How to Use AI for Research Safely (Without Fake Sources)

Using AI safely for literature search comes down to a single rule: a language model may search, but it must never be the last link between a database and your reference list. It can suggest what to look for and where. It must not retype a title, a year, a volume number or a DOI — those have to reach your paper straight from the source. Everything below is an expansion of that one rule, because every practical step follows from it.

This is not excessive caution. It is what the measurements say.

Why a better prompt will not fix it

A study in Scientific Reports evaluated citations generated by ChatGPT and found that 55% of the GPT-3.5 citations, but just 18% of the GPT-4 citations, were fabricated (Walters and Wilder, 2023, p. 1). Newer models are clearly better — but the improvement does not solve the problem, because the errors migrate to the real entries. Among citations pointing to genuine works, 43% of the GPT-3.5 references and still 24% of the GPT-4 references contained substantive citation errors (Walters and Wilder, 2023, p. 1). Pooling six studies, 51% of 732 citations were fabricated (Walters and Wilder, 2023, p. 2).

Two findings from that paper matter most in daily practice:

• Fabricated citations look legitimate at first glance — you cannot spot them by their form (Walters and Wilder, 2023, p. 2).
• ChatGPT often gives incorrect responses when asked directly "Is this citation correct?" or "Do you fabricate citations?" (Walters and Wilder, 2023, p. 2).

The second one eliminates the defence most students reach for first. Asking the model to check itself is not verification; it is querying the same system about the same thing a second time.

Nor is it enough to simply wire a model up to a real database. As Szeider notes, connecting a language model to a bibliographic database API does not solve the problem, because the model may still fabricate or corrupt bibliographic data while processing the search results (2025, p. 2). Hallucination does not disappear because correct data exists nearby. It disappears when the correct data bypasses the model.

The principle: AI searches, the database supplies the data

The same paper shows what this looks like architecturally. The key move is to bypass the language model at the point of final data export, fetching entries directly from the authoritative source (Szeider, 2025, p. 1). The effect was measured on 104 obfuscated citations: an 82.7% perfect match rate with unmediated export, against 28.2% for standard web search — with zero cases of metadata corruption (Szeider, 2025, p. 3).

Translate that to your own workflow. A model is genuinely good at what it is suited for: interpreting an informal description ("I need something on how remote work affects teacher burnout"), proposing search terms, suggesting synonyms and adjacent concepts, summarising an abstract you found. It is bad at exactly one thing — reproducing from memory data that has to be correct character by character.

A safe workflow, step by step

1. Use the model to build the query, not to name the paper. Instead of "give me five sources on X", ask "what terminology does the literature use for X, and which journals publish it?" The first request invites fabrication; the second does not.
2. Search in a database. Google Scholar, Crossref, PubMed, JSTOR, your university's discovery service. Here the database decides, not the model.
3. Copy metadata from the publisher's page or export a citation file. Never retype it out of a chat window — that is precisely where the 24% error rate in genuine entries comes from. A BibTeX or RIS export from the publisher is safer than any amount of careful copying.
4. Resolve every DOI. If it lands somewhere other than it should, the entry is unverified. The full procedure is in how to check if a source really exists.
5. Get the file before you cite it. Minimum standard: you do not cite what you do not have. A page number in a citation must come from a document you actually opened.

What not to do

• Do not ask the model to verify its own citations. Reason above.
• Do not keep an entry because it "probably exists". Among model-generated citations, even those pointing to real papers carried a correct DOI in only 16–20% of cases (Chelli et al., 2024, p. 5).
• Do not assume a newer model settles it. Reported hallucination rates range from roughly 18% to over 70% depending on the model and the field (Misra and Udandarao, 2026).
• Do not generate the bibliography at the end. A reference list is built while you read, from files you actually hold.

Where cytado fits

We built our tooling on exactly this principle. The source finder takes a paragraph of your own writing, searches a corpus assembled from real bibliographic records (Crossref, OpenAlex, DOAJ and academic library indexes), and returns existing publications together with page numbers drawn from the pagination of the original files. The metadata never passes through a "model retypes it from memory" stage. If you already have a finished reference list, run it through the bibliography checker.

This article was produced the same way: every study cited above was found by cytado in its own corpus, with page numbers taken from the originals. It is the cheapest test of credibility we could submit to — provided we are telling the truth.