Guide · Researching
How to use AI for research
AI genuinely accelerates research at four points: finding literature, pressure-testing a method, exploring data, and drafting prose. It cannot be the authority on any factual claim, and it fabricates citations in a form that looks correct. Used with a verification step attached, it is the largest practical gain research workflows have had in years.
How long does it take to learn?
A week to change how you read literature. A term to trust it near your own methodology, which is the right pace: the failure mode here is not slowness, it is a retraction.
The fastest path, in order
- 01
Use it to map a field, not to state facts
Ask what the debates are, who disagrees with whom, which papers everyone cites. It is very good at shape and unreliable on specifics, and directing it at the thing it is good at is most of the method.
- 02
Find papers with tools built to find papers
Elicit, Consensus and Semantic Scholar search real indexed literature. A general chatbot asked for references is guessing what a reference would look like, and it will produce one.
- 03
Verify every citation, mechanically
DOI, title, authors, year, and the sentence you are citing it for. Fabricated references are correctly formatted and plausibly titled, so eyeballing does not catch them. Only lookup does.
- 04
Use it against your method, not for it
Ask what a reviewer would attack, what confound you have missed, what the counter-hypothesis is. As an adversary it is excellent. As the author of your design it is a liability.
- 05
Let it write code, not conclusions
Analysis scripts, cleaning, plots: strong, and checkable by running them. Interpretation of what the numbers mean stays yours, and that boundary is the one worth holding hardest.
- 06
Draft with it, and read every line
It is good at structure and transitions and prone to confident filler. Know your institution's disclosure rules and your journal's policy before submission, not during.
Tools worth your time
| Tool | What it is for |
|---|---|
| Elicit | Literature search and extraction over real papers, built for systematic review workflows. |
| Consensus | Searches findings across papers and shows where the evidence agrees and where it does not. |
| Semantic Scholar | Free index with citation graphs, the fastest way to confirm a paper exists. |
| Connected Papers | Visual map of what a paper is related to, useful for finding what your search missed. |
| NotebookLM | Grounds answers in documents you upload, so it quotes your sources instead of its memory. |
| Zotero | Reference manager. Unglamorous, and the thing that makes citation checking a lookup. |
Grounded tools that read documents you provide are categorically safer than open-ended chat for anything factual, because the answer is tied to a text you can open. That single choice removes most of the fabrication risk before any checking begins.
Mistakes that cost people weeks
Asking a general chatbot for references
It will produce a formatted, plausible, non-existent paper. This is the most reported AI failure in academic work and it is entirely avoidable by using tools that search real indexes.
Letting it summarise a paper you have not read
Fine for triage, dangerous the moment you cite it. Summaries lose the qualifications, and the qualifications are usually the finding.
Not checking the disclosure rules
Institutions and journals now have explicit policies, and they differ. Reading yours takes ten minutes and is the difference between a legitimate tool and a misconduct case.