Gains:
- Ability to use artificial intelligence to produce a search strategy, concept map and synthesis outline for an economics question
- Ability to recognize the risk of sources, fake citations and false findings that artificial intelligence can make up, and verify each reference from the original source
- Ability to transform findings into a balanced literature summary with opposing views, methodological differences and context boundaries.
No economic analysis is born in a vacuum. Before starting a policy question, a forecast, a report, ask "what is known about this issue, what is discussed?" It is necessary to answer the question. That's what a literature review is: an organized compilation of existing knowledge, findings, and debates on a topic. Artificial intelligence seems like a dazzling accelerator in this business — summarizing a topic, mapping concepts, sorting out opposing views. But this is where it harbors the most dangerous trap: Artificial intelligence can produce non-existent sources, fake authors, and fabricated findings with complete confidence. The immutable rule of this unit: Every reference, every finding, every name is verified from the original source; No unverified references enter the text.
What does artificial intelligence do well in the literature and what does it never do?
What it does well (helpful): Mapping a topic and its subheadings, suggesting key concepts and search terms to look for, summarizing and comparing texts you provide, framing “sides” of a debate, extracting the argument structure of a text.
Never to be trusted (dangerous): Producing real citations and credits. Artificial intelligence can make extremely realistic but completely fabricated sentences such as "Yılmaz (2019) found that the inflation-growth relationship is negative." The author may not exist, the article may not exist, or the actual article may say the opposite. This is called hallucination in the literature and is fatal to academic/professional reputation.
Attention: Just because an ID given by artificial intelligence "looks realistic" does not prove anything. The journal name may be correct, the year is reasonable, the author's name may be Turkish — and the article may never exist. You only accept that it exists when you find it in the actual database (journal site, DOI, library).
The right workflow: AI maps, you verify
- Clarify the question. What are you looking for? Scope, period, geography, method?
- Create a concept map. Ask the AI for topic subheadings, key concepts, and search terms — these are verifiable pointers, not sources.
- Make a real call. Search for these terms in real databases (academic search engines, journal archives, institutional publications).
- Summarize the found texts to artificial intelligence. Now you have the real text; AI can summarize and compare it.
- Synthesize and balance. Opposing findings, method differences, and context boundaries are presented together.
- Verify each attribution. Does each reference included in the text really exist, is its attribution correct, does it really contain the claimed finding?
Tip: Use AI as a “search strategist and summarizer,” not a “source finder.” You find and verify the source; AI helps you understand and compare the source you find.
Balanced synthesis: without taking sides
A good literature summary does not present a single finding as absolute truth. There are opposing findings on most questions in economics: one study finds a positive relationship and another does not. This is usually due to differences in method, period, country or data. Honest synthesis makes these differences visible: "Studies using method A find this, studies using method B find that; the difference probably stems from this assumption."
three mini cases
Case 1 — Fabricated attribution caught. One researcher called the AI "summarize 5 studies on the relationship between minimum wage and employment." The AI gave five decent tags. The researcher searched for each in an academic database: three were real, two were not at all—the author-year-journal combination was completely made up. He used the real three and took out two. Without verification, two ghost sources would have made it into the article.
Case 2 — Adverse finding. The AI cited a real article with the correct byline, but said "this study found X." The researcher opened the original paper: it actually found the opposite. The artificial intelligence attributed the wrong finding to the real source. Even if the imprint is true, the lesson that it cannot be used without verifying the claim was reinforced.
Case 3 — Balanced synthesis. One analyst collected literature on the impact of a tax policy. In the first draft, the AI drew a one-sided conclusion (“impact positive”). The analyst also added opposing studies and showed that the difference comes from the method (short-term vs long-term data). The result was a balanced summary of “conditional”; The policy rating has become more reliable.
validation table
AI output
Verification step
Citation/imprint
Does it actually exist in the database/with DOI?
"This study found this"
Is this really the finding in the original text?
Numerical result
Does it match the table/summary in the article?
"The general acceptance is this way"
Have opposing findings been suppressed?
Search terms
Does it give meaningful results in real search?
Four copyable templates
1) Concept map and search strategy:
I will investigate the following economic question: [question]. Give me the subheadings of the subject, key concepts and search terms (Turkish + English) that I can use in the academic database. PRODUCING SOURCE/CITATION; just tell me where to look and in what terms.
2) Summarizing the text provided:
Below is the text/summary of an actual article. Summarize this: research question, method, data, main finding, limitations. Do not add anything that is NOT in the text; Mark the part you are not sure of as "unclear in the text".[text]
3) Comparative synthesis:
Compare the summaries of these 4 real studies I gave you: which ones show similar findings, which show contradictory findings, and what (method, period, country, data) is the difference likely to account for? Do not impose a one-sided conclusion; make opposing findings equally visible.
4) Citation verification list:
List all references in this draft. For each, state three things I need to verify: (1) does the source actually exist, (2) is the attribution accurate, (3) does it contain the finding attributed to it? Don't consider any of them as "certain" until I verify them.
Weak prompt / Strong prompt
Weak prompt:
Write a literature summary on this subject and give sources.
AI produces text that looks realistic but is filled with partially made-up references; If you do not verify, fake sources will enter the report.
Powerful prompt:
I will research the minimum wage-employment relationship. First, give me the subheadings and search terms (PRODUCING sources/citations). I will find the real works and bring their texts to you; You just summarize and compare what I brought. Present opposing findings equally and discuss whether the difference comes from the method or the period. Do not add any numbers or findings that are not in the text.
Common mistakes
- Asking for sources/citations directly from artificial intelligence and using them without verification. Risk of hallucinations.
- Mistaking realistic imprints for real. Appearance is not proof.
- Not noticing the attribution of the wrong finding to the correct tag. There is a source, but the claim may be false.
- One-way synthesis. Concealing contrary findings.
- Bypassing the method/context difference. Comparing apples and pears.
- Number leakage into the abstract that is not in the text. "Stuffing" of artificial intelligence.
In summary
In literature review, AI is a powerful search strategist and summarizer but an unreliable source generator. The right workflow: AI maps concepts and search terms, you find and verify actual sources, then AI summarizes and compares your findings. Every reference and finding is confirmed from the original source; The synthesis honestly reflects opposing views and methodological differences.
Application task
Choose an economics question. Get concept map and search terms from AI with template 1 (without requesting sources). Find at least three real studies in a real academic database with these terms. Summarize and compare them with templates 2 and 3. Then, in a separate attempt, ask the artificial intelligence for a direct citation and search the database for the citations it gives to determine how many are real and how many are fabricated and note them down.
checklist
- [ ] I used AI as a search strategist/summarizer, not a resource generator.
- [ ] I found the real sources myself and verified them in the database.
- [ ] I have confirmed that each reference actually exists and that its attribution is correct.
- [ ] I checked the attributed finding from the original text.
- [ ] I presented opposing findings and methodological differences in a balanced manner.
- [ ] I checked that no numbers/findings that were not in the text were leaked into the summary.
- [ ] I did not include any unverified references in the text.