ChatGPT is the most widely-used AI tool in the world, and millions of students and researchers reach for it first. It is genuinely useful for brainstorming, outlining, and editing. But there is one task where using it naively can quietly wreck your work — and your credibility: citations. ChatGPT has a well-documented tendency to fabricate references that look completely real but do not exist.
This guide explains why that happens, what the latest 2026 data actually shows, how the newer "Deep Research" mode changes the picture, and — most importantly — how to use ChatGPT for research without putting fake citations in your paper.
Why ChatGPT invents citations
To understand the problem, you have to understand what ChatGPT is. A large language model is a next-word prediction engine trained to produce plausible text. It does not have a database of papers it looks things up in; in its base chat mode, it generates what a citation should look like based on patterns in its training data. The result is a reference with a realistic-sounding title, plausible authors, a real-looking journal, and a confidently-formatted DOI — that is entirely invented.
This is called a hallucination, and citations are especially vulnerable to it because their format is so regular. The model has learned exactly what a reference looks like; it just has no mechanism, in plain chat, to guarantee that the specific paper exists.
What the data shows
This is not a hypothetical risk — it is measured. A widely-cited 2023 study in Scientific Reports found that a large share of the bibliographic citations ChatGPT generated were fabricated or contained errors. More recent samples suggest the base hallucination rate has improved but remains high — on the order of a third or more of references in some tests still being problematic.
More worrying is what is happening downstream, in the published literature. According to an analysis covered by Retraction Watch in 2026, roughly one in 277 papers indexed in PubMed in the first weeks of 2026 cited a reference that did not exist — up sharply from about one in 458 in 2025 and one in 2,828 in 2023. The trend is unmistakable: as AI use rises, fabricated citations are leaking into the formal scientific record. Even elite venues are affected — analyses of papers at top machine-learning conferences in 2025–2026 found confirmed hallucinated citations that peer reviewers had missed.
The takeaway is blunt: a fabricated citation is not a quirky AI mistake. It is research misconduct if it ends up in your submitted work, whether you intended it or not.
Does "Deep Research" fix it?
Partly — and it is important to be fair here. ChatGPT's Deep Research mode is a real improvement. Instead of generating text from memory, it actively searches the web, reads sources, and assembles a multi-page report with citations and links you can click through to the original material. Because it is grounded in retrieved sources, its citations are far more likely to point to real documents, and you can verify each one.
But three cautions remain:
- Deep Research is not the default. Most casual ChatGPT use is still plain chat, which fabricates. And Deep Research requires a paid plan.
- Grounded does not mean correct. Even when a source is real, the model can misrepresent what it says — citing a real paper for a claim it does not actually make.
- It is built for reports, not your manuscript. Deep Research produces a standalone summary; it does not write your paper with citations woven into your argument.
So Deep Research narrows the citation problem but does not eliminate the need to verify — and it does not turn ChatGPT into a research-writing tool.
How to use ChatGPT for research safely
You do not have to avoid ChatGPT. You have to use it for the right jobs and guard the citations:
- Use it for thinking, not sourcing. Brainstorming, outlining, clarifying concepts, and improving your prose are great uses.
- Never paste an AI-generated citation into your paper unverified. Treat every reference as fake until you have personally found the real paper.
- Verify against a real index. Check each citation in PubMed, Google Scholar, or the publisher's site. If you cannot find it, it does not exist.
- Don't ask it to "add citations" to claims. That is exactly the prompt that triggers fabrication. Find the source first, then write the claim.
- Prefer tools grounded in real papers for anything citation-critical.
A better approach for the writing itself
The deeper issue is that ChatGPT is a general tool. It was never built for the specific job of writing a research paper with accurate citations — which is why citations are the seam where it fails. For the writing itself, you want a tool designed around citation integrity from the start.
That is the principle behind PaceResearcher. It is a research writing workspace that drafts with you and inserts real, verifiable inline citations as you type, synthesizing genuine sources from a 200M+ paper corpus — never fabricated references. The difference is architectural: where ChatGPT generates plausible-looking citations from memory, PaceResearcher is built to cite real papers and put them directly into your manuscript, correctly formatted.
Use ChatGPT to think. Use a verification step on every reference. And for the actual writing-and-citing, use a tool that cannot fabricate by design.
For the related issue with Google's model, see Gemini for research: the citation problem. For the full landscape of research tools, read our best AI tools for research guide. And when you are ready to write a paper whose every citation is real, start with PaceResearcher free.