Responsible use of chatbots to advance the laboratory hematology scientific literature: Challenges and opportunities

Abstract

Large language models (LLMs) are artificial neural networks, a particular type of artificial intelligence (AI), trained to recognize complex patterns in natural language or text to generate written content similar to what a human would produce.1 Chatbots are specific LLM-based applications or interfaces designed to simulate human conversation, generating prompt responses to a variety of inquiries. Chatbots are now offered by several developers, but often differ in data set processing, architecture, training, number of parameters, cost, licensing, and ease of availability, among others.1 Chat Generative Pre-trained Transformer (ChatGPT, Open AI, San Francisco, CA, USA), publicly released in 2022, is a widely known chatbot that can respond to questions in English and other languages, is straightforward to operate, and can be easily used by nonexperts (Available at: https://chat.openai.com/chat [accessed March 11, 2024]). As ChatGPT and other chatbots have significant potential to alter how medical and scientific information is taught, communicated, and disseminated, they have become the objects of increased interest and scrutiny in scientific and technological fields. An emerging body of literature highlights ChatGPT's diverse medical applications, including medical education and practices, patient care, and interactions among healthcare professionals, patients, and data, as well as its limitations.1-5 Although there is currently little information regarding the use (and abuse) of LLM tools in laboratory hematology, promising applications in general laboratory medicine include domain-specific education, assistance in interpreting or integrating laboratory results, facilitation of communication between the laboratory, patient-care teams, and patients, as well as support of laboratory operations and regulatory/compliance inquiries.6, 7 However, there are concerns regarding their general accuracy, repeatability, and ability to interpret complex data sets or situations. For example, a recent study by the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group on Artificial Intelligence (WG-AI) demonstrated that although ChatGPT recognized simulated laboratory reports of common parameters and detected whether results deviated from the reference intervals (RI), overall interpretations were superficial, occasionally incorrect, and mostly judged to be incoherent, and no meaningful suggestion regarding follow-up diagnostics or further procedures were made.8 In this editorial, we would like to briefly explore the potential roles of chatbots in scientific manuscript preparation with particular emphasis on their potential impact on the laboratory hematology literature. Assistance by ChatGPT and other chatbots is being recognized as a useful resource at various stages of the research and publication process and can support planning, preparation, and editing of different manuscript types, including case reports, research papers, and systematic reviews (Table 1).9-11 For example, if prompted with appropriately worded questions and context, ChatGPT is capable of suggesting an appropriate study design or framework. Chatbots are also emerging as helpful tools in developing an appropriate literature search strategy,12-14 in assessing the results of the search15 or the quality of papers,14 and in extracting data from papers for further synthesis and analysis.14, 16 A recent concise review and SWOT analysis on the role of ChatGPT in the medical literature during the first 3 months after its release emphasized that it excels in its ability to express ideas and formulate general contexts clearly and comprehensibly, thereby offering an opportunity to assist non-native English speakers in the medical writing process.11 Chatbots have therefore been employed to assist in coordinating between individuals or groups involved in collaborative projects, including studies produced by the COVIDSurg Collaborative the GlobalSurg Collaborative, and the National Institute for Health and Care Research Global Health Research Unit on Global Surgery.12 However, employing chatbots to create specific scientific literature is associated with several limitations. In a systematic review, ChatGPT was found to occasionally perform so well that even experts in the field had difficulty identifying abstracts generated by ChatGPT, while on the other hand exhibiting a time-limited scope and substantial need for corrections.11 Other shortcomings include the production of texts that have a superficially fluent appearance but contain meaningless collections of medical terms, described as “artificial hallucinations,”17 as well as including faulty and/or fictitious references18, 19 and inaccurate conclusions.12, 13, 20, 21 Also, because chatbots are in general developed by for-profit companies, there is a lack of transparency in their methodology (“black boxes”), making the assessment of generated data less transparent to the untrained user.14, 21, 22 There is currently no clear consensus whether AI-detecting programs commonly used by journal publishers can reliably detect texts prepared by ChatGPT.23-26 Because of the ease of use, manuscripts prepared with this chatbot are already appearing in the literature,27 raising the question of whether chatbots such as ChatGPT can be authors of scientific papers. Statements have been issued by journals such as Nature and Science,28, 29 and are surrounded by active ethical debates and the increasing need for innovation and federation of knowledge—all of which also need to be addressed by the laboratory medicine community. A fundamental question is the degree to which chatbot assistance is allowable, given the issues regarding transparency of chatbot methodology and the quality of chatbot-generated text. Nevertheless, these texts may support the writing process by serving as preliminary drafts or outlines that can be further developed by human intelligence, critical thinking, and reasoning.17 AI assistance may also extend into the realm of peer review: as such, a recent review article highlighted the use of LLMs to assist peer reviewers in writing constructive reports that provide useful feedback to authors, an approach that could greatly facilitate the often time-consuming and cumbersome manuscript review process.30 Again, as with manuscript preparation, the lack of transparency of the chatbot process and the unclear quality of the output are important limitations to the use of LLMs in peer review, and clear instructions and expectations need to be communicated by journals, editors, and the community alike, including the need for critical evaluation of LLM-provided content.30 One area where chatbots are likely to add benefit to the medical literature is by improving the overall quality of medical writing. As manuscripts are submitted by authors practicing around the world, including in countries where languages other than English are spoken, a lack of native-level English fluency may be a barrier for many researchers.31 Although for-profit companies have exploited this niche and market their editing services to authors, researchers practicing in resource-limited environments may be unable to access good-quality English language editing. For this reason, AI-assisted scientific writing or proofing may democratize the scientific literature and facilitate the production of manuscripts from non-native speakers in high-quality idiomatic English that is appropriate for a scientific publication.14, 24 This might range from situations that are ethically unambiguous and would be essentially indistinguishable from using spelling/grammar checkers in word processing programs to more problematic situations where AI assistance might alter the context or content substantially, thereby creating factually wrong or nonsensical scientific text.14 Chatbots could also do the work of expert librarians, who routinely assist in data synthesis and systematic reviews, and could be an additional advantage to researchers working under limited circumstances.14, 16 Chatbot-assisted manuscript preparation is likely to be a permanent fixture of the medical literature, and a range of open access chatbots in addition to ChatGPT has already entered the market, including models specifically tailored to domain-specific knowledge bases. Based on the current state of the literature on this topic, it will be important for researchers in laboratory hematology and related fields to be responsible stewards of the data generated and facilitated by chatbots, and to report their methods and findings in a transparent manner. Chatbots carry great potential to enhance the writing experience, while being prone to misuse that might in part be driven by a deep enthusiasm for AI tools and the promise they hold in our professional and personal lives. Users should therefore be cognizant of the risk of misinformation, inconsistencies, and lack of human-like reasoning abilities in the use of LLMs. Similarly, the entire community including reviewers and journal editors must remain vigilant yet open to both risks and possibilities of LLMs and remain at the forefront of ethical discussions and evaluations. The strengths and weaknesses of this approach and the ethical implications of chatbot assistance will continue to be debated as AI-based systems become increasingly sophisticated. Chatbots using text only are just the beginning phases of development, and additional inputs to allow for the integration of images, videos, or other data will greatly expand what answers we will be getting—for better, or for worse. The authors declare no conflict of interest. Data sharing not applicable to this article as no data sets were generated or analyzed during the current study.

Ähnliche Arbeiten