AI for economic research

A literature review is where most research begins and where a surprising amount of it goes wrong. Read too little and you reinvent a wheel or miss the identification problem that sank an earlier paper; read carelessly and you misreport what a study actually found. AI can compress the weeks of discovery and summarising into days — but it introduces a new failure mode of its own, and the whole discipline of using it well is the discipline of verification. This module is about doing the work faster without lowering the standard.

Start with the question, not the search box

The most common mistake is searching before you have a question. 'Mobile money in Africa' is a topic; it is not researchable. A researchable question names an outcome, a treatment or variable of interest, a population, and — for empirical work — the kind of variation that could credibly identify an effect. AI is genuinely good at this refining step: it will interrogate a fuzzy interest, propose sharper versions, and surface the scope decisions you have not made yet. It proposes; you decide.

I am an economics student in Nairobi interested in mobile money and household finances. Help me turn this into a researchable question.

Propose five precise, researchable versions. For each one, state:
- the OUTCOME (and how it would be measured),
- the TREATMENT or variable of interest,
- the POPULATION and setting,
- the kind of IDENTIFICATION a credible causal answer would need (RCT, difference-in-differences, IV, RD, panel fixed effects).

End by listing the three biggest feasibility risks (data availability, endogeneity, external validity).

Finding the literature — and the honest limits of AI search

There is now a spread of tools, each good at a different job. Used together they are powerful; used naively they mislead. And be clear about the honest limits: a general chat model's knowledge is frozen at its training cut-off and skewed towards English-language, open-web, and heavily-cited work — which means recent papers, paywalled journals, working papers, and African grey literature (Central Bank of Kenya notes, KNBS reports, World Bank country studies) are exactly what it is most likely to miss or misremember.

• ChatGPT (OpenAI) and Claude (Anthropic): general reasoning assistants. Excellent at restructuring, summarising and critiquing text you give them; unreliable as a source of citations from memory. Best pointed at documents you supply, not used as a search engine.
• Perplexity: a search-first assistant that runs live web queries and answers with inline links to the pages it used. The citations are real URLs you can click — but they point to web pages, not always peer-reviewed papers, so you still judge the source's authority.
• Semantic Scholar (Allen Institute for AI): a free academic search engine and citation graph over 200M+ papers. Use it to find work, follow citations forwards and backwards, and confirm a paper actually exists — the direct antidote to fabricated references.
• Elicit: a research assistant built on academic databases that finds papers and extracts a structured table (sample, method, outcome) across many studies at once. A strong first pass; it does not excuse you from reading the papers.
• Google NotebookLM: answers only from the sources YOU upload — your own PDFs — and cites the exact passage. Ideal for interrogating a set of papers you have already collected, because it will not wander off into invented facts.

Reading a paper through AI: methodology, data, identification, findings

For an empirical economics paper, a useful summary is not a paragraph of prose — it is four specific things extracted cleanly: the DATA (sample, unit, country, period), the IDENTIFICATION STRATEGY (the exact source of variation used to claim a causal effect, and the assumption it rests on), the FINDINGS (estimates with magnitudes, signs and units), and the LIMITATIONS the authors themselves concede. Ask for those explicitly, ask for the page or section behind each, and you get something you can actually check.

You are helping me review an empirical economics paper. The full text is below. Use ONLY the text provided — no outside knowledge. Where the paper does not say, write "not stated".

Extract:
1. RESEARCH QUESTION - the precise question the paper answers, in one sentence.
2. DATA - dataset(s), sample size, unit of observation, country/region, time period.
3. IDENTIFICATION STRATEGY - how the authors claim a causal (not merely correlational) effect: the exact design (RCT, difference-in-differences, IV, regression discontinuity, panel fixed effects) and the identifying assumption it relies on.
4. MAIN FINDINGS - key estimates with magnitudes, signs and units.
5. STATED LIMITATIONS - threats to validity the authors acknowledge.

For every item, give the section or page where you found it so I can verify.

PAPER:
[paste full text, or attach the PDF]

Two things make that prompt work. It forbids outside knowledge and demands a section or page reference, so you can check each claim in seconds. And it works far better when the model can actually see the paper — paste the full text, or better, upload the PDF to a grounded tool like NotebookLM. Never ask a chat model to summarise a paper 'from memory'; that is precisely when it invents a methodology the authors never used. In economics the identification strategy is where memory-based summaries fail hardest: a model will cheerfully call a difference-in-differences design a randomised trial, or drop the instrument entirely, and your review inherits the error.

Comparing competing schools on a question

AI is genuinely useful for mapping a debate. Ask it to lay out the competing explanations for a phenomenon — the schools, their assumptions, their evidence, and where they part company — and you get a fast scaffold for the theoretical-framework section of a review. On mobile money and savings, for instance, you can have it contrast the transaction-cost view (mobile money lowers the cost of moving and storing value, so households save more and share risk better) with more sceptical readings (gains accrue mainly to agents and wealthier users, or the effect merely displaces informal savings groups). But watch two failure modes. Models are trained to be agreeable, so they drift towards false balance — dressing a fringe position as the equal of a settled one — and they over-generalise from one famous study to a whole continent. Treat the map as a set of hypotheses to test against the actual papers, never as a finding.

Building an annotated bibliography

An annotated bibliography is a list of sources, each with a short paragraph on what the study did and why it matters to your question. This is where an AI workflow earns its keep — but only in a specific order. Collect the real PDFs first (Semantic Scholar or Elicit to find them, then download from the journal or an open repository). Load them into NotebookLM, or feed each in turn to Claude with the methodology prompt above, and have the model draft a four-to-five-sentence annotation per paper: question, data, identification, finding, relevance. Then edit every annotation against the paper yourself, and store the reference in a manager like Zotero, whose metadata comes from the publisher rather than the chatbot. The AI drafts; the databases and the PDFs are the source of truth.