This website uses essential cookies for authentication, language preferences and remembering that you dismissed this notice. No tracking or advertising cookies are used. Privacy policy

Why can AI invent references?

By · Content reviewed

Check references against publication records, including references suggested by a writing assistant. Research such as GhostCite examines generated and unreliable citations. Treat a suggested source as a lead to investigate, and establish its publication record before citing it.

Generated text needs a publication check

When a language model generates a reference from learned patterns, it predicts likely text rather than retrieving a publication record. The GhostCite paper describes how this process can combine real author names, plausible titles and familiar venue names. Citation styles provide a recognizable template, so a model can fill in a title, year and journal in the expected format. Without a separate lookup, that fluent output does not establish that a publication exists with the complete set of details.

Similar titles, authors and venues can also be combined into a new entry: an author associated with a research area may be paired with a title resembling another paper and a journal that publishes that subject. Each part sounds credible, but their combination may describe no actual work. A fluent answer is therefore not evidence of retrieval. Establish the complete publication record independently before relying on a suggested reference.

Why the fakes look convincing

Search each field against the publication you find. A real journal name or author name does not establish a particular paper. A working DOI also needs a title and author comparison: the DOI Foundation describes an identifier for an object, not a certification of a bibliography attached to it.

Does web search fix it?

If a writing tool supplies a source link, open it and read the source. Check that the reference describes the linked work, then check the passage you want to cite. A publication lookup and a claim-support check are separate tasks; Hallucite describes that boundary in its methodology and limitations.

How often does it happen?

Do not apply a headline rate to your own bibliography. Consult the setting and definitions in the original citation research before interpreting a study. No public Hallucite accuracy figure is claimed here; benchmark results await review, as explained in the measurement summary.

What reduces the risk

Verify references against publication records before relying on them: confirm the work, compare identifiers and metadata, and read its relevant passages. Keep the original suggested reference in your notes while you investigate. Hallucite can help you review a bibliography against available evidence; a flag does not establish authorship or intent. See how verification works.

Check a reference list against publication records

Hallucite compares references with available publication records and shows the evidence. Review flags and unresolved entries; a result does not establish authorship or claim support. Read the methodology and limitations