Introduction
The rapid proliferation of large language models (LLMs) has transformed numerous fields, including biomedicine and bioinformatics, where they offer potential for accelerating literature synthesis, reference retrieval, and knowledge discovery.1,2 LLMs can assist in literature synthesis tasks such as generating relevant citations for research articles.3,4 LLM-based ensemble strategies with retrieval-augmented generation prompts can also help literature screening.5 However, LLMs often produce plausible but fabricated or inaccurate content, including fictitious references and incorrect digital object identifiers (DOIs).6-9 Some scholars refer to this phenomenon as referential hallucination.7,10-13
High rates of referential hallucination have been documented across early LLM versions, with studies reporting fabrication in 50%–80% of generated citations in biomedical contexts.13,14 Literature search errors have been observed not only in factual claims but also in bibliographic metadata, including DOIs and journal citations.7,13 This observation raises concerns about reliability in high-stakes applications like medical research, where inaccurate references could propagate errors or undermine scholarly integrity.15,16 However, there are few quantitative comparative or association analyses of the error/hallucination rates of contemporary LLMs.
Therefore, we quantified the errors of five widely used free-version LLMs (Grok, ChatGPT, Google Gemini, Perplexity AI, and DeepSeek) in searching and retrieving references relevant to original research articles published in four leading medical journals, namely British Medical Journal (BMJ), Journal of the American Medical Association (JAMA), The Lancet, and The New England Journal of Medicine (NEJM). These leading medical journals were chosen to present high-quality abstracts and scientific content in general medicine. By randomly sampling articles from January 2024 to July 2025 and rigorously validating 2,000 retrieved references using a multimetric framework, we aimed to provide an objective and reproducible assessment of errors in LLM-assisted general medical literature search and retrieval. These errors include fabricated references, real references with incorrect bibliographic data and irrelevant real references with correct bibliographic data. Our findings highlight significant variability in LLM performance and provide insights into the factors associated with these errors.
Materials and methods
Random selection of original research articles
Five original research articles published in 2024 and 5 in early 2025 (January 1 to July 15) were randomly selected from each of the four leading medical journals: BMJ, JAMA, Lancet, and NEJM (Fig. 1). Specific inclusion and exclusion criteria were thus not used due to the random selection process. Specifically, all original articles published in these journals in 2024 and early 2025 were assigned a number sequentially according to their PubMed indexing times, separately for each journal and publication year. For each journal, we used Google’s random number generator to generate five random numbers for the 2024 cohort and another five for the 2025 cohort. Only the articles corresponding to these random numbers were included in the analysis.
Retrieval of relevant articles using LLMs
For each of the 40 original research articles, we prompted 5 commonly used LLMs (all free versions) with the article’s abstract to retrieve 10 relevant references between July 15 and August 30, 2025. The LLMs were Grok (Grok-2), ChatGPT (GPT-4o), Google Gemini (2.5 Flash), Perplexity AI, and DeepSeek (V3). Each query focused on one article at a time. All 5 LLMs were accessed through free accounts without paid subscriptions and were evaluated under identical settings. The same web browser (Safari 18.6) without plug-ins or user-activated web search was applied to all 5 platforms at a given time.
The prompt was “Generate a list of 10 key references related to this abstract with the title, year published, DOI, PubMed ID, and Google Scholar URL; no description; clearly indented for each reference” with the article’s abstract appended to the prompt. We then instructed the LLMs to confirm these bibliographic identifiers using the prompt “Please confirm the DOI, PubMed ID and Google Scholar link.” The results confirmed by the platform were recorded and subjected to manual checking. The same prompting process was repeated in all experiments.
Manual validation of retrieved references
Four validity metrics were manually assessed using two strategies (Fig. 1).
One reviewer first searched the DOI registry (doi.org), PubMed (pubmed.gov), and Google Scholar with the retrieved reference’s title and authors. If a DOI, PubMed ID or Google Scholar link was found and matched to that provided by the LLM, one point was assigned to the respective metric. If found but not matched, no points were assigned. If no entry was found, the metric was considered not applicable. For example, an engineering reference/article understandably might not be indexed in PubMed and hence had no PubMed ID. Three of the four metrics were assigned using this strategy.
In the other strategy, one reviewer compared the retrieved references with the title and reference list of the index article to assign a relevance score. If the title and first author of the retrieved reference were found in the reference list, one relevance point was assigned. If not found, no points were assigned. However, if the retrieved reference was the index article, two points were assigned to distinguish retrieval of the index article from retrieval of other relevant references.
A second reviewer independently evaluated 500 randomly selected retrieved references (100 from each platform) between December 20 and December 30, 2025. The scores were compared with the original scores. A score difference in any of the 4 metrics was considered a scoring discrepancy and was resolved by the two reviewers via discussion and fact-checking. No major LLM version changes occurred, while DeepSeek had a minor version upgrade (from V3.1 to V3.2).
Calculation of the score ratio and complete miss status
Because some retrieved references had nonapplicable metrics, we used a composite ratio and complete miss status to more objectively evaluate the retrieved references.
First, we generated the total score by summing the scores for DOI, PubMed ID, Google Scholar link and relevance score for each retrieved reference (Fig. 1). Second, we generated the score cap by counting the number of applicable validity metrics. The retrieved references with nonapplicable metrics then had a lower maximum score than those without. Finally, we divided the total score by the score cap to generate the score ratio for each retrieved reference. For example, one reference retrieved by ChatGPT for a BMJ article was indeed the index article. DOI, PubMed ID and Google Scholar link were all correct. Therefore, the reference’s score cap was 4, while its total score was 5 including a bonus point for relevance. The score ratio was thus 1.25 (5 divided by 4). To highlight the retrieved references without any validated metrics, the complete miss status was assigned if the retrieved reference failed all metrics, indicating that none of the applicable validity metrics received a point.
Regression analyses and statistics
To account for the hierarchical clustering of multiple retrieved references within individual index articles and LLMs, we used mixed-effects multivariable linear and logistic regression models for the score ratio and complete miss outcomes, respectively. Index articles and clusters (article-by-LLM-platform) were modeled with random intercepts to account for baseline article-level and cluster-level heterogeneity. The fixed-effect factors of interest were journal, publication year, and LLM. These factors were evaluated as fixed predictors to quantify their specific associations if they met the inclusion criterion. We included only the factors potentially associated with either of the two outcomes in univariable regression. This liberal screening threshold of P < 0.25 was used according to the purposeful-selection approach described previously.17,18 A likelihood-ratio test was then used to examine a nested random intercept for LLM within article.
The chi-square or Fisher’s exact test was used to compare categorical data by metrics. We performed all statistical analyses and plotting with Stata (version 18; StataCorp LLC, College Station, TX, USA) and defined statistical significance using P < 0.05.
Results
Random selection of index articles
The 4 leading medical journals published 1,187 original articles from January 1 to December 31, 2024 and 692 from January 01 to July 15, 2025, from which we randomly selected 40 original articles (Fig. 1). The two reviewers had score differences in 9 (1.8%) of the 500 retrieved references reviewed by both. Their agreement rate was 98.2%. There were 4 score differences in Google Scholar link including 3 correct links mislabeled as incorrect and 1 incorrect link mislabeled as correct (net gain of 2 correct ones), 3 in PubMed ID and 2 in DOI.
Score ratio
The average score ratio of the 5 LLMs (n = 400 for each platform) was 0.29 (standard deviation, 0.35; range, 0–1.25), with a higher score ratio indicating greater accuracy in retrieving bibliographic data of medical literature. The highest and lowest score ratios were achieved by Grok (0.57) and Gemini (0.11), respectively, while the journal BMJ (0.33) appeared to have a higher score ratio than Lancet (0.29), NEJM (0.25) and JAMA (0.29).
Journals and LLMs were both independently associated with the score ratio in multivariable regression analyses (Table 1), while publication year was not associated with score ratio even in univariable regression. Compared with BMJ, NEJM was associated with a lower score ratio (coefficient = −0.08, 95% confidence interval [CI], −0.14 to −0.02; P = 0.013), but no significant differences were observed for JAMA or Lancet (coefficient = −0.04, 95% CI, −0.11 to 0.02; P = 0.187 for JAMA and coefficient = −0.06, 95% CI, −0.12 to 0.00; P = 0.065 for Lancet). Grok was linked to a higher score ratio than ChatGPT (coefficient = 0.14, 95% CI, 0.07 to 0.21; P < 0.001), while all other platforms were associated with lower score ratios (coefficient = −0.21, 95% CI, −0.28 to −0.14 for DeepSeek, coefficient = −0.31, 95% CI, −0.38 to −0.24 for Gemini, and coefficient = −0.30, 95% CI, −0.36 to −0.23 for Perplexity; P < 0.001 for all).
Table 1Multilevel mixed-effects multivariable regression on the associations of LLM and medical journal with the score ratio
| Coefficient | 95% CI | P |
|---|
| LLM | | |
| ChatGPT | Reference | |
| DeepSeek | −0.21 | (−0.28 to −0.14) | <0.001 |
| Gemini | −0.31 | (−0.38 to −0.24) | <0.001 |
| Grok | 0.14 | (0.07 to 0.21) | <0.001 |
| Perplexity | −0.30 | (−0.36 to −0.23) | <0.001 |
| Journal | | | |
| BMJ | Reference | | |
| JAMA | −0.04 | (−0.11 to 0.02) | 0.187 |
| Lancet | −0.06 | (−0.12 to 0.00) | 0.065 |
| NEJM | −0.08 | (−0.14 to −0.02) | 0.013 |
The likelihood-ratio test on the models for score ratio showed a nested random intercept for LLM within article (chi-square = 170.29, likelihood-ratio P < 0.001). The cluster-level intraclass correlation coefficient (ICC) was 0.21 (95% CI 0.16–0.26), while the article-level ICC was <0.01 (95% CI <0.01–0.95) and was considered trivial.
There were differences in adjusted associations between LLMs and the score ratio in most pairwise comparisons by LLMs (Fig. 2). All of our pairwise comparisons were statistically significant except Gemini versus Perplexity (P = 0.528).
Complete miss rate
References retrieved by LLMs failed all metrics in 48.7% of cases, as indicated by the mean complete miss rate. The lowest (best) complete miss rate was achieved by Grok (11.3%), and the highest (worst) by Gemini (78.5%). NEJM had the highest complete miss rate (56.0%), whereas rates for the other journals were similar.
Journals and LLMs were associated with complete miss rate in univariable regression, but publication year was not. Our multivariable regression analyses showed that LLMs, but not journals, were associated with complete misses (Table 2). Compared with ChatGPT, Grok was associated with a lower complete miss rate (better performance, odds ratio (OR) = 0.23, 95% CI, 0.11–0.46; P < 0.001), but others were associated with higher rates (worse performance, OR = 3.49, 95% CI, 1.87–6.50 for DeepSeek, OR = 16.86, 95% CI, 8.69–32.71 for Gemini, and OR = 11.01, 95% CI, 5.79–20.94 for Perplexity; P < 0.001 for all).
Table 2Multilevel mixed-effects multivariable logistic regression on the associations of LLM and medical journal with complete miss
| Odds ratio | 95% CI | P |
|---|
| LLM | | |
| ChatGPT | Reference | |
| DeepSeek | 3.49 | (1.87–6.50) | <0.001 |
| Gemini | 16.86 | (8.69–32.71) | <0.001 |
| Grok | 0.23 | (0.11–0.46) | <0.001 |
| Perplexity | 11.01 | (5.79–20.94) | <0.001 |
| Journal | | | |
| BMJ | Reference | | |
| JAMA | 0.88 | (0.48–1.60) | 0.675 |
| Lancet | 0.87 | (0.48–1.58) | 0.657 |
| NEJM | 1.64 | (0.90–3.00) | 0.104 |
The likelihood-ratio test of the models for complete miss also showed a nested random intercept for LLM within article (chi-square = 119.97, likelihood-ratio P < 0.001). The cluster-level and article-level ICCs were 0.30 (95% CI, 0.23–0.38) and <0.01 (95% CI, <0.01–0.99), respectively. Given its small value and wide 95% CI, article-level ICC should be considered trivial and characterized by substantial uncertainty.
Individual validity metrics
Across the 400 references retrieved by each LLM (100 per journal), the performance of LLMs varied significantly across all individual validity metrics according to journal or publication year (Table 3), except relevance score by publication year.
Table 3Characteristics of LLMs’ performance in retrieving references relevant to the index articles randomly selected from four leading medical journals
| ChatGPT | DeepSeek | Gemini | Grok | Perplexity | Total | P value |
|---|
| DOI | | | | | | | |
| Journal | | | | | | | |
| BMJ (n = 100) | 62 | 23 | 3 | 73 | 9 | 170 | <0.001 |
| JAMA (n = 100) | 49 | 13 | 13 | 72 | 19 | 166 | |
| Lancet (n = 100) | 35 | 23 | 30 | 53 | 22 | 163 | |
| NEJM (n = 100) | 47 | 11 | 7 | 55 | 5 | 125 | |
| Year | | | | | | | |
| 2024 (n = 200) | 118 | 26 | 25 | 111 | 32 | 312 | 0.001 |
| 2025 (n = 200) | 75 | 44 | 28 | 142 | 23 | 312 | |
| PubMed ID | | | | | | | |
| Journal | | | | | | | |
| BMJ (n = 100) | 47 | 13 | 0 | 50 | 5 | 115 | <0.001 |
| JAMA (n = 100) | 25 | 4 | 4 | 29 | 12 | 74 | |
| Lancet (n = 100) | 31 | 10 | 19 | 34 | 10 | 104 | |
| NEJM (n = 100) | 26 | 6 | 4 | 33 | 8 | 77 | |
| Year | | | | | | | |
| 2024 (n = 200) | 73 | 9 | 16 | 66 | 24 | 188 | 0.003 |
| 2025 (n = 200) | 56 | 24 | 11 | 80 | 11 | 182 | |
| Google Scholar link | | | | | | | |
| Journal | | | | | | | |
| BMJ (n = 100) | 78 | 69 | 0 | 88 | 13 | 248 | <0.001 |
| JAMA (n = 100) | 66 | 31 | 4 | 92 | 25 | 218 | |
| Lancet (n = 100) | 16 | 43 | 40 | 86 | 5 | 190 | |
| NEJM (n = 100) | 60 | 36 | 12 | 80 | 4 | 192 | |
| Year | | | | | | | |
| 2024 (n = 200) | 115 | 86 | 39 | 166 | 16 | 422 | 0.005 |
| 2025 (n = 200) | 105 | 93 | 17 | 180 | 31 | 426 | |
| Relevance score | | | | | | | |
| Journal | | | | | | | |
| BMJ (n = 100) | 22 | 16 | 3 | 46 | 7 | 94 | 0.048 |
| JAMA (n = 100) | 26 | 12 | 9 | 32 | 12 | 91 | |
| Lancet (n = 100) | 12 | 5 | 9 | 15 | 12 | 53 | |
| NEJM (n = 100) | 21 | 9 | 9 | 25 | 8 | 72 | |
| Year | | | | | | | |
| 2024 (n = 200) | 41 | 21 | 20 | 65 | 21 | 168 | 0.621 |
| 2025 (n = 200) | 40 | 21 | 10 | 53 | 18 | 142 | |
LLMs significantly differed in retrieving valid DOIs, PubMed IDs, and Google Scholar links by journal (P < 0.001 for all) and publication year (P = 0.001, 0.003 and 0.005), respectively. They also differed in relevance scores by journal (P = 0.048), but not publication year (P = 0.621).
Among the 4 journals, Gemini retrieved accurate article bibliographic data least frequently for BMJ (0/100 for both PubMed ID and Google Scholar link) and JAMA (4/100 for PubMed ID and 4/100 for Google Scholar link). In contrast, Grok and ChatGPT overall retrieved accurate data more frequently than other LLMs.
Discussion
This quantitative comparative study of LLM-assisted medical literature showed modest overall performance of LLMs and significant variability in retrieval accuracy across platforms and journals. These findings align with prior reports documenting high rates of fabricated or incorrect citations in LLM outputs,7,8,13 and extend them by demonstrating that such issues persist even when prompts explicitly request structured bibliographic metadata and data verification. We also identified associations between LLMs and overall errors in LLM-assisted literature retrieval. Future studies are needed to differentiate and analyze the errors involving fabricated references, real references with incorrect bibliographic data and irrelevant real references with correct bibliographic data.
Our findings have several implications for LLM-assisted medical literature research. First, LLM-assisted medical reference generation and retrieval should be manually verified, particularly when relying on a single platform. Reference metadata, such as DOI, PubMed ID and Google Scholar links, are often inaccurate. That is why we included these three commonly used metadata elements in the composite score, while adding the relevance score to assess the retrieved article’s relevance. The bonus point assigned to the retrieved index article emphasizes the importance of retrieving the most relevant article (i.e., the index one). Variability in LLM performance, as shown here and in previous studies,4,19 also highlights the value of using multiple LLMs at a time. This practice will not only provide more references but also offer opportunities for cross-checking.
Second, bibliographic data should always be checked even in the LLM-retrieved references with correct titles and sound relevance. Earlier studies have reported the retrieval of irrelevant or less-relevant references by LLMs.10 Attention has therefore focused on the relevance of retrieved references. However, we show that contemporary LLMs approximate topical relevance modestly but struggle more with retrieving bibliographic data. The frequent retrieval of plausible references with incorrect bibliographic data could be misleading, offer false assurance, and reinforce the ethical and practical concerns about unsupervised use of LLM-retrieved references.15,20
Third, two index-article publication-year cohorts, 2024 and 2025, were used to account for possible association between publication timing (latency) and retrieval performance. However, our univariable regression analyses showed no association of publication year with score ratio or complete miss rate. Publication year was thus not included in the multivariable regression analyses and was unlikely to be associated with retrieval accuracy.
Limitations
This study has limitations. We evaluated only free versions of LLMs and focused on a simple prompt design, which may not capture the full range of model capabilities. LLM capabilities also evolve rapidly and may have reduced their errors as this study is being reviewed. Moreover, our relevance scoring, while systematic and mostly objective (e.g., citation by the index article), involved human judgment and may be subjective and biased. Citation by the index article is objective but may be too specific and restrictive, since some relevant articles may not be cited due to space limitations. However, additional manual assessment may introduce subjectivity and biases. Furthermore, only leading general medical journals were used here. Our findings thus cannot be generalized to other medical specialties without validation. They may not be directly applicable to the abstracts in middle- or lower-tier medical journals, either, due to differences in abstract quality and editorial preferences. In addition, we did not systematically examine the influence of integrated web search on free-tier LLMs, while recent studies show that web search may help improve performance of LLMs.21,22 Besides the limited scope of this study, the key reason is that the examined LLMs differ significantly by web search availability. Perplexity and ChatGPT offer paid integration of web search. DeepSeek requires users to take manual action (toggle the button) for its free web search function, while Grok and Gemini appeared to integrate the search through X and Google for free, respectively. Our choice of leaving out user-activated web search therefore provides a desired baseline assessment of these platforms, while future work is needed. The inherent differences in LLMs’ database coverage of DOI, PubMed, and Google Scholar may also affect the interpretation of individual validity metrics. Finally, only abstracts of the index articles were subject to the LLM search but may not be the best prompting strategy. However, recent studies show that abstract-based and compressed prompts lead to more accurate literature retrieval.23-25 We believe that abstract-based prompting is more efficient and may be more accurate than the full-text-based approach. Future studies using full articles and other prompting strategies will be interesting and warranted, but are beyond the scope of this study. The findings should therefore be interpreted within the evaluated platforms, prompts, journals, and study periods.
Future directions
Explicit prompts to double-check references were used here but appeared ineffective in improving retrieval accuracy. Therefore, such prompts should be used with caution. More research is needed to better understand and improve prompting in LLM-assisted literature retrieval. One closely related question is whether abstract is more beneficial for LLM-assisted literature search and retrieval than full-text in medical fields. Another question is how to best use yet safeguard LLM-assisted literature search and retrieval in medical and other fields. It will also be interesting to compare the performances of paid versions of LLMs with those of free ones, while the paid versions may afford the users more tokens, longer uninterrupted runs and better LLMs than the free ones.
Conclusions
While LLMs hold promise as tools for biomedical literature exploration, their performance in the medical literature search and retrieval tasks evaluated here is highly variable and often unreliable. Until robust hallucination mitigation, verification systems, and trustworthy AI frameworks are widely implemented, LLMs should be used as assistive, not authoritative, tools in biomedical scholarship.
Declarations
Funding
This work was supported by the National Cancer Institute, National Institutes of Health (R37CA277812 to LZ). The funder of the study had no role in study design, data collection, analysis, interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
Conflict of interest
Lanjing Zhang is a co-editor-in-chief of this journal. The authors declare no other conflict of interest, particularly with commercial LLMs.
Author contributions
Study conceptualization and design, ensuring data access, accuracy, and integrity (LZ), data collection, data analysis, and reference validation (JG, LZ), manuscript writing and revision (JG, LZ, YZ, MLD). All authors approved the final publication version.
Ethical statement
This study did not involve human participants.
Data sharing statement
The full data are available from the corresponding author upon reasonable request.