Lesson 11 — Scientific Research with AI in Social Sciences

(literature mapping, evidence synthesis, qualitative coding, and Capstone Proposal v1)

Why this matters (motivation)¶

AI tools can speed up:

• finding papers,

• organizing themes,

• drafting summaries,

• and coding text.

But social-science research is not “speed-writing.” Quality comes from:

• careful question design,

• transparent evidence handling,

• and critical interpretation.

This week also connects to what you have learned so far: you now have multiple tool families (EDA/visualization, probability/simulation, regression, clustering/PCA, trees, text signals, and forecasting). The research challenge is choosing the right tool for the question—and justifying that choice with evidence and transparency.

A research workflow (the course version)¶

Today we focus on Steps 1–3 and practice a small piece of qualitative coding.

Part A — Research questions: from topic to question¶

Topic vs question¶

• Topic: “Remote work”

• Research question: “How is remote work associated with self-reported productivity and job satisfaction across industries?”

A good question has:

• a clear population/unit (who? what?)

• a clear outcome (what is measured?)

• a feasible strategy (what data or evidence?)

A quick “question tightening” template¶

• Population / setting: (e.g., Thai firms, ASEAN consumers, university students)

• Outcome: (e.g., sales growth, job satisfaction, adoption probability, inflation)

• Mechanism or channel (optional): (e.g., flexibility, monitoring, cost savings)

• Comparison: (e.g., before/after, high/low exposure, group comparisons)

• Feasibility: data available? time constraints? permissions?

Part B — Literature mapping with AI (transparent, not magical)¶

What literature mapping is (and is not)¶

• It is not: “ask AI for 10 papers and copy the summary.”

• It is: a structured method to discover themes, debates, and gaps.

Recommended tools (choose your level)¶

• Semantic Scholar (search + metadata; API optional)

• Connected Papers (visual map; optional)

• Zotero (organize citations; optional)

AI usage: drafting search strings, summarizing abstracts, proposing a theme taxonomy — but you must verify.

Part C — Evidence synthesis: writing “what we know”¶

A good synthesis does more than list papers. It organizes evidence by:

• methods (survey, experiment, panel regression, text analysis, forecasting),

• context (country, industry),

• findings (consistent vs mixed),

• limitations (data, measurement, identification),

• and open questions.

Part D — Qualitative coding with AI (and why we compare to manual coding)¶

Why qualitative coding matters¶

Interviews, open-ended survey responses, policy documents, and news text are common evidence sources.

Coding turns text into:

• themes,

• categories,

• structured analysis.

AI-assisted coding: promise and risk¶

Promise:

• speed,

• consistency,

• quick theme suggestions.

Risks:

• hallucinated themes,

• overconfident labels,

• bias/stereotypes,

• missing nuance (sarcasm, context).

A simple coding exercise (what we do today)¶

We code a small set of quotes using:

1. manual coding (small groups)

2. AI-assisted coding (structured prompt) Then we compare:

• agreements/disagreements,

• where AI overgeneralized,

• where humans disagreed (coding ambiguity).

Mini-lab (Google Colab)¶

In-class checkpoints (Literature mapping)¶

1. Choose a research question (provided list or your own).

2. Create a search query list (5–10 keywords/phrases).

3. Collect 10–15 candidate papers:

   • title + year + venue/source

   • abstract (if available)

4. Screen down to 6–8 “most relevant” papers using explicit criteria.

5. Group the final set into 2–4 themes and write a 150–200 word map.

In-class checkpoints (Qualitative coding)¶

6. Use a small set of interview/open-ended responses (provided).

7. Define a codebook with 4–6 codes (short definitions).

8. Code the same set manually (group work).

9. Run AI-assisted coding using your codebook and compare results.

10. Write 5–7 lines:

• where AI matched humans,

• where it differed,

• what you would do next to improve reliability.

Capstone Proposal v1 (due after class)¶

Submission format: PDF or Markdown export via LMS. Include a short prompt/workflow log if AI tools were used.

AI check (responsible use for research)¶

Good prompt examples

• “Propose search terms for studies on remote work and job satisfaction across industries.”

• “Given these abstracts, group them into themes and explain your grouping logic.”

• “Here is my codebook. Apply these codes to the following quotes and flag ambiguous cases.”

Bad prompt examples

• “Write my literature review with citations” (risk of fake citations and shallow synthesis)

• “Decide the themes and final conclusions for me” (outsourcing research judgment)

Review questions (quiz / reflection)¶

1. What makes a research question “answerable” rather than just a broad topic?

2. What are two risks of AI-assisted literature summaries?

3. Why do we compare AI coding to manual coding?

4. What is one transparency practice you will use in your capstone?