Detecting Hallucinated and Suspicious Citations: What Current Tools Can and Cannot Do
Fidan Badalova, Philipp Mayr
cs.DL
2026-07-18
On 104 verified references, no hallucinated-citation detector combines high recall with low false positives; the most thorough one flags about half of real references too.
When models such as ChatGPT help draft papers, they routinely invent references that do not exist: wrong DOIs, mismatched authors, or citations fabricated outright. This is not a sporadic glitch. It is contaminating the published record.
A 2026 audit in The Lancet (Topaz et al.) found 4,046 fabricated citations across 2.5 million biomedical papers. The frequency rose from roughly one fabricated reference per 2,828 papers in 2023 to one per 277 in early 2026, a tenfold worsening in a little over two years. A separate multi-corpus study (Zhao et al.) estimated that about 146,932 hallucinated citations appeared across four preprint platforms in 2025, with SSRN the worst at 1.91%. Conferences are not immune: ACL 2026 found citations to non-existent literature in more than 100 accepted papers during camera-ready checks.
The severity is well documented. What is missing is an honest account of whether the tools built to catch these citations actually work. This position paper fills that gap, collecting the evidence on severity and running a controlled test of five open-source detectors.
This paper proposes no new algorithm. It runs a controlled evaluation. The authors built a test set by hand: 104 references drawn from three scholarly documents, each verified individually, with 33 flagged as problematic (wrong titles, mismatched authors or venues, DOIs that fail to resolve or point elsewhere, or references that cannot be found) and 71 as genuine.
The set breaks down as P1 with 24 references (3 problematic), P2 with 15 (4), and P3 with 65 (26). The long document concentrates most of the problems.
Five tools were run over the set: CheckIfExist, HalluCiteChecker, Hallucinator, HalRef, and RefChecker. The logic is broadly similar. Each parses references out of a PDF or BibTeX file (title, authors, year, DOI) and checks them against scholarly databases such as Crossref, Semantic Scholar, OpenAlex, DBLP, and PubMed to decide whether a citation exists and whether its fields match. The differences are in parsing robustness, how many databases are consulted, and whether matching relies on the title alone or combines DOI and metadata.
Each reference, for each tool, falls into one of four buckets: a real problem caught, a genuine reference correctly left alone, a genuine reference wrongly flagged, or a real problem missed.
| Tool | Fake citations caught (TP/33) | Missed (FN) | Real references wrongly flagged (FP/71) |
| RefChecker | 32 (97%) | 1 | 36 (51%) |
| CheckIfExist | 31 (94%) | 2 | 34 (48%) |
| Hallucinator | 29 (88%) | 4 | 28 (39%) |
| HalRef | 24 (73%) | 9 | 53 (75%) |
| HalluCiteChecker | 18 (55%) | 15 | 20 (28%) |
Recall and false-positive rates were derived from the absolute TP/FP counts in the paper; the original reports only raw numbers.
The pattern is clear: recall and false positives move together. RefChecker and CheckIfExist catch the most fabricated citations, missing only one or two of the 33 problematic references, but they also flag nearly half of the 71 genuine references as suspicious. HalluCiteChecker sits at the other extreme, with the fewest false positives (28%) but missing nearly half the bad citations. Hallucinator is the most balanced. HalRef produces the richest output, assigning a hallucination score to every reference, but flags 75% of genuine references, which makes it unusable as-is.
The authors repeat one point throughout: the output of these tools should not be treated as a final decision.
For anyone using LLMs to help write, two practical takeaways follow. First, do not trust a single tool as an oracle. Treat it as a coarse screen, and manually verify the DOI, title, authors, and venue for anything it flags. Second, pick the tool for the error you can tolerate: HalluCiteChecker if you would rather miss some, RefChecker or CheckIfExist if you would rather over-flag, Hallucinator for balance.
For tool builders, the paper locates the bottleneck upstream of matching. The reference parser that extracts citations from PDFs breaks often, BibTeX with slightly different formatting throws the tools off, and incomplete database coverage causes real but unindexed citations to be misjudged. The improvement is in more robust parsing (GROBID-style structured extraction) and querying more sources, not in tuning a single matcher harder.
One trap the paper names explicitly: general-purpose AI text detectors such as Pangram cannot catch citation hallucination. They judge whether prose looks AI-written; they never check whether a reference exists or whether its fields are correct.
The biggest limitation is the test set: 104 references, 33 problematic, only three documents. Percentages such as 97% recall or 75% false positives from that base show a direction, not a stable estimate. A different mix of disciplines or PDF layouts could reshuffle the ranking.
The evaluation is a one-off run on the authors' own machines, with no reported reproducibility. Different versions or configurations (whether Hallucinator has GROBID installed, for instance) would give different numbers. The P3 document alone contributes 26 of the 33 problematic references, so a single document weighs heavily on the totals.
Most tools are black boxes. HalRef does not disclose how its hallucination score is computed, and the LLM-assisted web-search step inside RefChecker is hard to reproduce. Trusting any one tool requires the multi-source, reproducible verification system the paper calls for at the end, and this test is far short of that.
The authors also concede that the large-scale Lancet and Zhao studies cannot prove every fabricated citation came directly from an LLM. Paper mills and deliberate manipulation are plausible sources too. Treating hallucinated citation as one unified frame, as this paper does, blurs the causes.