Can AI write a literature review for me?

AI can draft the prose and structure of a literature review, but it can't reliably do the sourcing. General chatbots fabricate citations and summarise papers they've never read, so any AI-written review has to be fact-checked claim by claim against the real sources before you can trust it.

Is it ethical to use ChatGPT in research?

Using ChatGPT to draft, outline, or polish is increasingly accepted, and many universities now permit AI-assisted writing. What's not acceptable is presenting fabricated citations or unverified claims as your own scholarship. The line is disclosure where required, plus verifying every source yourself.

Will ChatGPT make up references in a literature review?

Yes, frequently. In one analysis of 115 ChatGPT-generated medical references, only 7% were both real and accurate — 47% were entirely fabricated and 46% cited real papers incorrectly. The fakes look convincing, with plausible authors, titles, and journals, so they're easy to miss.

What's the safest way to use ChatGPT for a literature review?

Bring your own vetted sources rather than asking it to find papers, make it work from text you paste in instead of its memory, and check every citation against the actual page. Use it to speed up the writing, never to decide what the literature says.

Can ChatGPT Write a Literature Review?

Yes — ChatGPT can draft the prose of a literature review in seconds. But it routinely invents citations and misreads what sources actually say, so its output can’t be trusted as-is. Used carefully — with your own vetted sources, and every claim checked against them — it speeds up the writing, not the thinking.

That distinction is the whole game. Below is what ChatGPT is genuinely good at, where it fails dangerously, and how to use it without putting fake references into your work.

Key takeaways

ChatGPT can draft a literature review, but it can’t reliably source it. It writes fluent prose; it does not check whether the papers it cites exist or say what it claims.
The fabrication rate is high. In one analysis, only 7% of ChatGPT’s references were both real and accurate — and newer models still invent roughly 1 in 5.
The problem is accelerating. Fabricated citations in published papers have climbed from 1 in 2,828 (2023) to 1 in 277 in early 2026.
It’s usable if you stay in control. Bring your own vetted sources, check every citation against the page, and never paste an AI reference list in unverified — or work from a tool that grounds every claim in your library from the start.

What ChatGPT is genuinely good at

As a writing assistant, ChatGPT is capable and fast. When you supply the substance, it can:

Draft an outline and shape the structure of your review.
Summarise or rephrase a paper you paste in.
Smooth transitions, tighten wordy paragraphs, and vary your phrasing.
Suggest themes or groupings across notes you’ve already written.

In every one of these, you are the source of truth and ChatGPT is doing the typing. That’s the safe lane.

The problem: it makes up the citations

A literature review is its citations. And this is exactly where general-purpose chatbots fail.

In one analysis of 115 references generated by ChatGPT for medical questions, only 7% were both real and accurate. 47% were entirely fabricated, and 46% pointed to real papers but cited them incorrectly. (Cureus, 2023)

This isn’t a one-off, and newer models haven’t fixed it. A comparative study of ChatGPT and Bard for systematic reviews found that even GPT-4 hallucinated 29% of its references (GPT-3.5: nearly 40%; Google’s Bard: a staggering 91%), and a 2025 study of GPT-4o writing mental-health reviews found about 20% of its citations were fabricated — with 45% of the real ones containing errors. The reason is structural: a language model predicts plausible-looking text. A citation with real-sounding authors, a credible journal, and a fitting title is easy to generate and hard to catch — the fabrications are convincing precisely because they’re statistically typical.

And the damage is showing up in the published record. A 2026 Lancet audit of 2.5 million biomedical papers found that the share of papers with at least one fabricated reference rose from 1 in 2,828 in 2023 to 1 in 277 in the first weeks of 2026.

You don’t have to look far for the consequences. When the White House’s 2025 MAHA children’s-health report was published, journalists found seven of its citations pointed to studies that don’t exist — and one Columbia epidemiologist said a paper the report credited to her “is not a real paper that I or my colleagues were involved with.” The White House called the errors “minor citation and formatting errors.”

For a graded essay, that’s risky. For published research, it’s disqualifying.

It also hasn’t actually read your field

Even when a citation is real, ChatGPT often hasn’t read it. It’s the other side of those real-but-miscited references above — the paper exists; the summary is the model’s guess, not the page. Without your sources in front of it, it writes from a frozen, general training set: it misses recent papers, can’t open paywalled work, and will confidently summarise studies it has never seen. A review built that way isn’t a review of the literature — it’s a review of the model’s best guess about the literature.

How to use ChatGPT for a literature review safely

You can still get real value from it. The rule is simple: make it work from your sources, and verify everything.

Bring your own sources. Don’t ask ChatGPT to find papers — give it the ones you’ve already vetted.
Feed it text, not memory. Paste in the abstract or the relevant section, and ask it to summarise that, not what it “knows.”
Check every citation to the page. Confirm the author, year, journal — and that the paper actually says what’s claimed.
Stay in charge of the argument. Use it to draft sentences, not to decide what the evidence shows.
Never paste an AI reference list into your manuscript without verifying each entry individually.

Follow those and ChatGPT becomes a fast drafting tool instead of a fabrication machine. Skip them and you’re gambling your credibility on a model that invents sources more than half the time.

The grounded alternative

The deeper reason ChatGPT fabricates is that it isn’t connected to your library. The fix is a workspace that drafts from your sources and traces every claim back to the exact page and paragraph — so a citation is something you can click and verify, not something you have to fact-check after the fact.

That’s what we’re building Shubo for. A literature review is grounded writing at its most citation-dense — exactly where Shubo fits. Drop in your papers, and every sentence it drafts links back to the passage it came from, so the review is sourced as you write it, not fact-checked after; when the sources don’t support a claim, it tells you instead of inventing one. The same grounding carries through everything else you write — your next paper, a thesis chapter, a grant proposal. You stay the author; the argument and the judgment are yours.

Bottom line

Can ChatGPT write a literature review? It can write the words. It cannot be trusted to source them. Treat its citations as guesses until you’ve proven them real, keep your own judgment in the loop — and, better yet, work from a tool that grounds every claim in your library from the start.

Sources

Bhattacharyya M, et al. High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content. Cureus, 2023.
Chelli M, et al. Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis. Journal of Medical Internet Research, 2024.
Linardon J, et al. Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study. JMIR Mental Health, 2025.
Topaz M, Roguin N, et al. Fabricated citations: an audit across 2·5 million biomedical papers. The Lancet, 2026.
NOTUS. The MAHA Report Cites Studies That Don’t Exist. 2025.