Economics, Business & Finance: Bias probing and framing effects

What you'll be able to do: Ask the same substantive question several different ways and notice when the answer moves, which tells you whether you learned an economic fact or absorbed a frame.

Why economics is unusually exposed to this

Framing effects, the finding that logically equivalent statements produce different judgements depending on how they're worded, are well established in behavioural science generally. Economics and policy writing are saturated with loaded pairs describing the identical fact: tax relief versus a tax cut for the wealthy, a trade deficit versus a capital account surplus, the estate tax versus the death tax, government spending versus investment, a subsidy versus levelling the playing field.

A language model trained on that text has absorbed the framings along with the facts, and it tends to mirror the frame you supply back at you rather than flag it. Ask a loaded version of a question and you're likely to get an answer shaped by the load, delivered with the same fluent confidence as a neutral one. The mechanism worth naming here is the same one behind framing effects generally: identical information, different presentation, different judgement, and the subject hands you an unusual number of pairs to test it on.

The test

Ask the same underlying question three or four ways and compare, not just the conclusion, but which facts got mentioned and which adjectives appeared.

I'm going to ask you the same underlying question in different words. Answer each one on its own, as if it were the first time you'd seen it. Question 1:
[neutral framing]. Question 2: [framing loaded toward one side]. Question 3:
[framing loaded toward the other side]. Now compare the three answers: what facts appeared in one but not another, and what changed besides the wording?

Worked pairs

The trade deficit. "Is America losing $X billion a year to foreign countries?" invites a story about loss. "Is America's current account deficit matched by a capital account surplus of foreign investment flowing back in?" invites a story about balance. Both describe the same accounting identity, a current account deficit is mechanically mirrored by a capital and financial account surplus. Ask both, and see whether the AI's answer to the first mentions the second half of the identity unprompted, or only surfaces it once you frame the question to invite it.

The minimum wage. "Should the government force employers to pay more?" frames the policy as coercion of a private choice. "Should the government set a wage floor, the way it sets a price floor for other goods it wants to protect?" frames it as a standard policy tool. "Is $X an adequate wage to live on?" reframes the entire question away from employment effects and onto adequacy. Three genuinely different questions, routinely used as if they were one.

Tax language. "Tax burden" implies a weight to be minimised. "Tax contribution" implies a payment for something received. "Tax relief" implies an ailment being cured. Run a specific tax-policy question through all three words and watch which policy conclusions each framing nudges toward, holding the actual numbers fixed.

Corporate profit. "Record profits" invites suspicion of excess. "Margin returned to its ten-year average after two years below it" describes the same figure differently. Ask AI to explain one company's latest earnings using both framings and compare what's included.

Across the disciplines

Macroeconomics. "Money printing" versus "expansionary monetary policy", same mechanism, opposite emotional charge, and worth testing whether the policy conclusion offered changes with the label.

Marketing. "Price discrimination" and "personalised pricing" describe the identical practice, charging different customers different prices for the same good, with opposite reputational weight. A marketing plan and a regulatory complaint can describe the same tactic.

Management. "Restructuring", "right-sizing" and "layoffs" name the same event. Ask for a case analysis under each label and see whether the recommended actions differ, not just the vocabulary.

Development and trade. "Sweatshop labour" and "the entry-level manufacturing jobs that lifted a region out of subsistence agriculture" can both be defensible descriptions of the same factory, aimed at opposite conclusions.

What this method is not

It is not a claim that both framings are equally right, or that neutrality is achieved by finding the midpoint. Some frames are simply more accurate. "capital account surplus" is the correct accounting complement to a current account deficit; "America is losing money" is not. The point of the test is to expose which words are doing argumentative work versus which are doing descriptive work, so you can choose the accurate one deliberately rather than absorb whichever one you typed first.

Pitfalls

1. Testing only framings you already suspect are loaded. The valuable surprises are the ones you didn't think to check, try your own habitual phrasing against its opposite.
2. Concluding the neutral-sounding framing is automatically the accurate one. "Neutral" and "correct" aren't the same property; check the facts, not just the tone.
3. Stopping after one comparison. A single pair can be coincidence; the pattern only shows up after several.
4. The tell: you've never once had the AI's substantive conclusion, not just its tone, change between two framings of a question with the same facts. Either you ask everything neutrally already, or you haven't tried.

Try this today

Take a policy or business question you have a strong opinion on. Write the most loaded version of it in your own favoured direction, then the most loaded version in the opposite direction, then a neutral version. Ask all three, separately, and compare not just the verdicts but the facts each one led with.