Gains:
- Being able to distinguish where artificial intelligence saves real time in economic analysis (data compilation, draft, code, summary) and where (forecast responsibility, policy decision, official figure) the decision is left to the competent economist, according to the task risk level
- Ability to apply a discipline that verifies each artificial intelligence output through the steps of connecting it to the original source (TURKSTAT/EVDS/IMF), recalculating it and passing it through the filter of economic common sense.
- Ability to recognize the risks of fake numbers (hallucinations), old data cut-off dates and confidential data leakage, and acquire the habit of choosing safe tools and resources.
An economist's desk is actually a data and text factory: you interpret the inflation data from TURKSTAT in the morning, write a policy note for a development agency in the afternoon, clean a time series and try to make a forecast in the evening. Much of this work is mechanical, not mental: downloading data, converting units, editing tables, summarizing long reports, writing code. Artificial intelligence (AI — large language models and auxiliary tools) alleviates exactly this mechanical burden. But let's start with a sentence that we will repeat throughout this module: Artificial intelligence is the economist's assistant, not the expert who replaces him. The responsibility for an inflation forecast, a monetary policy proposal or a publicly available figure always lies with the person and institution that signed it.
Think of this unit like a “user manual”: here we'll lay out where to put AI, where never to put it, and how to validate each output. The other ten units will put flesh on this skeleton.
What does artificial intelligence accelerate in economics and what does it not accelerate?
Let's divide the tasks into three boxes according to risk level. This distinction is the compass of the entire module.
Green box — AI is powerful, adds speed (still reviewed). Writing data download code, suggesting steps to cleanse data, summarizing a long report or article, generating code for a chart, simplifying and formatting text, explaining terms, translation, draft policy note skeleton. Here the cost of error is low and errors are easily seen.
Yellow box — AI helps but human verifies. Building a time series model, regression interpretation, scenario calculation, indicator interpretation, literature synthesis. Here the AI produces a blueprint; But numbers, assumptions and economic logic don't go anywhere without being checked line by line.
Red box — Responsibility lies with the human, AI cannot decide. Final approval of the official inflation/growth figure, ownership of the monetary or fiscal policy proposal, signature of a forecast to be made public, an opinion that turns into an investment/credit decision. Here AI only produces input; Judgment, responsibility and signature belong to the person.
Tip: If you're not sure which box to put a task in, ask this question: "Who gets hurt if this output turns out to be wrong, and will the error be easily noticed?" If the damage is great and the error is hidden, that task is close to red.
Three fundamental limits of artificial intelligence
In order to use it safely in economics, one must first know where it stumbles.
1) Fabrication (hallucination). The big language model sometimes makes up a number or source it does not know, in confident language. The sentence "2023 current account deficit was $45.2 billion" seems true but may be completely wrong. In economics, this is fatal: an incorrect figure poisons a report, a forecast, a decision.
2) Data cutoff date. The training data of the model stops at a certain date. The latest inflation, the latest interest rate decision, and yesterday's exchange rate may be excluded from this information. The AI may actually give you a number that is several months or years old as "latest data".
3) Causation blindness and lack of context. AI captures patterns in language; does not "understand" economic causality. He makes sentences such as "When interest rates increase, inflation decreases" fluently, but he cannot evaluate under which conditions they are valid or in which country the opposite occurred in which period. It cannot see structural breaks (change of pattern with a crisis).
Attention: Just because a sentence is fluent and confident does not mean it is correct. Artificial intelligence at its most convincing can be at its most wrong.
Verification discipline: SOURCE–ACCOUNT–FIlter
The heart of this module is a single habit. Take each AI output through three steps:
- SOURCE. Every number and factual claim is linked to the original source. Inflation from TURKSTAT, exchange rate and balance of payments from CBRT, global projection from IMF. You reference the official publication, not the figure given by AI.
- BILL. You manually recalculate each derived number at least once. Are the growth rate, real value and index derived with the correct formula? Check out a period on paper.
- SIEVE. You pass the result through economic common sense: Is this size reasonable? Is the sign correct? In historical range? Is there any contradiction? An unreasonable result is often a signal of error.
three mini cases
Case 1 — Capturing the fake figure. A development agency analyst told YZ to "make a table of 2023 export figures on a provincial basis." AI gave a decent picture. In the SOURCE step, the analyst looked at TUIK foreign trade statistics and saw that the figures for three provinces were twice the actual data; The AI had filled in the missing information with “reasonable looking” numbers. The analyst threw the table away and pulled the data directly from TURKSTAT. The fabricated figure was caught before it was included in the report.
Case 2 — Outdated data trap. An economic journalist asked YZ, "What is the current policy rate?" he asked; The answer he received was the rate of eight months ago. The journalist, who has a habit of ACCOUNT/SOURCE, looked at the last MPK (Monetary Policy Committee) decision of the CBRT and used the correct rate. If he hadn't asked, wrong interest would have gone to the publication.
Case 3 — Economic filter. A student-researcher had the AI calculate a real wage series. The result said that real wages increased by 140 percent in one year. The FILTER step came into play: this size was unrealistic. In the code, the AI had used the deflator (price corrective index) with the wrong base. The error was found, the formula was corrected, and the result fell within the reasonable range.
Four copyable templates
1) System prompt defining roles and boundaries:
Your role: assistant analysis assistant to a senior economist. Rules: (1) Don't make up any numbers I don't give you. If you don't know, say "you need to verify this data". (2) Specify from which official source (TURKSTAT/CBRT/IMF, etc.) it should be confirmed for each number you provide. (3) State the condition and uncertainty clearly when making a causality claim. If you understand, confirm in one sentence.
2) Source-confirmation request:
Analyze the following claim: "[claim]". List which official/primary source (institution, publication name, series) I should look at to verify this. Identify which definition and period each number in the claim belongs to. Don't give numbers from your own memory.
3) Prompt to filter output:
Check the following analysis result for economic consistency: does the sign make sense, is the magnitude within the historical range, are the units consistent, is there a hidden assumption? Mark anything you find suspicious with a "verify" tag.
4) Task risk classification request:
I will do the following task: "[quest]". Classify it as green (mechanical, low risk), yellow (draft, human verifies) or red (decision/responsibility human) and explain why in one sentence.
Weak prompt / Strong prompt
Weak prompt:
Write an analysis about the Turkish economy.
Contextless and unlimited; AI produces a generic, unsourced, unverifiable text.
Powerful prompt:
Your role: analysis assistant. I have the 2019-2024 annual CPI and GDP growth data from TURKSTAT (below). Write a 3-paragraph draft comment based on this data; Take each number from the table I gave you, do not add numbers from outside; Where you are unclear, write "must be verified". Finally list which points I need to manually check.[table]
This prompt gives the role, resource, constraint, and verification expectation together; The output becomes auditable.
Common mistakes
- Treating the number given by the AI as a source. AI is not a resource; TURKSTAT/CBRT/IMF is the source.
- Mistaking fluency for accuracy. A well-written sentence can be wrong.
- Bypassing the task box. Passing the red box job as quickly as the green is the most expensive mistake.
- Forgetting the data cutoff date. Asking for "current" and getting the old number.
- Indiscriminately sticking confidential/market-sensitive data into the tool. A privacy violation that we will see in the next unit.
- Skip the manual calculation step. Using the derived number without checking it at all.
In summary
Artificial intelligence is a powerful assistant that significantly reduces the mechanical burden of the economist; but the responsibility for forecasting, policy and official figures remains with man. Sort tasks into green-yellow-red boxes, keep in mind the three fundamental limits of AI (fabrication, data cutoff date, causality blindness), and pass each output through the SOURCE–ACCOUNT–FIlter discipline. These three steps are the safety belt for the rest of the module.
Application task
Choose a real week from your own job and list 10 economic tasks you did that week (e.g. "interpret inflation data", "prepare a provincial export table", "produce a report summary"). Classify each task as green/yellow/red. Then select a task in the green box and have the AI do it using templates 1 and 3 above, and inspect the output with SOURCE–ACCOUNT–FIlter and note at least one error you find.
checklist
- [ ] I placed the task correctly in the green/yellow/red box.
- [ ] I confirmed every number given by the artificial intelligence from the main official source.
- [ ] I recalculated at least one derived number by hand.
- [ ] I filtered the result through economic common sense (sign, magnitude, unit).
- [ ] I checked the data cutoff date and recency risk.
- [ ] I did not use confidential/sensitive data or used secure/approved tools.
- [ ] I have claimed ownership of the output, maintaining that I have final judgment and signature.