Rachel Miles summarizes a recently published journal article that cautions researchers on the use of automated citation recommendation tools, which are designed to help academic authors cite literature in the writing phase of their research project while potentially skipping crucial steps of searching, reviewing, and appraising literature.
In an increasingly competitive academic environment, researchers are understandably keen to find smarter, faster ways to complete their projects. However, faster does not always mean smarter. One daunting task all researchers face is finding, managing and tracking citations for their manuscripts. As a result, citation management tools have been developed, such as Mendeley, Zotero, and EndNote. This was followed by the emergence of literature mapping tools to help with literature searching, such as LitMaps, Citation Gecko, Inciteful, and Connected Papers (see this blog post for an in-depth review of such tools). In order to use these literature mapping tools, one must first have a seed paper to plug into the tool which then finds other relevant papers based on algorithms of co-citations and bibliographic coupling. More recently, citation recommendation tools have been developed, such as Citeomatic and its successor Specter. While promising to accelerate the writing process, citation recommendation tools come with major drawbacks, prompting questioning of what exactly are their benefits and perhaps unintended implications of using.
Collectively, literature mapping tools, databases, and search engines help researchers at the literature search phase and act as paper recommendation tools; they use the input of search terms or seed paper(s) to find relevant results; search engines and databases rely on the user inputting search terms, whereas literature mapping tools rely on the user inputting a seed paper to generate relevant papers using a similarity algorithm based on bibliographic coupling and co-citations.
In contrast, automated citation recommendation tools find relevant literature based on the input of a piece of text, such as a statement or paragraph, and then retrieve relevant citations to back the claims made by the author. According to a recent article published in Research Evaluation , automated citation recommendation tools replace a crucial stage of the research process: the literature searching and critical appraisal of the literature, and they facilitate a rather thoughtless way of citing, requiring little scrutiny from the authors. The infographic below demonstrates how automated citation recommendation tools can be implemented in an accelerated research process, omitting several phases of a more valid/rigorous/systematic/reliable process. The tools included in the bulleted lists of the infographic are merely examples and not exhaustive.
During the writing stage of a scholarly manuscript, the citations in the literature review serve a crucial purpose in supporting or disputing the authors’ own results or conclusions. They can also provide background information, identify methodology, identify original research and/or seminal articles, criticize a previous work or correct it, and substantiate claims, among others [2, 3]. The theories behind motivations for citing practices vary, but overall, there is an expectation that authors will provide high quality citations to accurately represent the literature on their research focus or topic. Although literature review sections of original research papers do not have to meet the strict requirements of systematic reviews and meta-analyses, editors, peer reviewers, and readers assume authors are thorough and diligent in conducting their literature reviews.
With the use of automated citation recommendation tools, authors can rely on software to automatically cherry pick the literature to support their claims without actually reading and reviewing the papers they retrieve.
Distressingly, the peer review and editorial process does not typically review papers for the accuracy and reliability of their citations, a tedious and time-consuming task. Instead, editors, peer reviewers, and readers assume authors fulfill this crucial responsibility.
Encouraging questionable citations
While the social effects of academic incentives and reward structures cause questionable citation practices [4, 5], automated citation recommendation tools embolden such practices through the following:
Lazy or perfunctory citing
If a piece of text is used as input to find relevant literature, the algorithm of the tool will tend to find literature to substantiate the claims of the text. Essentially, the author can relinquish their responsibility to systematically search and identify relevant literature, skip the literature search phase, and write their paper more quickly while padding the reference section with the help of the citation recommendation tool.
Previous research  shows that citation recommendation tools narrow their attention on a subset of the literature that substantiates the authors’ claims, ignores contradicting literature, and focuses on similar research within a specific field or community. Thus, these tools tend to amplify existing (positivity) biases and may even introduce new ones.
A new Matthew Effect
The Matthew Effect, in which advantage begets more advantage, is exacerbated by automated citation recommendation tools. These tools’ algorithms are based on current and past citation practices and function similarly to Google’s PageRankTM, retrieving the most “popular” results rather than the most relevant or accurate ones. In academia, scholarship from the Global North, in the English language, from high impact or top ranked journals, and scholarship with higher citation counts may be further benefited by these recommendation tools whereas publications in non-English languages, from the Global South, and in newer or emerging journals will be disproportionately disadvantaged.
Lack of transparency
The complexity of an algorithm to accurately identify relevant citations based on manuscript text introduces a few issues. First, the complexity alone would require considerable investment, development, and maintenance, which may inevitably depend on a business model that would favor certain results over others for its own gain. Second, the complexity would also make it difficult to demonstrate why certain papers are recommended while others are not. Third, there is a risk of well-resourced publishers, journals, and individuals gaming the algorithm to skew attention towards their own papers.
Although authors may be tempted by the convenience of citation recommendation tools, they should proceed with caution or avoidance of using such tools. Searching for literature before designing a research project is part of the research process, and attempting to pad a reference list to support claims and conclusions of an already finished research project is sloppy and irresponsible. Such practices shift the burden of vetting the reliability of the research literature to the reader, which is dangerous territory for other researchers, journalists and the media, clinicians, and members of the public, who are consuming the literature through either primary (e.g., research articles) or secondary sources (e.g., news articles, YouTube videos, Wikipedia, etc.). Therefore, implementation and marketing of such tools should be approached cautiously to avoid potentially boarding an already sinking ship.
The author would like to thank Dr. Serge Horbach, lead author of the journal article summarized for this blog post, for his valuable feedback and suggestions. She would also like to thank Robyn Price, the LIS-Bibliometrics Committee Co-Chair, and Cozette Comer, the Evidence Synthesis Librarian at Virginia Tech, for their reviews and feedback on this blog post.
- Horbach, S. P. J. M., Oude Maatman, F. J. W., Halffman, W., & Hepkema, W. M. (2022). Automated citation recommendation tools encourage questionable citations. Research Evaluation, 31(3), 321–325. https://doi.org/10.1093/reseval/rvac016 or https://pure.au.dk/portal/files/269565745/Citation_Recommendation_Tools_Author_version.pdf [Author’s Accepted Manuscript]
- Garfield, E. (1964). Can Citation Indexing Be Automated? Statistical Association Methods for Mechanized Documentation Symposium Proceedings, 1, 189–192. https://www.govinfo.gov/content/pkg/GOVPUB-C13-14e5c8213b5ad089bb2a05f24cc09e94/pdf/GOVPUB-C13-14e5c8213b5ad089bb2a05f24cc09e94.pdf
- Small, H. G. (1984). Citation context analysis. In B. Dervin & J. M. Voigt (Eds.), Progress in communication science (Vol. 3, pp. 287–310). Ablex.
- Horbach, S., Aagaard, K., & Schneider, J. W. (2021). Meta-Research: How problematic citing practices distort science. MetaArXiv. https://doi.org/10.31222/osf.io/aqyhg
- Edwards, M. A., & Roy, S. (2017). Academic Research in the 21st Century: Maintaining Scientific Integrity in a Climate of Perverse Incentives and Hypercompetition. Environmental Engineering Science, 34(1), 51–61. https://doi.org/10.1089/ees.2016.0223
- Färber, M., & Jatowt, A. (2020). Citation recommendation: Approaches and datasets. International Journal on Digital Libraries, 21(4), 375–405. https://doi.org/10.1007/s00799-020-00288-2
Rachel Miles is the Research Impact Librarian at Virginia Tech. She is the blog manager of The Bibliomagician for the LIS-Bibliometrics Committee. She assists administrators and researchers with research assessment, analytics, and communication. She also grounds her work in fair, ethical, and responsible research assessment and has taken efforts in university governance to advocate for equitable and responsible research assessment.
Unless it states other wise, the content of the Bibliomagician is licensed under a Creative Commons Attribution 4.0 International License.