Leveraging generative artificial intelligence errors to teach appropriate citation usage Authors: Rishita Shah, JJ L. Miranda Published: 2026 ABSTRACT ABSTRACTEducators would benefit from developing course material that increases understanding of generative artificial intelligence (AI). General-purpose large language models (LLMs) popular among students often handle citations incorrectly. These LLMs produce fabricated references referred to as hallucinations and often cite even real citations erroneously. These citation hallucinations and errors offer an opportunity to correct mistakes, understand reference interpretation, and learn appropriate source type usage. Here we repurpose generative AI errors as examples for teaching the rigorous citation of primary and secondary sources in scientific writing. The designed exercise simultaneously achieves learning objectives with regard to AI and information literacy. Seeing the factual errors in writing and citation usage firsthand demonstrates a key weakness in generative AI. Also, our exercise gives students a structure in which to practice appropriate source type usage. The errors produced by generative AI provide opportunities to practice and improve scientific writing skills. Our article proposes a conceptual framework within which to implement this approach in the classroom. INTRODUCTION Generative artificial intelligence (AI) in the form general-purpose large language models (LLMs) has disrupted the science education landscape. In many use cases, the benefits of these LLMs are questionable because of critical weaknesses. Specifically, citation handling remains an area in which LLMs make notable technical errors. Inappropriate or fake citations often indicate that the text is machine-generated. Rigorous citation discovery and usage, however, is a critical element of expert scientific writing (1). Examples displaying a lack of command of this skill pervade all levels of scholarship—while we observe undergraduate students struggling with senior capstone papers, citation usage errors have even been reported at the gold standard science level of the federal government (https://web.archive.org/web/20250522215748/https://www.whitehouse.gov/wp-content/uploads/2025/05/WH-The-MAHA-Report-Assessment.pdf). Generative AI is likely here to stay, and educators would benefit from developing course material that increases understanding of limitations while simultaneously improving scientific writing proficiency. Citation hallucinations and errors offer an opportunity to correct mistakes, understand reference interpretation, and learn appropriate source type use. General-purpose LLMs frequently generate fabricated references referred to as hallucinations. Research reports a rate of approximately 20 to >90% (2–9). Newer versions may be improving, as ChatGPT, a popular LLM, demonstrated decreases in fabrication rates between versions in comparative studies (8, 10). We also want to emphasize the distinction between fabrication of fake citations and incorrect interpretation of real citations. Even with real references, LLMs often cite them incorrectly. Approximately 30–70% of references were used erroneously (4–6, 9). Here, we repurpose generative AI errors as examples for teaching the rigorous citation of primary and secondary sources in scientific writing. Our article proposes a conceptual framework within which to implement this approach in the classroom. PROCEDURE Learning objectives This exercise is designed to achieve multiple learning objectives at different levels of complexity. After completing this exercise, students should be able to (i) remember how to use citation information to look up a paper, (ii) understand the difference between a primary research paper and a review article, and (iii) analyze if an AI-generated interpretation accurately reflects the data in a published study. Generating text for evaluation We used ChatGPT version GTP-5.1 Instant (https://chatgpt.com), ChatGPT version GTP-5.1 Thinking (https://chatgpt.com), Gemini version 3 Pro (https://gemini.google.com), Grammerly Pro (https://app.grammarly.com), and Perplexity Pro (https://www.perplexity.ai) to generate text for the scientific writing exercise. Other AI chatbots may also be used. The zero-shot instruction-based prompt given follows the format "Write a paragraph about [microbe] [topic]. Include in-text citations and a bibliography." Here, [microbe] is a specific organism, and [topic] is a pathway, structure, or process. Examples include Epstein-Barr virus reactivation, Saccharomyces cerevisiae kinetochores, and Aquifex aeolicus evolution. Students are encouraged to choose prompts for topics with which they are familiar. Prompts may also be selected by an instructor. Although initially designed for microbiology, this activity may involve any realm of biology. The prompt generates text that includes approximately 2–8 citations, a good number for the learning exercise. As an addition to the core exercise, prompt engineering may be utilized to attempt to change the source type used in the generated texts. The instruction-based prompt given can be modified to include addition of the direction "Do not cite review articles." or "Cite only review articles." Citation editing Students evaluate every citation generated by AI to determine if it is the most appropriate for the text. Our decision tree guides this process (Fig. 1). In our seminar classes, we emphasize properly distinguishing between a primary research article and a secondary review article. Flowchart illustrating sequential steps in learning proper citation usage from generative AI mistakes. The decision tree is evaluated once per cited reference. Step components that include primary research articles are emphasized in green; step components that include secondary review articles are emphasized in blue. Ending terminal nodes are shown as a rounded rectangle. Students begin by understanding whether the AI-generated statement of fact describes a single original discovery or multiple different discoveries. Then, students are asked to identify if there is a citation present and if one can confirm it is real. The answer may be no because the citation is a hallucination and does not actually exist. If no, students are directed to replace the citation. In the case of a single original discovery, a research article with the original experimental data is necessary. In the case of multiple different discoveries, a review article with a summary of papers is necessary. We acknowledge that multiple primary research articles can be cited together to account for the different discoveries. We chose to prioritize the option of a review article to give students the experience of identifying secondary sources popularly used in many fields. If the AI-generated citation is real, students next determine whether or not the citation provides the appropriate support for the text. If the statement describes a single original discovery, students ask if the reference contains the original experimental data that established the cited fact. If yes, the citation is validated and kept. If no, students are directed to find a research article with the original experimental data and replace the citation. In many cases, LLMs choose a secondary review article instead of the primary research article to describe a single original discovery. In the editing process, students often find the correct research article in the bibliography of the review article. If the statement describes multiple different discoveries, students ask if the reference contains a summary of papers that each support the cited facts. If yes, the citation is validated and kept. If the answer is no, students are directed to find a review article with a summary of papers and replace the citation. CONCLUSION We recommend pilot testing different LLMs because models vary, are frequently updated, and, therefore, produce a myriad of results. ChatGPT version GTP-5.1 Instant, ChatGPT version GTP-5.1 Thinking, Gemini version 3 Pro, Grammerly Pro, and Perplexity Pro all fabricated citations by generating hallucinations that were the incorrect combination of real authors, titles, journals, and/or volume/issue/page numbers. We observed a wide range of fabrication rates similar to previous systematic studies (2–7, 9, 10). ChatGPT version GPT-5.1 Thinking, Gemini version 3.0 Pro, and Perplexity Pro provided links to the references. ChatGPT version GPT-5.1 Thinking often cited multiple sources in the same parentheses. ChatGPT version GPT-5.1 Thinking and Perplexity Pro often generated >8 citations. Perplexity Pro grouped all in-text citations at the end of the paragraph. Prompt engineering can also alter LLM output. We chose to focus on prohibiting or forcing use of review articles. The LLMs we tested mostly, although not always, substituted citation types as directed. We caution that these were our observations at the time, but the field is rapidly changing. For example, retrieval-augmented generation significantly improves citation accuracy when synthesizing responses that reference recent literature in biomedicine (8). Testing a variety of LLMs is useful because different deficiencies in the results affect how the material can be leveraged to maximize effectiveness of the exercise for critical examination of citation usage. Correctly generated responses with real citations allow focus on appropriate primary versus secondary source type use. Though not directly addressed in our original learning objectives, instructors and students can further identify which model generates the most useful text for a chosen purpose. Potential variations can facilitate or hinder specific learning goals. Fabricated citations with incorrect components could facilitate teaching the constituent elements of a citation. Links to references accelerate validation of the bibliographic information. Multiple citations in the same parentheses are more challenging to evaluate. Large numbers of citations in the paragraph require more time to evaluate. Lack of in-text citations could enable teaching appropriate placement of references. Prompt engineering can increase the number of examples for either primary research articles or secondary review articles as starting material. In addition to the ability to generate useful text for teaching information literacy, we also advise considering student engagement, as we suspect familiarity would strengthen the impact of the exercise on AI literacy. Our work initially focused on ChatGPT because it is the most popular AI chatbot among higher-education students (11). Many companies advertise directly to and offer discounts for students and faculty. Our institution provides Gemini Pro free to our students. Using the lens of Bloom’s taxonomy, a current challenge with AI literacy is moving from the more common stages of knowing, understanding, using, and applying AI toward the next steps of analyzing and evaluating AI (12) (https://er.educause.edu/articles/2024/6/a-framework-for-ai-literacy). We hypothesize that analyzing and evaluating LLMs frequently used in higher education will be more effective at developing and expanding AI literacy. Future directions include a systematic evaluation of the effectiveness of this approach. Our exercise is a hypothesis-driven educational tool and not a proven intervention. The effectiveness of this approach has not yet been quantitatively assessed by prospective studies. Work remains to rigorously determine if this exercise indeed improves appropriate citation type usage of primary and secondary sources in scientific writing. Even if effective, it will also be important to assess if this approach offers advantages over other methods for teaching citation literary. We hope that this exercise provides both the impetus and groundwork for the future study of the rapidly evolving interface between generative AI and scientific writing. The designed exercise simultaneously achieves learning objectives with regard to AI and information literacy. Seeing the factual errors in writing and citation usage firsthand demonstrates a key weakness in generative AI. Also, our exercise gives students a structure in which to practice appropriate source type usage. The majority of first-year science and health science undergraduate students can correctly identify primary research articles but are less adept at recognizing secondary review articles (1). Biology majors are still learning literature-searching skills up to senior year (13). The errors produced by generative AI provide opportunities to practice and improve scientific writing skills.