HallucinationNerd checks whether real citations support the claims attached to them

The easiest fake citation to catch is a paper that never existed. Search the title, and the trail collapses. A more dangerous error is a genuine scientific paper used to justify something its authors did not find. The author name, journal and DOI may all be correct, and the cited paper may even address the same disease. Only a reader who compares the specific claim with the supporting passage notices that the evidence does not travel as far as the sentence. Researchers at New York University have built a tool around precisely that gap.[1]

Their system, HallucinationNerd, was described in the journal Information in October 2026. It breaks a document into factual assertions, matches each assertion to the citation attached to it, retrieves relevant portions of the cited document and asks whether the text really supports the statement. The research is accompanied by another system, ProtocolNerd, intended to help scientists find laboratory procedures suitable for a proposed experiment. Together they reflect an important change in the ambition of research AI: not just composing a fluent report, but checking the evidence structure beneath it.[1][3]

Six verdicts expose different problems

HallucinationNerd distinguishes fully supported, partially supported, unsupported and contradicted claims, alongside references that do not exist and documents it cannot access. Those categories prevent a failed retrieval from being mistaken for a false statement. Evidence pointers let a reader inspect the material behind the verdict.[3]

The distinctions matter. Suppose a paper finds that a treatment improves a laboratory biomarker in one small study. A summary stating that it saved patients' lives would not become supported merely because the paper itself was real and the keyword matched. It might be a much more serious overstatement than accidentally mistyping a publication year. A useful checker must examine the relationship between a sentence and a source, not only whether the bibliography contains valid records.[1]

A citation swap is a useful first test

The researchers built a benchmark from arXiv papers, replacing some legitimate citations with randomly selected papers from unrelated areas. The resulting claim–source pairs let them test whether the auditor detected a broken connection. The authors report stronger performance than the systems used for comparison on this dataset.[1]

A separate demonstration examined a scholarly AI assistant’s answers to five health questions. HallucinationNerd extracted 43 cited claims; it could not assess three because the required evidence was unavailable. It classified 37 of the remaining 40 as supported and three as partially supported. Those were the tool’s own classifications, not an independent human audit.[1]

These examples test two different things. Replacing a citation with an unrelated paper exposes an obvious break in the evidence chain; the small health-question demonstration shows how the auditor categorises ordinary output. Neither, by itself, establishes how reliably it detects a closely related paper whose population, endpoint or conclusion has been subtly changed. That harder task is the natural next benchmark.

Finding an experiment before checking the write-up

ProtocolNerd takes a different approach. A researcher describes a desired experiment, and the system asks for details that could change which procedure is appropriate. For example, identifying a test of drought tolerance in rice depends on the plant's growth stage. It then retrieves candidate protocols from scientific records and explains why each might fit or not fit. According to the NYU account, the team evaluated this retrieval task using material drawn from PubMed and Europe PMC and compared the system with alternative approaches.[2][3]

This is attractive because research articles often contain method descriptions buried far from the sentence someone searches for. Finding a broadly related experiment is not the same as finding a usable protocol. Equipment, specimen preparation, reagents, calibration and safety procedures can make two studies that look similar on a title page impossible to reproduce using the same steps. A conversational search tool can narrow the candidate set, but a lab still needs to read the original method and determine whether it fits its particular setting.[1]

ProtocolNerd starts with the question a scientist hopes to answer; HallucinationNerd starts with the answer someone has written. Their common contribution is to make the intermediate evidence visible. A researcher can inspect the proposed protocol or the retrieved passage rather than accept an unexplained score. In research writing, that makes the citation useful again: a route to the evidence behind a sentence.

Report an error or suggest a correction