Articles collecting and analyzing

In the medical field, any new idea regarding pathogenesis, diagnosis, or treatment must be supported by strong evidence from scientific literature [28]. To provide physicians and medical practitioners with the best possible information in the form of reviews or guidelines, authors should perform a comprehensive literature search to familiarize themselves with the field and recent advancements made by other researchers. Integrating AI tools has notably reshaped the traditional approach to this task in qualitative data collection and analysis. A literature review typically begins with a superficial screening followed by a full-text review in search of valuable facts. Identifying the most significant papers in the field of interest might be difficult as the number of documents received in PubMed or other search engines is often large and includes papers vaguely or unrelated to the subject matter. AI can be implemented to improve this process by identifying highly influential papers and demonstrating the connections between them. Tools such as Scholar [29, 30] or SciteAI [30, 31] use text mining to identify important documents and generate short summaries useful during screening.

Rules of evidence-based medicine (EBM) require authors to identify all studies that investigate the patient population of interest. Most commonly, it is performed using the PICO format (P—patients, I—intervention, C—comparison, O—outcomes) [32]. Typically, authors would have to assess whether each study matches the chosen requirements. To facilitate this process, machine learning algorithms can be implemented to predict the relevance of articles according to specific frameworks, such as the PICO format [33, 34]. A tool called the PICO Portal extracts PICO elements directly from the articles with the help of a LLM, thereby improving the screening process [35]. Similar function is also offered by Rayyan, a web and mobile app focused specifically on systematic reviews [36]. However, it is crucial for human authors to verify whether the identification performed by AI was correct.

A proper literature search can result in a large amount of text piling up in front of an author. Despite being recommended, thorough reading might be time-consuming and only sometimes beneficial because just the subpart of initially retracted papers is genuinely significant. AI can optimize literature review by offering a streamlined approach to condensing research papers, thereby reducing the time needed for reading (tool example: Genei) [37]. Summaries received from AI are short and comprehensible; however, it is essential to remain mindful of potential misinformation [38]. Commercially available AI tools are trained on publicly available data, which often excludes articles behind the paywalls of the most impactful journals. Consequently, the responses generated by AI can be derived from incomplete information and, therefore, may be misleading. Additionally, the datasets used by AI are not continuously updated; instead, the data is available only up to a certain point in time. This creates a risk that some novel and crucial findings will not be incorporated into the responses, potentially leading to outdated or incomplete conclusions [39]. To address this issue and prevent possible publication bias it would be advisable to perform secondary confirmation with manual information searching.

Another approach to understanding the articles more straightforwardly is to ask AI questions. Readers can utilize various tools (e.g., ChatPDF [40], Chatwithpdf [41]) to inquire about the study’s primary outcomes, methods, or any part. In this instance, the limiting factor is mainly the clarity of the reader’s question. More tools with similar functions may be practical for literature review (examples: PerplexityAI [30, 42], Elicit [43]). However, choosing the right tool designated for the task is crucial. Using ChatGPT in each case is not always optimal as it is prone to misinterpretation.

In one study [44], ChatGPT was asked to create a summary of a systematic review of cognitive behavioral therapy (CBT). ChatGPT reduced the number of studies used in that article (from 69 to 46) and exaggerated the effectiveness of CBT. Indeed, both GPT-3.5 and GPT-4 need help with citations and are prone to make blunders in the bibliography. AI tends to change the authors of cited articles, journals in which those papers were published, and even the titles. However, it is worth noting that GPT-4 makes fewer mistakes than GPT-3.5, proving that this technology is improving. When fabricated citations were used, GPT-4 achieved a result of 18%, whereas GPT-3.5 up to 55%. Substantial errors in the accurate citations made by GPT-4 occurred in 24% and GPT-3.5 in 43% of all the cases [45]. This technology is still developing, and significant upgrades are required before we can use it without hesitation or doubt. As for today, authors are advised to double-check any information provided by the AI.

Given all these imperfections, a more streamlined and time-efficient literature review is worth considering. The mentioned tools collectively enhance qualitative data collection and analysis and provide ready-to-use information that may be particularly useful at the beginning of the writing process.

Articles writing

From drawing the first draft to refining the finished manuscript, AI tools have emerged as strategic allies offering various functionalities beyond conventional grammar checks. AI can help biomedical researchers in the writing process by advising on the structure of their manuscripts, providing references, or even suggesting an article title [46]. It can be particularly helpful for researchers whose English is the second language to translate biomedical articles or to edit their own work [47].

AI can serve as a writing assistant that improves general language structure and the overall sophistication of written content. This function is represented by several tools, with Grammarly being one of the most popular and having over 30 million users. It is a versatile guide that elevates writing quality through real-time grammar checks and style improvements [48, 49]. TrinkaAI tool offers a similar function mainly designated for academic and technical writing [30, 48]. Quillbot, on the other hand, serves as a paraphrasing tool, ensuring the preservation of originality while offering additional features such as summarization and grammar checking [30, 50]. Utilizing AI-powered tools facilitates writing and has been proven to enhance engagement and self-efficacy in learning [51]. Students generally perceived it well, praising their ease of use [52]. Once the draft is set, the more challenging and time-consuming phase takes place—the refining of the manuscript. An author might seek evaluation on grammar, style, and choice of words. Writefull [30, 53] or Wordtune [54, 55] offer instant feedback. To ensure adherence to journals’ rigorous quality standards, an author can also implement Paperpal, which specializes in academic papers [30, 56]. Finally, JenniAI offers autocomplete in-text citations, providing documents for stated thesis, paraphrasing, and a built-in plagiarism checker [30]. With all these functions, the sluggish and stagnant process of polishing a manuscript can be smoother and more time-efficient.

The benefits arising from using AI are not limited solely to academic writers or medical researchers. Reviewers constitute another group that can utilize AI to save time and enhance the efficiency of their work. For instance, with the help of accessible tools, reviewers can evaluate papers to determine whether they meet specific criteria (tool example: Penelope.ai) [30, 57]. Additionally, during the peer review process, papers may undergo scanning to detect possible plagiarism (tool example: Copyleaks) [58]. Moreover, platforms built on LLMs can quickly and efficiently identify methodological and statistical errors in the evaluated manuscripts (tool example: StatReviewer) [57]. For obvious reasons, such a trait may also benefit writers, as they can polish their manuscript to satisfy the reviewers in the mentioned aspects.

As previously stated, most journal policies do not cover non-generative AI and do not require disclosing information about their usage in the manuscript. Regardless of the lack of obligation to do so, it is worth mentioning it in the method sections or the acknowledgments. Users take full responsibility for the content created or changed by these tools. Therefore, this practice would ensure scientific transparency and encourage authors to assess carefully the quality of their refined work.

Plagiarism and copyrights in AI-generated text

The benefits presented by AI are tempting. However, additional questions must be addressed to ensure that AI usage remains ethical and that the generated content is unbiased. Since AI tools offer a wide range of content-generating capabilities, plagiarism is a significant and increasing concern for users. The integration of AI also raises an issue regarding authorship and meets challenges in formulating a complete legal framework for this technology. The terms “plagiarism” and “copyrights” in the age of AI-generated content may require redefinition or further clarification.

Simplified, plagiarism means using another person’s ideas without giving attribution [59]. Content generated by the AI may contain misappropriated information, which may be perceived as a form of plagiarism. Studies revealed that language models like GPT-2 tend to plagiarize from their training data, often memorizing and implementing parts of the information in the final product [60]. The types of plagiarism vary among language models; however, the findings indicate that paraphrasing, idea plagiarism, and verbatim were prevalent in the tested models. Paraphrasing entails using someone else’s idea in one’s words; idea plagiarism involves stealing statements and solutions [60], whereas verbatim plagiarism involves word-for-word use without proper attribution. Despite great effort, different forms of plagiarism are still present in medical literature across different disciplines [61–63]. At the same time, plagiarism stands for a substantial percentage of reasons for biomedical article retraction [64–66]. Proper citing is crucial, signaling readers to seek more information, lending credibility to the text, and indicating thorough research [67]. ChatGPT, upon request, provides references for most answers. The information provided appears authentic through the citation of reputable journals, well-known authors, and the provision of seemingly genuine Wikipedia URLs. However, when checked thoroughly, it becomes clear that the provided information is inaccurate; it refers to well-known magazines, but the titles cited by AI do not align with journal entries. While specific authors were authentic, they did not contribute to the mentioned research. Provided URLs and DOIs were mostly fabricated. Furthermore, even when the source was genuine, it frequently failed to support the claims made by the AI [68]. Surprisingly, new versions of the software trained on larger datasets performed less accurately in these regards compared to their predecessors [60].

Another concern that typically follows an issue of plagiarism regarding AI-generated content is whether copyrights protect that output. The analysis of the Court of Justice of the European Union’s (CJEU) approach to this matter suggests that EU law sufficiently addresses this issue [69]. For the content to be protected by the copyright, it must fulfill some requirements. Essential are human creative choices that must be present in the final output. Based on the CJEU’s law, the outcome cannot be regarded as a protected work if there is a lack of human engagement in design, specifications, and revision [69]. As long as humans play a crucial role at both the conception and redaction stages, obtaining copyrights should pose no difficulty. This topic seems to be also acknowledged in the United States, and the matter of granting copyrights for AI-generated content was addressed by Congress in 2023. The report emphasizes that the U.S. Copyright Office recognizes copyrighted work only as created by a human being [69]. However, the office acknowledges specific work with copyright if substantial human arrangements and modifications have been applied to AI-generated material. The U.S. Copyright Office also emphasizes that AI contributions should always be identified, and the authors should disclaim such parts [70]. Previously mentioned journals’ policies follow this conviction and almost undisputedly forbid acknowledging AI as coauthors of the manuscripts [70]; nevertheless, some exemptions still appear. In 2023, the Canadian Urological Association published an editorial about using AI in medical publishing, written entirely by ChatGPT [71]. An editor could receive a comprehensive article on a chosen topic with a proper prompt described in detail. Surprisingly, the editorial was very self-aware of AI and highlighted its limitations and the value of human input. Because the editorial included solely the AI-generated response that remained unchanged by the user, the first author position remained vacant and no “human author” was listed. One would argue whether this was a proper solution; nevertheless, authors might protect themselves from excessive criticism and skepticism by being transparent and appropriate reporting.

The presence of AI in biomedical publishing and ineffective detection tools

The obligation of reporting the use of AI in the writing process is crucial to maintaining publishing quality, reliability, and integrity. Multiple journals implemented this rule, however, several cases of abuse were still reported. Namely, repeated use of AI was not indicated in papers, giving the author all the credit [72]. What is even more disturbing is that in the submitted manuscripts, authors occasionally forgot to remove prompt phrases that indicated the use of AI, such as “Regenerate response” or “As an AI language model, I …”. This represented not only the unethical usage of AI by authors but also insufficient attention of the reviewer or editor. The fact that even peer-reviewed articles sometimes contain unevaluated text calls for urgent action [72]. Fortunately, as it turns out, the cause of this problem may also be its solution. Machine learning algorithms can be applied to spot AI-generated writings [73]. The effectiveness of this method was variable and depended on the ChatGPT version used to create the analyzed text. The overall efficiency reached 99%, indicating that it may be an effective solution to detect and report the usage of AI in scientific writing. Other available detection tools are insufficient, with 35–65% and 10–55% effectiveness for ZeroGPT and a tool by OpenAI, respectively [73]. Before the development of such algorithms, a careful reviewer could have suspected the usage of AI by noticing so-called “tortured phrases”. These are uncommon descriptions or paraphrases of well-established concepts [74]. For instance, using “colossal information” instead of “big data”. These strange-looking phrases are most likely from reverse-translation software employed to avoid plagiarism or the detection of the use of AI. This phenomenon was reported in hundreds of articles [75]. Taking notice of those “tortured phrases” might be one of the approaches that may lead the article’s recipient to raise suspicion towards the provided information. With proper guidance, AI can generate convincing full medical articles, which might be hard for an unskilled reader to distinguish from real ones [76]. Lack of attentiveness and criticality to the information encountered may result in spreading misinformation which can be destructive in the medical field. In the era of AI, it is crucial to maintain alertness toward the possibility of fraud performed with AI, as the novel technology creates a new area of abuse.

The research misconduct and unethical practices caused by AI

With the increasing popularity of generative AI, new forms of research misconduct emerge. Unethical organizations or individuals began to exploit loopholes in the law and guidelines for authors and see this as a profit opportunity. They offer to write articles for a specific price, often relatively significant [77]. These companies are called “papermill farms” and it’s becoming clear that they usually apply AI on a large scale. Unlike ghostwriters—who can generate accurate request information, papermills tend to release falsified, biased articles. These papers are rich in the previously mentioned “tortured phrases”. A “papermill alarm” tool was created to track and report these suspicious articles. This solution also implements machine learning to search, detect, and report AI input in the article [78]. Unfortunately, even with such tools, the growing number of papermills farms has become a considerable problem in scientific publishing [77]. Falsified articles pose the most significant threat in medicine and biology [79], which can lead to incorrect treatment of patients or confusion in the field. Until recently, AI was mainly capable of generating text. Now, this is no longer a limitation. There are reports of generated microscope images [77], Western blot results [80], or even transforming a routine imaging test result into pathology [81]. To better visualize this, we used a tool called Midjourney [82, 83] to falsify some results (Figure 2). While with more complex tasks, like RNA-seq, AI might still struggle (Figure 2A and B), more straightforward images represented with Giemsa or immunofluorescence staining show many similarities to the actual results (Figure 2C and D). This is particularly worrying because just as there are regulations against AI utilization in generating text, such rules often do not exist regarding pictures or photos. This is true even among the most essential publishers, as only 5.9% prohibit implementing AI-generated images. Across the top 100 journals’ policies, only 5.7% forbid this practice [27]. Falsified Western blots, immunochemistry images, or any other visualization of results might provide misleading preclinical or clinical evidence that can set the direction of research to the wrong tracts. The lack of clear guidelines regarding the implementation of AI-generated images in biomedical articles might allow misconducted articles to undergo a reviewing process, which does not always detect such procedural improprieties. In 2023, more than 10,000 articles were retracted due to possible manipulation and compromised peer review process [84]. The lack of guidelines on this issue opens a door for fraud and research misconduct and might exacerbate those numbers even more. The risks posed by using AI should be given special attention, as up to 90% of the Internet is expected to be synthetic within a few years [85].

Display full size

Figure 2. 

AI-generated pictures using MidJourney [83] with our slight editing. (A) Fake single-cell RNA sequencing UMAP plot. Prompt: single-cell RNA sequencing data presented as a UMAP plot (with x and y axis) showing four distinct cell populations represented as four distinct groups of clustered dots (all dots are in the same size) in 4 different colors (in the same shade). Axes added post-generation; (B) fake RNA-bulk sequencing heatmap. Prompt: the heatmap of the RNA sequencing data in shades of blue and pink. Legend and gene names added by authors; (C) AI-generated Giemsa staining of white blood cells. Prompt: Giemsa staining of leukocytes in microscopy images Unedited; (D) artificially generated immunofluorescence staining of GFP positive cells. Prompt: immunofluorescence microscopy images of the hypothalamus with eGFP signal. GFP: green fluorescent protein; eGFP: enhanced green fluorescent protein