Module 02: AI-Assisted Research — Literature Review, Synthesis, Citation Management & Academic Integrity

"The literature is a conversation. Your job is to join it with evidence." — Paraphrasing common guidance from research-methods faculty

Duration: 6–8 hours · Difficulty: Intermediate · Prerequisites: Web search, basic academic writing, library VPN and Zotero (or similar) strongly suggested.

Learning Objectives

By the end of this module, you will be able to:

1. Design a literature workflow: discovery → screening → extraction → synthesis → citation, with human checkpoints.
2. Use Semantic Scholar, Connected Papers, OpenAlex, and assistants like Elicit without confusing retrieval with having read the paper.
3. Apply PRISMA-style transparency (search strings, counts, exclusions) even outside medicine.
4. Verify every bibliographic fact before it enters Zotero—ghost citations are routine LLM failures.
5. Integrate citation managers with AI-assisted cleanup you still audit.
6. Disclose AI steps where your syllabus requires it.

Concept 1: Why Workflows Matter When AI Is Fast

Generative models compress text quickly; scholarship is an argument on evidence. AI cannot close the access gap (paywalls), design judgment (confounds), or integrity (fabricated refs).

Comparison table: Discovery tools

Concept 2: Discovery vs Understanding

Discovery finds candidates; understanding requires reading. Responsible AI uses: synonym expansion, Boolean strings you run, table templates you fill, summaries of text you paste (policy permitting).

Workflow map

Concept 3: Summarization and Integrity

Use structured prompts on text you have rights to paste; tag numeric claims [VERIFY IN PDF]. Never accept a bibliography from a model without Crossref/library checks.

Data and methodology

Concept 4: Writing workflow

You thesis + outline from your notes → AI rough expansion → you rewrite with real citations → reverse outline to catch “AI soup.”

Concept 5: PRISMA spirit and Zotero

Prespecify databases and dates; log queries; report screen-out reasons. Hallucination flavors: ghost papers, wrong DOI, merged citations, misquoted stats—delete or verify.

Hands-On: Python — Fetch Metadata by DOI (Crossref)

Key Example: Calling Crossref is the fastest way to kill a suspicious reference before it pollutes your bibliography. Use a polite User-Agent with your email per Crossref etiquette. pip install requests.

[object Object], requests

HEADERS = {,[object Object],: ,[object Object],}

,[object Object], ,[object Object],(,[object Object],) -> ,[object Object],:
    url = ,[object Object],
    r = requests.get(url, headers=HEADERS, timeout=,[object Object],)
    r.raise_for_status()
    msg = r.json()[,[object Object],]
    authors = [,[object Object],.strip()
               ,[object Object], a ,[object Object], msg.get(,[object Object],, [])]
    ,[object Object], {
        ,[object Object],: (msg.get(,[object Object],) ,[object Object], [,[object Object],])[,[object Object],],
        ,[object Object],: (msg.get(,[object Object],, {}).get(,[object Object],) ,[object Object], [[,[object Object],]])[,[object Object],][,[object Object],],
        ,[object Object],: authors,
        ,[object Object],: (msg.get(,[object Object],) ,[object Object], [,[object Object],])[,[object Object],],
    }

,[object Object], __name__ == ,[object Object],:
    ,[object Object],(fetch_crossref(,[object Object],))

Batch verification, OpenAlex search, and CSV loops follow the same pattern: polite headers, timeouts, small delays—describe those steps in words if you skip extra code.

Try This Now

1. Synonym burst — AI suggests 15 terms; you dedupe and map to database fields.
2. Abstract discipline — Compare AI summary to the abstract; mark added claims.
3. Integrity rehearsal — Write your school’s AI policy in your own words in ~100 words.

Practice Exercises

Exercise 1 — Narrow a broad topic; pre-register what evidence would prove the question too big.
Exercise 2 — Seed paper + five neighbors; justify reading order.
Exercise 3 — Five LLM citations on a niche topic: verify each; classify misses (fabricated / merged / wrong year).
Exercise 4 — Three papers → extraction table with “threat to my claim” column.
Exercise 5 — Messy bibliography → clean export via verified metadata.

Mini-Project

Annotated bibliography + synthesis memo (1500–2000 words): 12–15 sources, ≥8 peer-reviewed; PRISMA-style paragraph; AI disclosure with verification steps.

Key Takeaways

• Discovery tools reduce friction; you own rigor and ethical access.
• Verify or delete references—hallucinations are common.
• AI helps structure after you know the field, not as a substitute for reading.

Resources

• https://www.semanticscholar.org/ · https://openalex.org/ · https://www.connectedpapers.com/
• Crossref: https://www.crossref.org/documentation/retrieve-metadata/
• PRISMA: https://www.prisma-statement.org/ · Zotero: https://www.zotero.org/