Introduction

Generative artificial intelligence (AI), such as ChatGPT (OpenAI), has revolutionized all sectors, including academia, with the majority of ChatGPT-related articles published in the medical field since its emergence in late 2022 [1]. The estimated use of articles by large language models (LLMs) has been increasing since ChatGPT’s release [2]. While LLMs possess the capability to generate human-like content across multiple modalities, including text, images, audio, and video, their dependence on pretrained knowledge can lead to inaccuracies or hallucinations [2]. Moreover, AI is having such a significant impact on academia that LLMs are now even being used to assist in systematic literature reviews (SLRs) to minimize human intervention.

With the rising adoption of AI platforms for SLRs (SLR-AIs) that support SLR functionalities in academic writing, researchers face the challenge of selecting the most appropriate SLR-AI for high-quality research. As various SLR-AIs continue to appear, each offering distinct features, choosing and using the right SLR-AI has become a key determinant of research productivity. Researchers now face a new dilemma: should they opt for an SLR-AI trained on bibliographic databases, or should they utilize the SLR functionalities of general-purpose AIs? This is an increasingly relevant concern in the research community.

This study aims to provide a concise introduction to SLR-AIs that support SLR processes, to compare their features and performance, and to present this information to researchers as guidance for their selection and use of SLR-AIs. I intend to focus exclusively on major SLR-AIs, with no conflicts of interest regarding any products, and to share personal experiences as objectively and fairly as possible. This analysis draws on my experience as a practicing researcher and librarian, in addition to a review of relevant literature, in order to present a balanced overview of both theoretical and practical aspects of SLR-AIs.

For this essay, I initially gathered foundational materials by searching Google and bibliographic databases. Subsequently, using major SLR-AIs that were freely or affordably accessible, I collected additional data for comparison by reviewing responses to targeted queries. By comparing the theoretical orientation of SLR-AI as outlined in recent research with the real-world characteristics of available SLR-AIs, this essay is written from the perspective of practical utilization rather than technical development.

Review of key research on SLR-AIs since the release of ChatGPT

Usefulness and limitations of SLR-AIs

ChatGPT-like AIs can serve as effective SLR-AIs to assist novice researchers due to their breadth and accuracy of knowledge. ChatGPT demonstrates potential in automating article screening during SLRs, achieving high sensitivity and substantial reductions in workload, although it exhibits lower specificity compared to human reviewers [3]. GPT-4, in particular, has proven to be an effective tool for enhancing the efficiency of article screening, demonstrating “human-like” and “almost perfect” performance comparable to human experts [4]. Additionally, LLMs contribute to SLR processes by selecting relevant studies and extracting key information [5]. In the SLR workflow, which encompasses literature search, review, data extraction, and synthesis, a retrieval-augmented generation (RAG)-based LLM, which combines real-time information retrieval with generative capabilities, can markedly improve accuracy, relevance, and contextual understanding [6]. Although LLMs have not yet fully supplanted human experts in abstract screening, ChatGPT 4.0 has shown considerable promise for improving SLR with a good balance of sensitivity, specificity, and high overall accuracy [7]. Therefore, LLM-based SLR-AIs are increasingly used to enhance research discovery, visualization, and efficient summarization. A comparative assessment of academic AIs is needed to support the implementation of robust AI-driven scholarly information services, facilitate reliable data-driven research, and improve academic writing [8]. Thus, more detailed quality comparisons of LLMs for SLR should employ qualitative metrics such as accuracy, response time, consistency, depth of knowledge, contextual understanding, and transparency [9]. While LLMs for SLR can reduce the time required and increase the value of analytical results, human oversight remains necessary to ensure the accuracy and reliability of findings [10].

Present and future of SLR-AIs

General generative AI models, such as ChatGPT, originally trained on broad, nonacademic data from the internet, are now offering SLR functionalities due to rapid technological advancements. In response, bibliographic database providers have released SLR-AIs trained on the abstracts within their databases and are offering these as additional subscription services to libraries. Consequently, recent research on SLR-AIs has primarily focused on the screening and extraction phases of article review [11]. According to an analysis of articles retrieved from PubMed, ChatGPT and other GPT-based LLMs represent the most promising architectures for SLR-AIs, despite their relatively low recall, and are expected to transform SLR methodologies [12]. However, SLR-AIs often generate divergent queries even when provided with the same prompt, resulting in suboptimal recall. Therefore, caution is advised when employing these tools for rapid systematic reviews. Continued technological development of SLR-AIs is necessary to enhance the precision of article retrieval, enable systematic integration of retrieved data, and support learning from a wide array of specialized and emerging information sources to generate value-added insights—such as identifying research gaps. Producing reliable and actionable SLR-AI results remains a significant technical and ethical challenge, necessitating increased accuracy for sophisticated analytics. In domains where accuracy and reliability are paramount, such as medicine, there is also a pressing need to implement the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines when utilizing SLR-AIs [13].

Comparison of the characteristics of major SLR-AIs since the release of ChatGPT

Table 1 summarizes the characteristics, advantages, and disadvantages of major SLR-AIs that I have used or that have been described in the literature; only the key highlights are presented here. DeepSeek R1 (DeepSeek) achieves performance comparable to OpenAI’s o1 on reasoning tasks, but with transparent, explainable reasoning, and structured problem solving with break down complex problems. I had no direct experience with it, because Korea restricts the use of DeepSeek R1.

Bibliographic database–based SLR-AIs

Scopus AI

Scopus AI (Elsevier) was developed using abstracts from articles published after 2003 in the Scopus database. The aim of Scopus AI is to help avoid hallucinated references and ensure there are no “fake” citations [14]. Scopus AI formulates queries and performs vector searches to construct a summary of the query, subsequently generating answers by checking cited references for validation and transparency. Scopus AI may be particularly useful for the initial stages of SLRs or for fields with less stringent methodological requirements, as it offers a level of quality and reliability grounded in peer-reviewed articles that surpasses that of general LLM-generated answers. Now available in a mobile version, Scopus AI may also impact competing AIs by introducing new approaches to SLR [15].

Research Assistant

The Research Assistant (RA; Literature Review 2.0, Clarivate) has been offered as an add-on to the Web of Science (WoS) subscription libraries since 2025. Clarivate developed RA for the WoS Core Collection, which has been compiling a bibliographic database of articles since before Scopus. Clarivate trained the RA model on WoS article abstracts to ensure high quality and reliability of the learning sources. However, unlike Scopus AI, RA displays, “I’ll create a search strategy with keywords and synonyms covering these aspects,” during the question-and-answer process. It works by generating answers using GPT technology through an expanded search based on existing keywords. While this approach is similar to Scopus AI, it differs in that RA first automates the selection of search keywords and then uses an LLM to generate answer texts for the most relevant studies in the search results, with results restricted to WoS subscription years.

Dimensions Research GPT

Dimensions (Digital Science) manages research data that includes more than 70% of publications with full-text indexing, and has built one of the most comprehensive collections of grants, publications, patents, clinical trials, and datasets [14]. Since 2000, the Dimensions database has grown more rapidly than WoS and includes many smaller publishers and open access articles [16]. As a result, Dimensions Research GPT has access to a broader range of references, enabling it to generate summaries, insights, and citations based on scientific evidence, including datasets, and thus offer new perspectives.

Elicit Reports

Elicit (Ought Inc) relies on article abstracts from bibliographic database [17]. Elicit Reports (Elicit Systematic Review) feature generates answers based on highly-cited articles sourced from Semantic Scholar (Ai2), PubMed, arXiv, JAMA, and other platforms. Elicit Reports uses semantic search, focusing on search intent rather than simple keywords, and presents results as summaries of eight relevant articles selected from up to 500 retrieved results.

SciSpace Deep Review

SciSpace (Business Integra and d3i) is promoted as an all-in-one AI platform for students and researchers, offering functionalities such as PDF chatting and AI writing support. It allows users to create custom columns, with both free standard and premium Deep Review options, and produces reports based on 50 highly relevant papers. While SciSpace Deep Review features a well-organized table of contents compared to other SLR-AIs, its answers tend to be overly long and detailed, and their accuracy is sometimes uncertain.

Generative AI–based SLR-AIs

Deep Research

OpenAI, Perplexity, and Google have all released SLR-AIs called “Deep Research,” intensifying technological competition and, at times, causing confusion for users, even among LLM-oriented SLR-AIs. These tools summarize findings using logical reasoning and provide citations from both scholarly sources and publicly available web information to support up-to-date SLR. “Deep Research” tools reorganize, expand, modify, and adjust search terms in response to user prompts, but their repeatability, reproducibility, and result consistency for identical prompts tend to be weak. Nonetheless, the use of OpenAI’s Deep Research is strongly recommended [18]. Perplexity and Google’s “Deep Research” were omitted from the comparison because we did not find any significant differences with OpenAI’s “Deep Research.”

Felo Agent

Felo Agent (Felo Inc) is not technically an SLR-AI, as it is not built on a bibliographic database, but is included here for convenience. It searches across numerous academic publications, incorporating models such as DeepSeek R1, o4-mini (OpenAI), GPT-4o (OpenAI), and Claude 4.0 Sonnet (Anthropic), enabling users to easily select and utilize responses from multiple AIs with a single prompt. When Felo Agent provides integrated responses, its answer organization and PPT generation system are logical and well-structured, offering an agent-like SLR-AI experience. In my experience, Felo Agent was stronger at non–peer-reviewed, recent academic evidences than other SLR-AIs, but also more likely to present fictitious references.

Implications derived from the reviewed articles and major SLR-AIs in use

Understand the accuracy and reliability of the SLR-AI training sources

With the emergence of bibliographic database–based AI, SLR-AIs now have the capability to generate the introduction section of a research article without human intervention. However, even the most accurate and reliable LLMs should function as helpful adjuncts to the SLR process, rather than as leading contributors [2]. Although bibliographic database–based SLR-AIs are highly accurate and reliable in terms of their learning sources, questions remain regarding whether the AI technologies used to generate answers are sufficiently robust. For SLR-AIs trained on internet-sourced information, there are concerns that the accuracy and reliability of the results cannot be guaranteed. When evaluating the quality of SLR-AIs, it is essential to consider not only the reliability of their training sources but also the frequent technical updates and rapid evolution of these systems, which can impact the consistency of their learning sources [19].

Be aware of possible plagiarism, copyright law, and research ethics violations

Although bibliographic database–based SLR-AIs claim to have been trained on the abstracts of all the articles within their databases, it remains unclear whether all abstracts have actually been used, including those with copyright restrictions [14]. This issue is even more pronounced for SLR-AIs from Clarivate and Digital Science, which are not themselves publishers and therefore face greater limitations in accessing copyrighted abstracts, compared to Elsevier’s Scopus AI [14]. Generative AI technologies used for manuscript writing, data analysis, peer review, and editorial activities can introduce a range of ethical concerns for authors, reviewers, and editors alike [17,20]. To avoid these issues, outputs generated by SLR-AIs should be thoroughly rewritten in human-understandable language and properly incorporated into manuscripts. Overreliance on AIs has already resulted in article retractions due to research ethics violations [21]. Therefore, it is crucial to disclose the use of SLR-AIs, such as by providing a dedicated methods section to increase transparency [17].

Weigh the usefulness of SLR-AI’s capabilities against potential hallucinations

The risk of hallucinations by SLR-AIs increases when they are tasked with new topics or fields that lack sufficient article coverage, leading to optimism for academic AI models trained on full-text and web-based information rather than just abstracts. Because of the inherent limitations of SLR-AIs, human intervention remains indispensable. SLR-AIs should ideally allow users to upload their own PDFs and interact directly with the results. To promote widespread adoption, it is necessary to thoroughly evaluate both the usefulness and potential hallucinations of SLR-AIs, as well as to support question-and-answer interactions in languages beyond English.

Keep an eye on new technologies such as RAG or AI agent

Frequent updates to incorporate new features can result in rapid obsolescence of SLR-AI models, imposing a continuous learning burden on users. After an initial SLR-AI run, it may become necessary to re-prompt for supplementary research as time advances. This does not mean relying on a single SLR-AI; rather, users should understand the strengths and weaknesses of various SLR-AIs and select the most appropriate tool according to the topic and research context. By leveraging advanced technologies such as RAG or AI agent systems, SLR-AI can become more accurate and widely used, ultimately enhancing research productivity [6]. For deep research tasks, agent-based AI (autonomous problem solving without user intervention) can even outperform RAG and prompt engineering-based approaches, so further development of agent-based SLR-AIs is anticipated. With new releases like OpenAI’s ChatGPT agent, which boasts deep research capabilities similar to Felo Agent, competition among SLR-AIs is intensifying.

Conclusion

A variety of SLR-AIs have emerged since ChatGPT, accelerating the processes of literature search and summarization. Bibliographic database–based SLR-AIs, in particular, are effective as starting points at the outset of research. They improve researchers’ understanding of new concepts, enable the collection of reliable articles, and generate summaries by automating keyword selection. However, they may produce incomplete answers in emerging fields. Nonbibliographic database–based SLR-AIs, on the other hand, often overlook key references and important findings. More SLR-AIs offer natural language summaries, visualized results, and suggested follow-up questions. The ongoing technological race in the SLR-AI landscape, marked by advances in LLM reasoning, agent technology, and RAG, remains fierce. Current SLR-AI technology cannot fully replace researchers but rather serves to augment human expertise. Human researchers provide critical analysis, uniquely identify trends, and offer perspectives on future research directions that go beyond mere summarization. Exclusive reliance on SLR-AIs does not foster the growth necessary to become an independent researcher. Based on my experience, it remains difficult to precisely determine the reliability of SLR-AIs. The future of SLR-AI should be grounded in trustworthy and efficient content, aided by responsible AI practices. This study suggests that careful selection of SLR-AIs can optimize the efficiency of literature reviews. Solutions such as Felo Agent, in conjunction with adherence to the four implications discussed above, represent a prudent choice. However, excessive reliance on SLR-AIs that draw from public sources poses immediate research ethics risks, making final, comprehensive human review essential. As more articles are published utilizing SLR-AI, increasing effort will be required to verify their authenticity, which risks becoming a societal burden.

Notes

Conflict of interest

No potential conflict of interest relevant to this article was reported.

Funding

The author received no financial support for this article.

Acknowledgments

Artificial intelligence was used for testing products, but not for writing articles.

Data availability

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

Supplementary materials

References

• 1. Kim SJ. Trends in research on ChatGPT and adoption-related issues discussed in articles: a narrative review. Sci Ed 2024;11:3-11.https://doi.org/10.6087/kcse.321. Article
• 2. Tang BL. Will widespread use of artificial intelligence tools in manuscript writing mark the end of human scholarship as we know it? Sci Ed 2025;12:231-3.https://doi.org/10.6087/kcse.366. Article
• 3. Issaiy M, Ghanaati H, Kolahi S, et al. Methodological insights into ChatGPT’s screening performance in systematic reviews. BMC Med Res Methodol 2024;24:78. https://doi.org/10.1186/s12874-024-02203-8. ArticlePubMedPMC
• 4. Khraisha Q, Put S, Kappenberg J, Warraitch A, Hadfield K. Can large language models replace humans in systematic reviews? Evaluating GPT-4’s efficacy in screening and extracting data from peer-reviewed and grey literature in multiple languages. Res Synth Methods 2024;15:616-26.https://doi.org/10.1002/jrsm.1715. ArticlePubMed
• 5. Lieberum JL, Toews M, Metzendorf MI, et al. Large language models for conducting systematic reviews: on the rise, but not yet ready for use: a scoping review. J Clin Epidemiol 2025;181:111746. https://doi.org/10.1016/j.jclinepi.2025.111746. ArticlePubMed
• 6. Han B, Susnjak T, Mathrani A. Automating systematic literature reviews with retrieval-augmented generation: a comprehensive overview. Appl Sci 2024;14:9103. https://doi.org/10.3390/app14199103. Article
• 7. Li M, Sun J, Tan X. Evaluating the effectiveness of large language models in abstract screening: a comparative analysis. Syst Rev 2024;13:219. https://doi.org/10.1186/s13643-024-02609-x. ArticlePubMedPMC
• 8. Park J. A study on the current status analysis and service strategies of LLM-based conversational AI academic information services. J Kor Soc Inf Manag 2025;42:29-57.https://doi.org/10.3743/KOSIM.2025.42.1.029. Article
• 9. Mostafapour M, Fortier JH, Pacheco K, Murray H, Garber G. Evaluating literature reviews conducted by humans versus ChatGPT: comparative study. JMIR AI 2024;3:e56537. https://doi.org/10.2196/56537. ArticlePubMedPMC
• 10. Li Y, Datta S, Rastegar-Mojarad M, et al. Enhancing systematic literature reviews with generative artificial intelligence: development, applications, and performance evaluation. J Am Med Inform Assoc 2025;32:616-25.https://doi.org/10.1093/jamia/ocaf030. ArticlePubMedPMC
• 11. Bolaños F, Salatino A, Osborne F, Motta E. Artificial intelligence for literature reviews: opportunities and challenges. Artif Intell Rev 2024;57:259. https://doi.org/10.1007/s10462-024-10902-3. Article
• 12. Scherbakov D, Hubig N, Jansari V, Bakumenko A, Lenert LA. The emergence of large language models as tools in literature reviews: a large language model-assisted systematic review. J Am Med Inform Assoc 2025;32:1071-86.https://doi.org/10.1093/jamia/ocaf063. ArticlePubMedPMC
• 13. Bernard N, Sagawa Y, Bier N, et al. Using artificial intelligence for systematic review: the example of elicit. BMC Med Res Methodol 2025;25:75. https://doi.org/10.1186/s12874-025-02528-y. ArticlePubMedPMC
• 14. De Leon J, De Leon-Martinez S, Artés-Rodríguez A, Baca-García E, De Las Cuevas C. Reflections on the potential and risks of AI for scientific article writing after the AI endorsement by some scientific publishers: focusing on Scopus AI. Actas Esp Psiquiatr 2025;53:433-42.https://doi.org/10.62641/aep.v53i2.1849. ArticlePubMedPMC
• 15. Nogueira LA. Chatbots and scholarly databases: impressions from trying out Scopus AI. Inf Technol Libr 2025;44:1-8.https://doi.org/10.5860/ital.v44i1.17116. Article
• 16. van Bellen S, Alperin JP, Larivière V. Scholarly publishing’s hidden diversity: how exclusive databases sustain the oligopoly of academic publishers. PLoS One 2025;20:e0327015. https://doi.org/10.1371/journal.pone.0327015. ArticlePubMedPMC
• 17. Kim SJ. Research ethics and issues regarding the use of ChatGPT-like artificial intelligence platforms by authors and reviewers: a narrative review. Sci Ed 2024;11:96-106.https://doi.org/10.6087/kcse.343. Article
• 18. Haman M, Školník M. Fake no more: the redemption of ChatGPT in literature reviews. Account Res; 2025 Feb. 15. [Epub]. https://doi.org/10.1080/08989621.2025.2465619. ArticlePubMed
• 19. Daraio C, Di Leo S, Ferazzoli F. Pitfalls, benefits, and comparative analysis of artificial intelligence ChatBots in the systematic review process. Int Trans Oper Res; 2025 May. 25. [Epub]. https://doi.org/10.1111/itor.70054. Article
• 20. Yun CH. Research and publication ethics with generative artificial intelligence-assisted tools. Sci Ed 2025;12:1-3.https://doi.org/10.6087/kcse.362. Article
• 21. Travis K. Springer Nature retracted 2,923 papers last year [Internet]. Retraction Watch; 2025 [cited 2025 Jun 12]. Available from: https://retractionwatch.com/2025/02/17/springer-nature-journal-retractions-2024.

Figure & Data

References

Citations

Citations to this article as recorded by