AI for Research in Economics

Part 01 Foundations

The foundations — and the golden rules that come before any tool

What we mean by “AI” in this guide

“AI” is a broad word. For research purposes, three overlapping ideas matter:

• Large language models (LLMs) — systems like ChatGPT, Claude and Gemini that generate text (and increasingly code, tables and images) by predicting plausible continuations. They are pattern generators, not knowledge bases: brilliant at drafting, summarising and explaining, but capable of stating false things fluently.
• Retrieval-grounded AI — tools that first fetch real documents and then let an LLM answer only from those sources (e.g. NotebookLM, Elicit, Consensus). Grounding reduces, but does not eliminate, the risk of errors.
• Machine learning (ML) & code execution — the older statistical machinery (regression, clustering, classification) plus newer tools that write and run analysis code for you. This is where most quantitative economics work actually happens.

Where AI fits in the research lifecycle

AI is not one thing you bolt on at the end. It can assist at every stage — but the kind of help, and the level of caution, changes as you move from ideas to published claims.

Caution rises as you move right. A wrong word in a brainstorm costs nothing; a wrong number in your results section costs your credibility.

The golden rules (read these before anything else)

Those four rules sit underneath everything that follows. If a tool tempts you to break one of them to save time, that is exactly the moment to slow down.

Part 02 The toolkit

The AI toolkit for economists, by category

The market moves fast, so treat names and prices as a snapshot, not gospel. Always confirm the current free tier, price, and data policy on the official site before you rely on a tool. The categories below are stable even as individual products change.

① General-purpose AI assistants

Your everyday workhorses for explaining, drafting, brainstorming, coding help and quick analysis. They are generalists: flexible but not specialised, and they hallucinate, so verification is essential.

② Literature discovery & review

This is where AI saves economists the most time. These tools search across tens or hundreds of millions of papers, summarise, and extract structured data — most importantly, the good ones link every claim back to a source you can check.

③ Writing, editing & references

Economics writing is dense and precise. AI helps most with structure, clarity, grammar and formatting — and least with making substantive claims, which remain your job.

④ Quantitative analysis & econometrics

For most economists, this is the heart of the work. Two families of tools matter: those that help you write and run code, and “chat-with-your-data” tools that hide the code. Each has a sharp trade-off.

⑤ Qualitative analysis

Economics does qualitative work too — interviews, open survey responses, policy documents, archival text. Two paths exist: established CAQDAS packages with AI bolted on, and general LLMs used for thematic/content analysis.

⑥ Transcription & data collection

At a glance: which tool for which job

Part 03 By research type

AI across quantitative, qualitative & mixed-methods research

The same tools behave very differently depending on your paradigm. Here is how to use them — and where to be careful — in each.

Quantitative research with AI

Quantitative economics runs on data and models. AI can assist at each step, but the closer you get to the causal claim, the more your own judgment must dominate.

1. Data cleaning & wrangling. AI is genuinely strong here: reshaping, merging, handling missing values, fixing messy formats. Ask it to write the script rather than do it invisibly, so cleaning is reproducible.
2. Exploratory analysis. Summary statistics, distributions, correlations, first charts. Great for fast orientation — but confirm anything surprising.
3. Model building. AI can draft a regression, suggest specifications, and remind you of assumptions. It cannot decide your identification strategy for you — whether a coefficient is causal depends on your design, not the software.
4. Interpreting output. AI explains coefficients, standard errors, and diagnostics in plain language — excellent for learning. Always check it isn’t over-claiming significance or causation.
5. Visualisation. Quick, publication-style charts from a prompt; refine the code yourself for final figures.
6. Robustness & replication. Ask AI to propose robustness checks, alternative specifications, and placebo tests — then run them deliberately.

Qualitative research with AI

Qualitative work in economics — elite interviews, open-ended survey text, policy and archival documents — values interpretation and traceability. AI accelerates the mechanical parts; the meaning-making stays human.

1. Instrument design. Draft and critique interview guides or open-ended survey questions; ask AI to flag leading or ambiguous wording.
2. Transcription. Convert audio to text (mind the ethics rules from Part 2). Always proof-read transcripts; AI mishears names, jargon and accents.
3. Familiarisation & first-pass coding. AI can suggest an initial codebook or themes from your text. Treat these as hypotheses to test, not findings.
4. Thematic / content / framework analysis. Use CAQDAS (ATLAS.ti, MAXQDA) so each code links to a quotation, or build that link yourself if using a general LLM.
5. Synthesis. AI helps draft theme summaries and pull illustrative quotes — which you verify against the raw data before quoting.

Mixed-methods research with AI

Mixed methods combine quantitative and qualitative strands, and AI can help bridge them — provided you keep each strand’s standards intact.

• Design. AI can explain and help plan classic designs (convergent, explanatory-sequential, exploratory-sequential) and sketch how strands connect.
• Parallel analysis. Run the quantitative and qualitative workflows above, each with its own cautions.
• Integration. The hard, valuable step. AI can help build a joint display — a table placing quantitative results beside qualitative themes on the same question — and draft “meta-inferences” that you then scrutinise.
• Narrative. AI helps weave a coherent story across strands, which you fact-check against both datasets.

Part 04 Prompting

Prompting for economists — how to get useful, checkable output

The gap between a vague prompt and a precise one can be the difference between a helpful draft and something you have to throw away. Prompting isn’t magic — it’s clear briefing, the same skill you’d use delegating to a research assistant.

The five principles of good research prompting

1. State the role, method and constraints upfront

Tell the AI what it is (an economics research assistant), what method you’re using (e.g. OLS, thematic analysis, difference-in-differences), and what it must not do (e.g. invent citations, claim causation without a design). Front-loading constraints saves you from cleaning up overconfident output later.

2. Be specific about the output format

If you need a table with particular columns, say so. If you want code, specify the language and packages. If you need a paragraph of exactly 150 words for an abstract, say that. Ambiguous asks produce ambiguous output.

3. Provide context and data descriptions

AI cannot see what you haven’t given it. Describe your data structure, variable names and units — or upload the data directly where possible. The more the model knows about your setup, the less it guesses.

4. Ask for reasoning, not just answers

When you ask AI to “explain your steps” or “walk through the assumptions,” the output is far easier to verify — and you learn in the process. This is especially valuable for econometrics, where understanding the method matters more than the point estimate.

5. Iterate — prompt, check, refine

Your first prompt rarely produces the final output. Think of it as a conversation: get a draft, spot what’s wrong or missing, and ask a targeted follow-up. This cycle of prompt → check → refine is how you maintain quality control over AI-assisted work.

Prompts for specific research stages

Part 05 Ethics & integrity

Ethics, integrity & limitations you must understand

AI is a power tool, and like any power tool, it can hurt you if you don’t respect its limits. This part covers the non-negotiable ethical constraints and the technical limitations you’ll encounter.

Hallucinated citations: the #1 integrity risk

This problem deserves its own section because it can destroy a career before it starts. When you ask a general LLM to “cite sources,” it may produce entries that look impeccable — real-sounding authors, plausible journal names, correct-looking DOIs — for papers that do not exist. This is not occasional; it is routine. The model is generating text that looks like a citation, not retrieving one.

Retrieval-grounded tools (Elicit, Consensus, NotebookLM) are much safer here because they pull from real databases and link to the source. But even with these, click through and confirm — the summary might misrepresent the paper’s conclusion, or the paper might be a preprint that was later revised.

Disclosure: what your university and journals expect

Rules are evolving, but the direction is clear. Most major economics journals and universities now expect you to:

• Disclose that AI was used and how (e.g. “ChatGPT was used to draft interview question wordings, which were then revised by the researcher”).
• Take full responsibility for every claim, number and citation — AI cannot be listed as an author.
• Follow your specific institution’s AI policy, which may be stricter than the journal’s.

The safest practice: maintain a brief “AI use log” in your research notes — what tool, what task, what you did with its output — so you can answer any question later with full transparency.

Data privacy and confidentiality

Economics researchers handle data that ranges from public macro statistics to individual-level survey responses covered by ethics agreements. The rule is layered:

• Public, aggregated data (World Bank, FRED, census tables) — generally safe to use with any tool.
• Licensed data (firm-level COMPUSTAT, proprietary surveys) — check the licence; many prohibit sharing with third parties, which includes cloud AI tools.
• Human-subjects data (interview transcripts, individual survey responses) — governed by your IRB/ethics approval. If consent didn’t cover “processed by AI,” you may need to anonymise rigorously or use local-only tools.

Bias and representativeness

LLMs are trained on internet text, which over-represents English-language, Western, higher-income-country perspectives. This matters for economics:

• Literature coverage gaps. AI tools may under-surface research from non-English journals, working-paper series from developing-country institutions, or grey literature from government agencies.
• Interpretive bias. When coding qualitative data, the model may default to frames from the dominant literature and miss locally meaningful categories.
• Data analysis assumptions. AI may suggest methods common in North American/European empirical traditions without flagging that your data context may require different approaches.

Mitigation: supplement AI-discovered literature with manual searches in regional databases; cross-check qualitative themes with domain experts; and treat AI method suggestions as starting points for your own judgment, not final answers.

Reproducibility

Economics values reproducibility. AI introduces two new reproducibility challenges:

• Non-deterministic output. The same prompt may yield different results on different days, because models use randomness (“temperature”) in generation. Save both the prompt and the output.
• Model versioning. “ChatGPT” or “Claude” is not a fixed thing — models are updated. Record the model name/version and date of use in your methods section.

Part 06 Worked example

A worked end-to-end workflow: one project, many tools

Let’s walk through a realistic project so you can see how the pieces fit together. We’ll follow a hypothetical master’s student, Priya, writing a dissertation on the impact of broadband internet expansion on rural employment in India.

Stage 1 — Ideation & scoping

Priya starts with a broad question: “Does internet access improve jobs in rural areas?” She uses Claude to brainstorm more precise research questions and discuss what identification strategies others have used for similar topics. She settles on a difference-in-differences design comparing districts that received BharatNet fibre at different times.

Stage 2 — Literature review

Priya searches Semantic Scholar for “broadband internet rural employment developing countries” and saves 40 results to Zotero. She feeds the same query into Elicit and finds 15 more papers she missed, plus an auto-generated evidence table. She uses Research Rabbit to map the citation network around three seminal papers and discovers a working paper from NBER she hadn’t seen. She downloads 20 key PDFs, uploads them to NotebookLM, and asks it to compare their identification strategies side by side.

Stage 3 — Data assembly & cleaning

She downloads district-level employment data from the Periodic Labour Force Survey (PLFS) and BharatNet roll-out dates from government sources (public data — safe for cloud tools). She uses ChatGPT Data Analysis to write a Python script that merges the datasets by district code, creates treatment-timing variables, and generates summary statistics. She reads every line of the generated code, catches a merge-key mismatch, fixes it, and reruns.

Stage 4 — Quantitative analysis

Priya asks Claude to draft a commented Python script for a staggered diff-in-diff estimation (using the Callaway-Sant’Anna estimator from the did package in R). She pastes the draft into her own RStudio session, reviews it, runs it, and interprets the output. She then asks Claude to suggest three robustness checks — she runs a placebo test, an event-study plot, and a sensitivity analysis excluding the largest metro-adjacent districts.

Stage 5 — Qualitative component (mixed methods)

To explain the mechanism behind her quantitative result, Priya conducts 12 semi-structured interviews with rural entrepreneurs in treated districts. She transcribes the (Hindi/English) recordings using Whisper locally (she has ethics approval that requires data to stay on her machine). She uses MAXQDA to code the transcripts, with its AI Assist suggesting initial codes — which she reviews, merges and refines manually, anchoring every code to a quotation.

Stage 6 — Integration & write-up

Priya builds a joint display table (with Claude’s help) matching her quantitative findings to qualitative themes. She drafts each chapter in Google Docs and uses Writefull to check academic phrasing. She asks Claude to act as a “peer reviewer” and critique her introduction — it flags a gap in her framing that she fixes. Final references are managed entirely through Zotero.

Priya’s tool stack at a glance

Part 07 Wrap-up

FAQ, glossary & your final score

Frequently asked questions

Glossary of key terms

Your final score