Literature Review With LLMs: Scope First, Search Second

Classic lit review starts with keyword search. That works when you know the terrain. For a new domain, keyword-first means you find the loudest papers, not the most load-bearing ones. An LLM can compress a decade of conversation into a navigable map in ten minutes — but only if you make it scope the review, not summarize it.

The scoping prompt that pays for itself

1. 1Use the LLM to draft inclusion and exclusion criteria BEFORE searching
2. 2Ask it to name the rival camps in the field, not just the consensus
3. 3Force one-line criticisms on every seed paper — this surfaces the debate shape
4. 4Require an uncertainty flag on any claim about a specific paper

Handing off to a search engine

Once the LLM has given you the conceptual map, switch tools. Use Semantic Scholar or Consensus for actual papers — these are retrieval-backed and far less likely to invent sources. The LLM is a sparring partner for scope; it is not a database.

The big idea: scope first with an LLM, search second with a retrieval tool. Reverse that order and you inherit the LLM's biases instead of your own.

A lit review needs structure before content. AI can propose a thematic outline; you fill it with what you actually read.

What AI does well here

• Propose 3-5 themes from paper titles and abstracts you paste.
• Slot each paper under a theme.
• Suggest where contradictions between papers belong.

What AI cannot do

• Read papers it doesn't have access to.
• Verify which papers are seminal in the field.
• Substitute for your synthesis of the actual texts.

AI can take a topic and a list of sources and draft a literature review skeleton with thematic sections and topic sentences.

What AI does well here

• Group sources by theme rather than chronology
• Draft topic sentences and transitions between sections

What AI cannot do

• Verify a paper says what AI claims it says
• Generate genuinely new theoretical synthesis