How to Use AI to Write a Research Paper (Without Fake Citations)

Yes — you can use AI to write a research paper, and nearly half of researchers already do. But there is one thing you must never let AI do: invent your citations. In a 2023 study published in Scientific Reports, 55% of the references GPT-3.5 generated for short literature reviews were entirely fabricated — papers that do not exist, with plausible authors, journals, and page numbers. GPT-4 did better, but still invented 18% of its references. A fake citation in a submitted manuscript is not a small embarrassment anymore: conferences now desk-reject papers over it.

This guide shows the workflow that keeps you safe: which tools to use at each stage of a paper, a 3-step citation check I tested myself on real and fake references, what journals actually require you to disclose, and copy-paste prompts for every section. The short version: use AI to find real papers, draft your own words, and verify every single citation against a database before you submit.

The safe workflow in 5 steps

Find sources with a research tool, not a chatbot. Ask Elicit or Semantic Scholar a question in plain language and read the abstracts yourself.

Collect your papers in Zotero. Every claim you plan to cite should trace back to a PDF you actually have.

Draft each section yourself first, then use AI as an editor. You supply the facts and the argument; AI supplies cleaner phrasing.

Run the 3-step citation check on every reference. DOI resolves, exact title search matches, and the claim is really on the cited page.

Disclose your AI use the way your target journal asks. Keep a log of tool, version, and purpose as you write, so the disclosure takes two minutes.

Can I use AI to write a research paper?

Yes, with disclosure — and you are far from alone. A Wiley survey of 4,946 researchers found that 45% of respondents had used AI to help with their research, most commonly for translation, proofreading, and editing manuscripts. A Nature survey of about 3,800 postdoctoral researchers found around one in three use AI chatbots in their work: 17% use them daily and 43% weekly, most often to refine text (63%), generate or edit code (56%), and find or summarize literature (29%).

Every major publisher allows AI assistance, and every major publisher draws the same two lines: AI cannot be listed as an author, and substantive AI use must be disclosed. The details differ by publisher:

One exemption is universal: basic grammar, spelling, and punctuation checks need no declaration at any of these publishers. The moment AI changes sentence structure, organization, or content in a substantive way, it crosses into disclosable territory. And one rule applies to reviewers as well as authors: never upload an unpublished manuscript into a public chatbot. Elsevier, IEEE, Wiley, and Springer Nature all forbid reviewers from doing it, and the same confidentiality logic applies to your own unpublished work — a chatbot is a third-party system, not a private desk.

Which AI tools should I use at each stage?

Different tools fail in different ways, so the trick is matching the tool to the stage. A general chatbot is excellent at tightening a paragraph you wrote and terrible at finding real papers. A research search engine is the opposite. Here is the split I use:

Elicit’s free Basic plan is genuinely usable for a single paper: unlimited search across its corpus of more than 138 million papers, unlimited summaries, unlimited chat with papers, and Zotero import. The paid Pro plan adds the systematic-review workflow that can screen 5,000 papers and a research agent that writes cited reports — useful for a thesis, overkill for a term paper. If you only need quick answers with sources, my guide to using Perplexity for research covers the mode-by-mode workflow and a citation-verification method that applies here too.

How do I find real sources with AI?

Start with a question, not keywords. Elicit is built for this: type a research question in plain language, for example “Do AI chatbots improve student writing skills?”, and it returns real papers with one-line summaries of what each one found. Then work through them in this order:

Skim the summaries, then read the abstracts of the five most relevant papers yourself. AI summaries are a triage tool, not a substitute for reading. Elicit itself warns that it cannot judge study quality reliably.

Export the keepers to Zotero. Elicit’s free plan includes Zotero import, and Zotero pulls the real metadata — authors, year, journal, DOI — straight from the publisher. If you are choosing between reference managers, my Zotero vs Mendeley comparison covers the trade-offs.

For a paper you already have, use NotebookLM. Upload the PDF and ask it to summarize the methods, pull out the key findings, or explain a section in plain language. Because it only answers from your documents, it cannot hand you a paper that does not exist. My PDF summarization shootout tested NotebookLM against ChatGPT, Claude, and Perplexity on the same 30-page report, and NotebookLM won on citation quality for exactly this reason.

Never accept a reference from a general chatbot at face value. ChatGPT and Claude generate citations from patterns in their training data, not from a database. A reference that looks perfect — right journal, right year, plausible page range — can still be entirely made up. That is what the next section is for.

How do I check every citation AI gives me?

This is the single most important habit in the whole workflow, and it takes about 30 seconds per reference. Run every citation through these three steps before it goes into your bibliography:

Paste the DOI into your browser. A real DOI resolves to the publisher’s page for that paper. If it 404s or lands on a different paper, the citation is wrong.

Search the exact title on Semantic Scholar or Crossref. Title, authors, and year must all match. A close match is not a match.

Open the paper and check the claim. Even a real citation can be misused — confirm the finding actually appears on the cited page, not just somewhere in the paper.

I tested this check myself while writing this guide, using the public Crossref and Semantic Scholar APIs. First, a real citation: Walters and Wilder (2023), “Fabrication and errors in the bibliographic citations generated by ChatGPT,” Scientific Reports, DOI 10.1038/s41598-023-41032-5. Crossref returned the exact title, authors Walters and Wilder, publication date September 7, 2023, and journal Scientific Reports. Semantic Scholar returned the same paper with matching authors. Pass — this is a citation you can use.

Then a plausible-looking fake: “Chen, L., & Martinez, R. (2019). Graph Neural Approaches to Bibliographic Entity Matching. Proceedings of SIGIR, 412–421.” It has real-sounding authors, a real conference, and a plausible page range. A Crossref title search returns no paper with that exact title — the closest matches are different papers on entity matching, none of them by Chen and Martinez, none of them from SIGIR 2019. Fail — drop it from the bibliography, no matter how well it fits your argument.

That test matters because the errors are not rare. In the Walters and Wilder study, 55% of GPT-3.5’s references and 18% of GPT-4’s were fabricated outright — and among the references that did point to real papers, 43% (GPT-3.5) and 24% (GPT-4) still contained substantive errors such as a wrong year, volume, or page range. Newer models do better, but no model is at zero: a 2026 benchmark cited in recent coverage found every one of thirteen state-of-the-art models still hallucinating citations, at rates from 14% to as high as 95% depending on the model and field. The stakes have risen with the rates. ICLR 2026 desk-rejected submissions with confirmed hallucinated references, and the NeurIPS 2026 handbook treats hallucinated citations as a violation of its Code of Conduct. Treat every AI-suggested reference as a rumor until the DOI resolves.

How do I draft each section with AI?

Use one worked example through the whole paper so the prompts stay concrete. Say your research question is “Do AI chatbots improve student writing?” Here is how each stage looks:

Outline first. Prompt: “Here is my research question: [question]. Here are the five papers I have read, with one-sentence summaries of each: [list]. Draft a six-section outline for a 5,000-word paper. Do not invent sources or claims I have not given you.” AI is good at structure; it is dangerous at content.

Literature review from papers you have read. Prompt: “Here is the abstract of [paper]. Summarize its method and findings in three sentences, then tell me how it connects to [your question].” If you have not read the paper, do not let AI summarize it for you — upload the PDF to NotebookLM instead, where the summary is grounded in the actual text.

Methods: write it yourself, then ask for a clarity check. Prompt: “Here is my methods section. Flag any sentence a reader could interpret two ways, and any step a researcher could not reproduce from this text.” Methods are the section where an AI invention would be research misconduct, not just a bad citation.

Discussion: use AI as a devil’s advocate. Prompt: “Here is my thesis and my three main findings: [paste]. What are the strongest counterarguments, and which alternative explanation for my results have I not addressed?” This is AI at its best — stress-testing your thinking instead of replacing it.

Abstract last, never first. Prompt: “Here is my finished draft. Write three versions of a 200-word abstract: one plain, one emphasizing the method, one emphasizing the finding. Keep every number identical to the draft.” An abstract written before the paper exists will drift from the paper as you revise.

The golden rule for every section: you supply the facts, AI supplies the phrasing. The moment you ask AI to write a section from nothing — no sources, no data, no draft — you are asking it to fill gaps with invented details, and invented citations are the most common result. If you want a second worked example of this discipline in action, my brainstorming guide shows the same principle applied to idea generation, using the real Doshi and Hauser (2024) Science Advances study on AI and creativity — a paper I verified through Crossref (DOI 10.1126/sciadv.adn5290) before citing it.

What should I disclose about AI use?

Keep a running log as you write — tool name, version, and what you used it for. Then the disclosure is a two-minute job at submission. Elsevier gives authors a template to copy:

“During the preparation of this work the author(s) used [NAME OF TOOL/SERVICE] in order to [REASON]. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.”

For a Nature journal, the same information goes in the Methods section. For IEEE, it goes in the acknowledgments, naming the AI system and the specific sections it touched. Whatever the venue, the disclosure has the same three parts: the tool, the purpose, and your confirmation that you checked the output and take responsibility for it. One more thing to note: if you are a student rather than a publishing researcher, your institution’s or course’s AI policy wins over any journal policy — check it before you start, not after you submit. My guide to using AI to study for exams covers where that line usually falls for coursework.

What are the honest limits of using AI for research papers?

AI cannot judge methodology or interpret results. It can polish how you describe an experiment, but it cannot tell you whether the experiment was sound. That judgment is yours and your reviewers’.

AI summaries can miss nuance. A summary that drops a limitation or a caveat can quietly change what a paper actually claims. Always read the paper before you cite it — a classic study estimated only about 20% of citing authors actually read the original, and AI makes that shortcut easier, not safer.

Research tools have limits too. Elicit’s own help center warns that its outputs must be checked and that it cannot judge study quality reliably, and it works best on empirical research in biomedicine and social science. NotebookLM only knows what you upload.

Policies change. Journal AI policies have been revised every year since 2023. Check your target journal’s current author guidelines before you submit — do not rely on a guide, including this one, for the final word.

One honest note on how this guide was built: I verified the citation-check workflow myself through Crossref and Semantic Scholar’s public APIs, and the tool pricing comes from Elicit’s own pricing page. I did not test Elicit’s paid Pro plan, since I have no paid account, and the hallucination rates quoted are from the 2023 Walters and Wilder study of GPT-3.5 and GPT-4 — newer models perform better, but none is at zero, and no study has yet tested the exact models you will use on the exact task you will give them. That is why the verification step is not optional.

Frequently Asked Questions