How to Prompt AI for Research
Most disappointing AI output is a prompt problem. Some of it isn't — there are questions no phrasing makes safe to ask. Here are seven prompts that make an AI genuinely useful in a research workflow, three that reliably produce fabrications, and the check that separates them.
The rule underneath all of it: retrieval beats recall
Every prompt below is an application of one idea. A language model asked to recall a source produces text shaped like a citation. A model asked to work on a source you supplied produces an analysis of that source. The first mode is where fabrication lives.
The size of that gap is measurable. Walters and Wilder, writing in Scientific Reports in 2023, had ChatGPT produce short literature reviews on 42 topics and checked all 636 resulting citations: 55% of the GPT-3.5 citations and 18% of the GPT-4 citations referred to papers that do not exist. Among the citations that were real, 43% (GPT-3.5) and 24% (GPT-4) still contained substantive errors. Newer models score better, but the failure mode has not changed — and an 18% fabrication rate is still roughly one invented reference in every five.
So the goal of a good research prompt is rarely "get the model to know more." It is to move the work from recall into retrieval, critique, or structure — three things models are genuinely good at.
Seven prompts that work
1. Turn a question into a search, not into sources
"My research question is: does spaced repetition improve long-term retention in adult learners? Break it into concepts, give me synonyms and likely database subject headings for each, and combine them into three Boolean search strings — one broad, one precise, one for grey literature."
This is the single highest-value AI prompt in research, and it cannot fabricate anything: you take the strings to Google Scholar or a subject database and the results are real by construction.
2. Paste the paper, then ask
"Here is the full text of a paper. Summarize the design, the sample, the outcome measure, and the effect the authors report. Quote the sentence you drew each answer from. If the paper does not state something, write ‘not stated’ — do not infer."
Naming a paper invites reconstruction; supplying it invites reading. The "quote the sentence" clause matters as much as the question — it gives you a two-second way to check each claim against the text rather than trusting the summary wholesale.
3. Ask for the objection, not the verdict
"Here is my argument. Give me the three strongest objections a hostile reviewer in this field would raise, ranked by how hard they are to answer."
"Is this a good argument?" is close to useless as a prompt. Models fine-tuned on human feedback lean sycophantic: Sharma et al. (2023) found five leading AI assistants consistently matched users' stated views across free-form tasks, and that matching the user's view was one of the strongest predictors of which response human raters preferred. Ask a leading question and you will get a flattering answer. Ask for the attack and you get something you can use.
4. Reverse-outline your own draft
"Read this section and write one sentence per paragraph stating what that paragraph claims. Then list the paragraphs whose claim doesn't follow from the one before it."
A reverse outline exposes structure you can no longer see in your own text. This is grounded work — the draft is right there in the prompt — so there is nothing to invent.
5. Make it show the steps for anything analytical
"Work through whether this design can support a causal claim. Set out the reasoning step by step before giving a conclusion."
Wei et al. (2022) showed that prompting a model to produce intermediate reasoning steps — chain-of-thought prompting — substantially improved accuracy on arithmetic, commonsense and symbolic reasoning benchmarks. The practical benefit for you is different but just as useful: visible steps are auditable steps. You can see which link in the chain is wrong instead of accepting or rejecting a verdict whole.
6. Require explicit uncertainty
"Answer only from the text I provided. Mark anything you are inferring with [INFERENCE] and anything the source does not cover with [NOT IN SOURCE]."
Models default to fluent completeness — the register of confident prose is exactly what they learned. Giving uncertainty a label makes it cheap to express, and the flagged lines are precisely the ones worth your attention.
7. Use it as a rubric-driven reviewer
"Here is my methods section and here is my department's marking rubric. For each rubric criterion, quote the sentence in my draft that satisfies it, or say it's missing."
Generic "give me feedback" produces generic feedback. Supply the actual criteria — a rubric, a reporting checklist, a journal's author guidelines — and the response becomes specific and checkable.
Three prompts that backfire
- "Find me 10 sources on X." Pure recall. This is the prompt that produces plausible authors, plausible journals, plausible DOIs, and papers nobody wrote. Use prompt 1 instead.
- "Write my literature review on X." It will read well and be built on sources you never chose or read — which is both an integrity problem and a supervision disaster the first time you're asked about one of them. Build the review from papers you have actually collected.
- "Make this sound more academic." You get nominalizations, passive constructions and longer sentences — the surface features of academic prose without its precision. Ask for the specific fix instead: "replace vague quantifiers with the actual numbers," or "hedge any claim the data doesn't support."
The 60-second verification pass
Whatever the prompt, run this before a reference enters your bibliography:
- Paste the DOI after
https://doi.org/. If it doesn't resolve, the paper doesn't exist in that form. - No DOI? Search the exact title in quotation marks in Google Scholar or your library discovery layer.
- Open it and find the specific claim. A real paper cited for a finding it never reported is still a fabrication.
- Copy the metadata from the source itself, not from the chat window.
Once a source has passed the check, drop its details into our free citation generator — APA 7, MLA 9, Chicago, Harvard or IEEE — and copy a clean reference and in-text citation.
What no prompt can fix
Prompting shapes how a model uses what it has. It cannot give a model access to a paywalled paper it never saw, and it cannot make a confident sentence true. That is a tooling problem, not a wording problem: the reliable answer is to work from retrieved documents rather than remembered ones, which is why serious research tools search first and write second. If you want the reasoning behind that failure mode in detail, we wrote it up in why ChatGPT makes up citations.
Skip the prompting problem entirely
CiteGood researches from real papers and checks every claim against the source it cites — so there's no fabricated reference to catch.
Try CiteGood free →Frequently asked questions
What is the best prompt for finding academic sources?
There isn't one — asking a chatbot to recall sources from memory is the wrong job for the tool. Ask it to build the search instead: 'Turn this question into three Boolean search strings for Google Scholar, with synonyms and controlled vocabulary for each concept.' Then run the searches yourself in a database that only returns papers that exist.
Does asking AI for a DOI stop it from making up citations?
No. A DOI is a short string in a predictable format, so a model can fabricate one as easily as it fabricates an author list — and a plausible-looking DOI is harder to doubt than a plausible-looking title. The only fix is to resolve it: paste the DOI after https://doi.org/ and see whether it lands on the paper you were promised.
Is it better to paste the paper into the prompt or just name it?
Paste it. When the text is in the prompt, the model is summarizing something it can read; when you only name the paper, it is reconstructing what a paper with that title probably said. The second mode is where invented findings and misattributed quotes come from.
Why does AI agree with everything I say about my research?
Because models fine-tuned on human feedback tend toward sycophancy — Sharma et al. (2023) found five leading AI assistants consistently matched users' stated views across free-form tasks, and that agreement with the user was among the strongest predictors of which response humans preferred. Ask for the strongest objection to your argument rather than asking whether your argument is good.
Can I use AI prompts for research without breaking academic integrity rules?
Usually yes for search strings, outlines, critique and editing; often no for generating prose you submit as your own. The rule that applies is your institution's, and many journals now require a disclosure statement describing how AI was used. Check the assignment brief or the journal's author guidelines before you start, not after.