How to Use AI for Research: A Step-by-Step Workflow

Used carelessly, AI is the fastest way to write a confident research paper on sources that do not exist. Used as a pipeline, it compresses a two-week literature review into two days — with better citations than you had before. This is the exact workflow our research desk uses, stage by stage, with the tool for each job and the trap to avoid.

The pipeline at a glance

1. Scope — ChatGPT or Claude, for the map of the field
2. Search — Perplexity, Elicit, Consensus for real papers
3. Read — NotebookLM, grounded in your sources only
4. Organize — Zotero for citations, always
5. Write — your words, with AI as a structural editor

Step 1: Scope the field

Start by asking a general model for the landscape, not the answers:

“What are the main competing theories on X, who holds them, and what would a 2019 source miss?”

You are buying a map, and maps are allowed to be wrong — their job is to give you names, terms, and debates to verify in step 2.

The trap: believing the map is the territory. Anything factual from this stage gets re-verified or deleted.

Step 2: Search with grounded tools

This is where you switch to tools that retrieve real documents:

• Perplexity — fast, cited answers; excellent for recent developments and industry sources
• Elicit — searches actual papers, extracts claims into a table you can sort; built for literature reviews
• Consensus — yes/no-style questions answered from the literature with a confidence meter
• Semantic Scholar — free, with citation graphs that show which paper the field actually leans on

Export everything to Zotero as you go. Your future self does not want to reconstruct a bibliography from browser history.

Step 3: Read with NotebookLM

Upload your collected PDFs to a NotebookLM notebook and interrogate only those sources:

“Which of my sources contradict each other on dosage effects?”

Every answer cites the chunk it came from, so verification is one click, not one afternoon.

The Audio Overview feature — two AI hosts discussing your papers — is genuinely useful for absorbing a dense set on a walk.

The rule that keeps this whole workflow honest: the model may summarize, compare, and structure — but every claim that reaches the page traces to a PDF in Zotero.

Step 4: Keep citations mechanical

Let Zotero generate every citation.

AI can format a bibliography, but it will happily invent a plausible DOI at 1 a.m. The plugin does it perfectly, every time, in any style. There is no reason to do this part by hand in 2026.

Step 5: Write it yourself, edit with AI

Draft the argument in your own voice — this is the part that makes it research and not reportage.

Then use Claude as a structural editor:

“Here is my section. Where does the argument leak, and which claim is weakest?”

It will find the hole you have been politely ignoring.

The traps, collected

• Citation laundering — an AI summary of a source is not the source; always open the PDF
• Recency blindness — models love old consensus; check for 2025–2026 papers explicitly
• Confidence theater — fluent prose is not evidence; weight by study design, not vibes
• Single-tool pipeline — one retriever has one blind spot; two retrievers have two smaller ones

Research is judgment plus sources. The AI handles the fetching; the judgment remains, stubbornly and correctly, yours.