why you rejected any AI-suggested variables or specificationsintermediate
When AI proposes an additional regressor or functional form and you say no, record your reasoning. This creates accountability and helps you spot when you are avoiding complexity for the wrong reasons.
Challenge multicollinearity before accepting correlated predictorsadvanced
AI models often include multiple variables that move together. This inflates apparent precision. Ask the AI to calculate variance inflation factors and force you to justify including collinear variables on theoretical grounds alone.
Reclaim Forecasting Judgement from False Precision
Reject AI confidence intervals that are narrower than your theoretical uncertainty allowsbeginner
AI calculates uncertainty from historical data variance. But genuine economic uncertainty about policy shocks, structural breaks, or black swan events lives outside the data. Your judgement about these must widen the intervals, not the model output.
Produce three separate point forecasts using different models or assumptionsintermediate
Use the AI to build one forecast, then build another with different lag lengths or variable choices, and a third using a simpler method. Present all three to decision makers. This shows them the range of reasonable outcomes rather than false certainty from one AI model.
List the conditions that would make your forecast wildly wrongbeginner
Before publishing a forecast, write down what policy changes, external shocks, or behavioural shifts would break the model. If you cannot list them, you do not understand your model's limits. AI will not do this for you.
Check what past forecast errors looked like for the variable you are predictingintermediate
Pull historical forecasts of the same variable from your organisation or others. How often and how badly were they wrong? Use that empirical track record to set realistic confidence intervals, not the AI model's optimistic bands.
Set a forecast review date and document what you expected to seebeginner
Write down before the forecast period what would count as 'the forecast was right' versus 'the forecast was wrong'. This prevents you reinterpreting results after the fact or accepting vague vindication.
Ask the AI to explain which historical periods most resemble the forecast periodintermediate
The model learns from all past data equally. But some years are more relevant than others. Force the AI to identify analogues. Then decide whether those analogues support the forecast.
Defend Economic Theory Against Data Fitting
Insist on a causal story before accepting any policy recommendationbeginner
When an AI model suggests a policy lever, ask it to explain the causal mechanism. Not the correlation in the data, but the economic logic for why this policy changes behaviour. If the story is weak, the recommendation is weak.
Use AI to search for evidence against your preferred policy, not just for itbeginner
Ask the AI: what are the three strongest critiques of this policy in the literature? What populations does it harm? What unintended consequences have researchers found? Let AI help you stress test, not just support.
Distinguish between statistically significant and economically meaningfulbeginner
AI can tell you a coefficient is significant at p less than 0.05. But is the effect size large enough to matter for policy? A 0.001 percentage point change in inflation is not economically meaningful. You must make this judgement.
Require yourself to write the policy memo before showing it to AI for editingintermediate
Draft your recommendations first, in your own words. Then ask AI to tighten the language. Do not ask AI to draft the substance. The argument must come from your economic reasoning, not generated text.
Test whether your key finding survives when you remove the most recent dataintermediate
Recalculate your model without the last two or three years. Does the result hold? Recent data can create the illusion of a strong effect that is really just temporary. AI will reproduce whatever pattern exists in the data you give it.
Identify which variables in your model are measured with greatest erroradvanced
Some economic variables are precisely observed. Others are estimates or constructed from surveys. Ask the AI which inputs are least reliable. Then ask whether your conclusion depends on those unreliable inputs.
Compare your AI model's predictions to the predictions from published economic models by othersadvanced
Search the literature for forecasts or analyses of the same question. How do the AI results compare? If they diverge wildly, find out why. You may have missed something, or the AI may have.
Five things worth remembering
- Never present an AI-generated confidence interval without first widening it to account for genuine economic uncertainty that the model cannot capture.
- When an AI tool proposes a variable or specification, ask it to also show you what the results look like without it. Compare the two versions side by side.
- Document every choice you make against an AI recommendation. This creates a record of where your judgement overrode the model, which helps you learn whether you were right.
- Use AI for calculation and literature search, not for deciding what the data means. The searching and checking can be automated. The judgement cannot.
- Before a forecast enters policy circulation, present it to a colleague who built a competing forecast independently. If they diverge, investigate why before leadership sees either one.
Related reads
Common questions
Should economists write out the theoretical mechanism before asking ai to model it?
Specify the causal chain you expect: how variable A moves variable B, and why. Only then ask the AI to help build it. This stops you accepting assumptions the model embedded without your knowledge.
Should economists verify each coefficient makes economic sense?
When AI generates a model output, check whether the sign and magnitude align with economic theory and prior evidence. A negative relationship between unemployment and wage growth should match established labour economics, not just fit your data.
Should economists run the model on a subset of years and check stability?
Rebuild the same model using different time periods. If coefficients swing wildly between periods, the relationship is unstable. AI will not flag this unless you ask it to test across subsamples.