One of the most-shared AI research posts on Chinese academic social media is titled “How to let AI help you find research ideas from scratch.” The underlying method is real and learnable. Here it is, without the hype.
What AI Can and Cannot Do Here
AI doesn’t know your field the way you do after three years of daily reading. It can’t invent novel scientific contributions. What it can do: help you think systematically about gaps in a body of work you already understand, combine concepts across subfields in ways you hadn’t considered, and surface adjacent literature you weren’t tracking.
Step 1: Map Your Field’s Assumptions
Describe your research area to Claude at length. Include what methods are standard, what outcomes are considered worth studying, what assumptions underlie most work in the field. Then ask: “What assumptions in this field have the least empirical support?” or “What does this field assume that adjacent fields have challenged?” The gaps often live in unexamined assumptions rather than in “nobody has done X study.”
Step 2: Cross-Field Contamination
Name your field and 2-3 adjacent fields. Ask Claude: “What methods, frameworks, or findings from [adjacent field] haven’t been applied to [your field] yet?” This is productive because the literature in two adjacent fields almost never fully overlaps. The missing application is often the research idea.
Step 3: Generate a Long List, Then Filter Hard
Ask for 20–30 potential research directions based on your discussion. This is generative work where AI quantity helps. Then you filter with your expertise: Which is actually novel? Which is feasible given your resources? Which do I have a methodological advantage to execute? Which can be done in 2 years?
Step 4: Stress-Test the Surviving Ideas
For each surviving idea, ask Claude to argue against it: “What would a skeptical reviewer say about this research direction?” Then look up whether the objection is real in the literature. The ideas that survive this process are genuinely worth pursuing.
The Constraint That Matters
The process only works if you have real deep knowledge of your field first. AI pattern-matches on what you tell it. If your field knowledge is shallow, the ideas it generates will be shallow too. This tool amplifies existing expertise — it doesn’t substitute for building it.
