Reimagining biomedical science workflows in the age of large language models

Roberts, Kaleigh F.; Koutarapu, Srinivas; Melendez, Justin; Landsness, Eric C.; Reese, Justin T.; Elbert, Donald L.; Bateman, Randall J.

doi:10.1038/s44400-026-00126-3

Cited by 0Artificial Intelligence in Healthcare and EducationArtificial Intelligence ApplicationsScientific Computing and Data Management

Reimagining biomedical science workflows in the age of large language models

Open access

Roberts, Kaleigh F., Koutarapu, Srinivas, Melendez, Justin, Landsness, Eric C., Reese, Justin T., Elbert, Donald L., Bateman, Randall J.·2026·npj Dementia

DOI 10.1038/s44400-026-00126-3

Open this paper in CoChat — free account Open PDF

Abstract

Large language models (LLMs) are generative artificial intelligence (AI) models that are rapidly reshaping the practice of biomedical science. Their ability to synthesize literature, generate analytical code, and interface with multimodal data offers a new framework for accelerating discovery. Yet their integration into scientific workflows remains irregular, and the field lacks clear guidance for reliable and productive use. We review emerging evidence on researcher adoption, highlight common failure modes such as so-called hallucinations (i.e. confabulations) and overgeneralization, and provide practical recommendations for domain-informed use of LLMs in basic biomedical research. We structure this review around four domains in which LLMs increasingly augment biomedical science: administrative tasks, literature search and synthesis, data analysis, and scientific writing. For each domain we provide practical guidance, illustrative use cases, and examples of free or low-cost tools that researchers can readily adopt. Finally, we discuss the organizational and cultural changes for biomedical science to leverage LLMs responsibly, including transparent reporting, human-in-the-loop validation, and alignment with scientific rigor and reproducibility standards. Together, these recommendations provide a path for integrating LLMs into biomedical research in ways that enhance, rather than replace, human expertise and accelerate the path from biological insights to beneficial human impact.

AI snapshot

Written from the paper’s full text.

TL;DR

This narrative review argues that large language models can meaningfully accelerate biomedical research across administrative work, literature search and synthesis, data analysis, and scientific writing, but only when paired with human verification and domain expertise. It offers practical guidance, tool comparisons, and organizational recommendations for responsible adoption.

Key findings

Researcher interest in AI is high but adoption is uneven: in the Wiley ExplanAItions studies, more than 60% of researchers selected lack of training or clear guidelines as a major barrier to adoption.
Writing attitudes split sharply by task: about 10% of researchers believe AI use is never appropriate for editing, while 35% believe it is never appropriate for writing a first draft; roughly 28% report having used AI to edit a manuscript and only 8% to write a first draft.
Population-level estimates of LLM use in writing: ~10% of bioRxiv abstract sentences carried LLM-associated stylistic signals by September 2024, and 13.5% of all 2024 PubMed abstracts were estimated to be processed with LLMs; authors incorporating LLMs published more preprints per month than nonadopters.
Recurring failure modes are fabricated citations (hallucinations/confabulations), overgeneralization, statistically inappropriate test selection, hallucinated library functions, and drift toward overclaiming or generic phrasing; arXiv imposes a one-year ban on authors who submit hallucinated citations.

Methods

Narrative review reflecting user experience rather than a systematic review; the authors state they aimed to cover the landscape broadly without claiming completeness and to remain LLM-agnostic by drawing examples from a range of tools.
Organized around four domains — administrative tasks, literature search and synthesis, data analysis, and scientific writing — with "pearls and pitfalls" summaries, illustrative use cases, and tables of free or low-cost tools (literature tools current as of May 20, 2026).
No datasets were generated or analyzed; the paper synthesizes prior surveys and studies (e.g., Nature surveys, arXiv/bioRxiv/PubMed linguistic analyses) and cites emerging work on AI coscientists.

Limitations

Stated: the literature on AI for biomedical research is thin and mostly limited to preprints because tools evolve faster than peer review and developers have little incentive to publish; the article reflects user experience rather than systematic review and does not claim completeness.
Stated: tool coverage is uneven across niche subfields, map and recommendation quality depend heavily on seed-paper choice, and citation-validation efforts remain limited regarding citation quality and balanced coverage.
Evident from the text: Tables 4 (scientific writing prompts) and 5 (AI coscientist landscape) appear without content in the provided text, so those promised examples and tool listings cannot be evaluated; the review also reports no primary data of its own, and several cited adoption figures come from preprints or non-peer-reviewed surveys.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s44400-026-00126-3

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

36 of 79 cited sources checked · 10 supported · 1 flagged · accuracy not yet measured

Premisecitation
Interest in using AI to support research is extremely high.
“First, interest in using AI to support research is extremely high (Fig. 1).”
Quote in the paperNot a substantial claim3 of 9 sources checked
- cites [4] Van Noorden, R. & Perkel, J. M. AI and science: what 1,600 researchers think. Nature 621, 672–675 (2023). source
- cites [5] Ng, J. Y. et al. Attitudes and perceptions of medical researchers towards the use of artificial intelligence chatbots in the scientific process: an international cross-sectional survey. Lancet Digit. Health 7, e94–e102 (2025). source
  The Abstract (p. 2) reports that 69.7% of surveyed medical researchers wanted to learn more about and receive training on AI chatbot use in research. The Experience with AI Chatbots section (p. 11) reports that 58.1% were likely or very likely to use chatbots in future research. These findings show substantial interest, but interest in training is not the same as interest in using AI, and the paper does not establish that interest in use is ‘extremely high.’ (LLM, conf=0.83)
  Experimental — accuracy not yet measured
- cites [6] Ng, J. Y. et al. Attitudes and perceptions of biomedical journal editors in chief towards the use of artificial intelligence chatbots in the scholarly publishing process: a cross-sectional survey. Res. Integr. Peer Rev. 10, 19 (2025). source
  The Discussion (page unavailable) reports notable interest in using AI chatbots for research support, and the Results say 77% of surveyed biomedical editors in chief considered them important to future scientific research. Neither passage establishes that interest in using AI to support research is “extremely high” generally; the Discussion also notes concerns about accuracy and transparency. (LLM, conf=0.82)
  Experimental — accuracy not yet measured
- cites [7] Nordling, L. How ChatGPT is transforming the postdoc experience. Nature 622, 655–657 (2023). source
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
- cites [9] Researchers and AI: Survey Findings. https://academic.oup.com/pages/ai-survey-findings.
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [10] Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [11] ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
  Not checked in the free run.
- cites [12] Ruediger, D., McCracken, C. & Skinner, M. Adoption of Generative AI by Academic Biomedical Researchers. http://sr.ithaka.org/?p=321415 (2024) https://doi.org/10.18665/sr.321415. source
  The Introduction describes biomedical researchers as having a “moderate degree of interest” in using generative AI for research. The “Interest in generative AI as a research tool” section reports that about a third are very or extremely interested, rather than interest being extremely high overall. Page numbers are not available in the supplied passages. (LLM, conf=0.86)
  Checked by a reasoning model — may be wrong.
  Experimental — accuracy not yet measured
Premisecitation
AI-assisted data analysis has emerged as one of the most powerful and democratizing applications of LLMs in scientific research.
“AI-assisted data analysis has emerged as one of the most powerful and democratizing applications of LLMs in scientific research 19 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [19] Yin, H. et al. An Evaluation of Large Language Models in Bioinformatics Research. Preprint at https://doi.org/10.48550/arXiv.2402.13714 (2024). source
  The Introduction (page not provided) reports that LLMs can partially solve several bioinformatics tasks and can match competitive baselines on simple tasks, supporting their potential for AI-assisted analysis. It does not establish that data analysis is among the most powerful applications of LLMs across scientific research or that it democratizes research. (LLM, conf=0.85)
  Experimental — accuracy not yet measured
Premisecitation
Modern LLMs substantially lower the barrier to computational analysis by enabling researchers to describe analytical tasks in plain language and receive executable code in return.
“Modern LLMs substantially lower this barrier by enabling researchers to describe analytical tasks in plain language and receive executable code in return 20,21 .”
Quote in the paperSubstantial claim2 of 2 sources checked
- cites [20] Chen, M. et al. Evaluating Large Language Models Trained on Code. Preprint at https://doi.org/10.48550/arXiv.2107.03374 (2021). source
  The Codex and Codex-S Performance section (p. 2) demonstrates that a language model can generate Python functions from natural-language docstrings, with some outputs passing executable unit tests. H.3 (p. 34) suggests code generation could widen access to programming. The paper does not establish that researchers can obtain code for analytical tasks or that the barrier to computational analysis has been substantially lowered. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [21] Tian, H. et al. Is ChatGPT the Ultimate Programming Assistant -- How far is it? Preprint at https://doi.org/10.48550/arXiv.2304.11938 (2023). source
  Section 2.2.1 describes generating code from natural-language problem descriptions, and the Section 4.1 findings report that ChatGPT can generate correct code for common programming problems. The paper studies programming assistance for developers, not researchers’ computational analyses, and does not establish that it substantially lowers the barrier to that work; it also notes weaker performance on new, difficult problems. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- Rests on an assumption: Traditionally, computational analysis has required specialized programming skills, creating a barrier for many experimental scientists.
Premisecitation
New LLM-based tools supporting end-to-end scientific workflows are emerging.
“new LLM-based tools supporting end-to-end scientific workflows are emerging 34 .”
Quote in the paperSubstantial claim1 of 1 sources checked
- cites [34] Roberts, K. F. et al. Towards an AI biomedical scientist: Accelerating discoveries in neurodegenerative disease. J. Prev. Alzheimers Dis. 13, 100398 (2026). source
Premisecitation
13.5% of all 2024 PubMed abstracts were processed with LLMs.
“Another study which estimated LLM usage in scientific writing by quantitating vocabulary shifts suggested that 13.5% of all 2024 PubMed abstracts were processed with LLMs 27 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [27] Kobak, D., González-Márquez, R., Horvát, E. -Á & Lause, J. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Sci. Adv. 11, eadt3813 (2025). source
  The Discussion (Summary) says the excess-vocabulary analysis suggests that at least 13.5% of 2024 PubMed abstracts were processed with LLMs. The citing sentence gives 13.5% without the important lower-bound qualifier, making it sound like a point estimate. (LLM, conf=0.85)
  Experimental — accuracy not yet measured

Premisecitation
Most publishers now allow AI tools for editing, readability, and grammar, often with disclosure.
“Most publishers now allow AI tools for editing, readability, and grammar often with disclosure.”
Quote in the paperNot a substantial claim2 of 3 sources checked
- cites [29] Flanagin, A., Bibbins-Domingo, K., Berkwits, M. & Christiansen, S. L. Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge. JAMA 329, 637 (2023). source
  The paper’s main section says JAMA Network authors should disclose AI assistance with writing or editing, and it describes tools used for grammar and language. It does not establish that most publishers allow these uses; instead, it describes differing journal policies, including restrictions on AI-generated text. It also exempts basic grammar and spelling checks from JAMA’s disclosure requirement. No page number is available in our copy. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [30] Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613, 612–612 (2023). source
  The ground-rules section (p. 1) says Nature and Springer Nature journals require researchers to document LLM use, and that other publishers are likely to adopt a similar stance. It does not establish that most publishers already allow AI tools, or specify policies for editing, readability, and grammar. (LLM, conf=0.83)
  Experimental — accuracy not yet measured
- cites [31] Conroy, G. How ChatGPT and other AI tools could disrupt scientific publishing. Nature 622, 234–236 (2023). source
Premisecitation
Human readers have imperfect discrimination of human vs. AI-generated text.
“Because human readers have imperfect discrimination of human vs. AI-generated text 25 and there is some stigma associated with AI use, algorithmic approaches have become central for estimating LLM usage in scientific writing.”
Quote in the paperSubstantial claim1 of 1 sources checked
- cites [25] Gao, C. A. et al. Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers. Npj Digit. Med. 6, (2023). source
Interpretationcitation
The fabrication of citations by LLMs occurs because the model generates text from patterns rather than retrieving verified sources.
“This phenomenon occurs because the model is generating text from patterns rather than retrieving verified sources 18 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [18] Bhattacharyya, M., Miller, V. M., Bhattacharyya, D. & Miller, L. E. High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content. Cureus 15, e39238 (2023). source
  The Discussion says ChatGPT generates responses from learned statistical patterns and cannot distinguish accurate from false information. But it explicitly says the cause of fabricated or inaccurate references is unclear; it does not establish that failure to retrieve verified sources is the cause. The citing sentence presents that explanation as certain. (LLM, conf=0.87)
  Experimental — accuracy not yet measured
Premisecitation
Authors who incorporate LLMs into their writing publish more preprints per month than nonadopters.
“Adoption also appears to translate into measurable output gains, with authors who incorporate LLMs into their writing publishing more preprints per month than nonadopters 28 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [28] Kusumegi, K. et al. Scientific production in the era of large language models. Science 390, 1240–1243 (2025). source
Premisecitation
A frequently encountered hallucination in AI-powered literature tools involves fabricated citations.
“A frequently encountered hallucination (i.e.confabulation) in this context involves fabricated citations [14][15][16][17][18] .”
- Also stated in: § Literature search and synthesis · plain
  “General purpose LLMs can readily produce citation-like strings that appear plausible and conform to familiar formatting conventions, yet correspond to papers that do not exist.”
- Also stated in: § Pearls Pitfalls · plain
  “General-purpose LLMs fabricate plausible-looking citations; every reference needs verification against an authoritative database.”
Quote in the paperSubstantial claim3 of 3 sources checked
- cites [14] Chelli, M. et al. Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis. J. Med. Internet Res. 26, e53164 (2024). source
- cites [15] Alkaissi, H. & McFarlane, S. I. Artificial Hallucinations in ChatGPT: Implications in Scientific Writing. Cureus 15, e35179 (2023). source
  The Figure 2 caption and the “PMID: PubMed ID” section report that ChatGPT supplied references whose PMID numbers belonged to different papers, supporting the point that AI-generated citations can be false. The paper describes ChatGPT examples, however; it does not establish how frequently this happens across AI-powered literature tools. No page number is available in our copy. (LLM, conf=0.84)
  Experimental — accuracy not yet measured
- cites [16] Aljamaan, F. et al. Reference hallucination score for medical artificial Intelligence Chatbots: development and usability Study. JMIR Med. Inform. 12, (2024). source
Findingnone given
Most of these systems are semi-autonomous.
“These systems combine literature search, data analysis, and writing in one interface, and most are subscription-based, more expensive, and semi-autonomous (Table 5).”
Quote in the paperNot a substantial claim
Findingnone given
These systems combine literature search, data analysis, and writing in one interface.
“These systems combine literature search, data analysis, and writing in one interface, and most are subscription-based, more expensive, and semi-autonomous (Table 5).”
Quote in the paperNot a substantial claim
Interpretationreasoning
Algorithmic approaches have become central for estimating LLM usage in scientific writing.
“Because human readers have imperfect discrimination of human vs. AI-generated text 25 and there is some stigma associated with AI use, algorithmic approaches have become central for estimating LLM usage in scientific writing.”
Quote in the paperSubstantial claim
- Rests on an assumption: There is some stigma associated with AI use.
Premisecitation
Early career researchers and scientists for whom English is not a first language show higher willingness to use AI for writing tasks.
“Early career researchers and scientists for whom English is not a first language show higher willingness to use AI for writing tasks.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Premisecitation
Self-reported AI use in writing is relatively modest.
“Self-reported AI use in writing is relatively modest.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Premisecitation
Many more researchers express openness to trying AI writing tools in the future.
“Many more express openness to trying these tools in the future.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Interpretationreasoning
These patterns suggest a gradual normalization of AI assistance with scientific writing.
“These patterns suggest a gradual normalization of AI assistance with scientific writing.”
Quote in the paperNot a substantial claim
Premisecitation
In bioRxiv abstracts, the estimated fraction of sentences carrying LLM-associated stylistic signals had risen to ~10% by September 2024.
“In bioRxiv abstracts, the estimated fraction had risen to ~10% of sentences by the study end in September 2024.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [26] Liang, W. et al. Quantifying large language model usage in scientific papers. Nat. Hum. Behav. https://doi.org/10.1038/s41562-025-02273-8 (2025). source
Interpretationreasoning
Institutional structures such as training programs, formal guidelines, and field-specific standards lag behind actual practice.
“Interest is high, capabilities are rapidly evolving, and adoption is accelerating, yet institutional structures such as training programs, formal guidelines, and field specific standards lag behind actual practice.”
Quote in the paperSubstantial claim
Premisecitation
These skills equip models with current best practices, proper API usage, and validated workflows for specialized tools such as pyOpenMS, Scanpy, RDKit, and biomedical databases.
“Rather than relying solely on training data, these skills equip models with current best practices, proper API usage, and validated workflows for specialized tools such as metabolomics processing software (pyOpenMS), single-cell RNA-seq analysis (Scanpy), cheminformatics packages (RDKit), and access to biomedical databases (PubMed, UniProt, ChEMBL).”
Quote in the paperSubstantial claim0 of 1 sources checked
- cites [23] K-Dense Inc. Claude Scientific Skills: A Comprehensive Collection of Scientific Tools for Claude AI. (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Interpretationreasoning
LLMs can explain code in plain language.
“Another advantage of LLM-assisted computation is the ability of models to explain code in plain language.”
- Also stated in: § Pearls Pitfalls · plain
  “Models can explain unfamiliar functions and statistics in plain language, supporting trainees and interdisciplinary review.”
Quote in the paperNot a substantial claim
Interpretationreasoning
For researchers, code generation can incorporate established methodologies from their specific subdiscipline, reducing the need to manually specify every analytical parameter or correct for common pitfalls in specialized analyses.
“For researchers, this means that code generation can incorporate established methodologies from their specific subdiscipline, reducing the need to manually specify every analytical parameter or correct for common pitfalls in specialized analyses.”
Quote in the paperSubstantial claim
Interpretationreasoning
The broader scope of these systems makes them harder to control and validate.
“The broader scope makes them harder to control and validate, so we recommend users first build fluency with narrower, task-specific tools.”
Quote in the paperSubstantial claim
Premisecitation
The datapoint for the Nature Van Noorden study excludes the researchers who study AI (~50% of respondents).
“The datapoint for the Nature Van Noorden study 4 excludes the researchers who study AI ( ~50% of respondents)”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [4] Van Noorden, R. & Perkel, J. M. AI and science: what 1,600 researchers think. Nature 621, 672–675 (2023). source

May combine two claims. 25
Classified as probably not a substantial claim. 31
The quote only partly states this claim. 1

May combine two claims.
Premisecitation
Surveys across the scientific community consistently show three dominant themes: high interest, uncertainty about how to begin, and concerns about accuracy, bias, reproducibility, and erosion of critical thinking.
“Surveys across the scientific community consistently show three dominant themes [4][5][6][7][8][9][10][11][12] .”
Quote in the paperNot a substantial claim3 of 9 sources checked
- cites [4] Van Noorden, R. & Perkel, J. M. AI and science: what 1,600 researchers think. Nature 621, 672–675 (2023). source
- cites [5] Ng, J. Y. et al. Attitudes and perceptions of medical researchers towards the use of artificial intelligence chatbots in the scientific process: an international cross-sectional survey. Lancet Digit. Health 7, e94–e102 (2025). source
  The Abstract (p. 2) reports strong interest in learning about AI chatbots, and the Implications (p. 21) describe limited training and concerns about errors and ethical issues. The survey also reports concern about biased outputs (Perceived Benefits and Challenges, p. 14). It does not establish that these are three consistent themes across surveys of the wider scientific community, or clearly identify uncertainty about how to begin, reproducibility, and erosion of critical thinking as dominant concerns. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [6] Ng, J. Y. et al. Attitudes and perceptions of biomedical journal editors in chief towards the use of artificial intelligence chatbots in the scholarly publishing process: a cross-sectional survey. Res. Integr. Peer Rev. 10, 19 (2025). source
  The Discussion (no page number in our copy) reports editors’ interest in AI chatbots, limited training, and concerns about bias and reliability; the Introduction also reports researchers’ concerns about accuracy and reproducibility. But this survey is limited to biomedical journal editors in chief, and our copy does not establish three consistent themes across the scientific community or identify erosion of critical thinking as one of them. (LLM, conf=0.84)
  Experimental — accuracy not yet measured
- cites [7] Nordling, L. How ChatGPT is transforming the postdoc experience. Nature 622, 655–657 (2023). source
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
- cites [9] Researchers and AI: Survey Findings. https://academic.oup.com/pages/ai-survey-findings.
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [10] Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [11] ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
  Not checked in the free run.
- cites [12] Ruediger, D., McCracken, C. & Skinner, M. Adoption of Generative AI by Academic Biomedical Researchers. http://sr.ithaka.org/?p=321415 (2024) https://doi.org/10.18665/sr.321415. source
  The Introduction and Discussion (pages not provided) report interest in generative AI, uncertainty about best practices, and concerns about accuracy; Social and research ethics also discusses bias. This paper does not establish that surveys consistently find these themes across the scientific community, and the cited passages do not identify erosion of critical thinking or reproducibility as survey themes. (LLM, conf=0.84)
  Experimental — accuracy not yet measured
Probably not a substantial claim: may not pass S2 (checkable).
Premisecitation
A structured workflow can help researchers integrate AI into scientific writing while maintaining rigor.
“A structured workflow can help researchers integrate AI into scientific writing while maintaining rigor 32 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [32] Lin, Z. Techniques for supercharging academic writing with generative AI. Nat. Biomed. Eng. 9, 426–431 (2025). source
Probably not a substantial claim: may not pass S2 (checkable).
Premisecitation
The danger is not that trainees use these tools, but that they use them in place of the foundational practice that develops critical evaluation.
“The danger is not that trainees use these tools, but that they use them in place of the foundational practice that develops critical evaluation 36 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [36] Parker, J. L. & Becker, K. P. Defining and assessing AI literacy for researchers across the research lifecycle. Front. Educ. 11, (2026). source
May combine two claims.
Premisecitation
Nearly all publishers agree that an LLM cannot be a co-author or bear responsibility for errors.
“While journal policies on AI use are evolving, nearly all publishers agree that an LLM cannot be a co-author or bear responsibility for errors 29 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [29] Flanagin, A., Bibbins-Domingo, K., Berkwits, M. & Christiansen, S. L. Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge. JAMA 329, 637 (2023). source
  In the main text (page not provided), the paper says Nature bars AI tools from credited authorship because they cannot take responsibility for the work, and that other journals and organizations are developing similar policies. It does not establish that nearly all publishers agree, so the claim overstates the breadth of that agreement. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
May combine two claims.
Premisecitation
LLMs can produce inaccurate or fabricated outputs (hallucinations/confabulations) and their underlying reasoning is often opaque.
“At the same time, researchers recognized that LLMs can produce inaccurate or fabricated outputs (i.e.hallucinations, which may better be described as confabulations) and that the underlying reasoning behind model responses is often opaque 3 .”
Quote in the paperSubstantial claim1 of 1 sources checked
- cites [3] Huang, L. et al. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Trans. Inf. Syst. 43, 1–55 (2025). source
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation
AI tools cannot be listed as authors, and human authors remain responsible for accuracy.
“AI tools cannot be listed as authors, and human authors remain responsible for accuracy.”
Quote in the paperNot a substantial claim2 of 3 sources checked
- cites [29] Flanagin, A., Bibbins-Domingo, K., Berkwits, M. & Christiansen, S. L. Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge. JAMA 329, 637 (2023). source
- cites [30] Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613, 612–612 (2023). source
- cites [31] Conroy, G. How ChatGPT and other AI tools could disrupt scientific publishing. Nature 622, 234–236 (2023). source
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingcitation
Different surveys report mixed, but generally increasing adoption of AI by the scientific community.
“Different surveys report mixed, but generally increasing adoption of AI by the scientific community.”
Quote in the paperNot a substantial claim3 of 9 sources checked
- cites [4] Van Noorden, R. & Perkel, J. M. AI and science: what 1,600 researchers think. Nature 621, 672–675 (2023). source
- cites [5] Ng, J. Y. et al. Attitudes and perceptions of medical researchers towards the use of artificial intelligence chatbots in the scientific process: an international cross-sectional survey. Lancet Digit. Health 7, e94–e102 (2025). source
  The Experience with AI Chatbots section (p. 11) reports that 52.1% of respondents had used a chatbot in scientific work, while Comparative Literature (p. 20) reports 77.1% AI use in a survey of a different research group. These surveys support variation in reported use, and the Background (p. 4) calls AI an increasingly important research tool, but the paper does not establish a general increase in adoption over time. (LLM, conf=0.84)
  Experimental — accuracy not yet measured
- cites [6] Ng, J. Y. et al. Attitudes and perceptions of biomedical journal editors in chief towards the use of artificial intelligence chatbots in the scholarly publishing process: a cross-sectional survey. Res. Integr. Peer Rev. 10, 19 (2025). source
  The Introduction (page not provided) says AI chatbots are increasingly used in academic contexts, and the Discussion (page not provided) reports mixed views among biomedical journal editors. But this cross-sectional survey found limited use in editors’ work; it does not show that different surveys establish a generally increasing adoption trend across the scientific community. (LLM, conf=0.82)
  Experimental — accuracy not yet measured
- cites [7] Nordling, L. How ChatGPT is transforming the postdoc experience. Nature 622, 655–657 (2023). source
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
- cites [9] Researchers and AI: Survey Findings. https://academic.oup.com/pages/ai-survey-findings.
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [10] Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [11] ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
  Not checked in the free run.
- cites [12] Ruediger, D., McCracken, C. & Skinner, M. Adoption of Generative AI by Academic Biomedical Researchers. http://sr.ithaka.org/?p=321415 (2024) https://doi.org/10.18665/sr.321415. source
  The Introduction and Adoption of generative AI sections describe differing survey findings and mixed levels of use: many biomedical researchers had tried generative AI, but few used it regularly. The Longitudinal data collection section calls this survey a snapshot and says future data are needed to measure adoption over time. The paper therefore does not establish a generally increasing trend across the scientific community, and its findings concern generative AI rather than AI broadly. (LLM, conf=0.84)
  Experimental — accuracy not yet measured
Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.
Premisecitation
The concern is not the use of LLMs per se, but whether they are integrated in ways that reinforce active reasoning and intellectual ownership, or instead encourage delegation of thinking itself.
“The concern, therefore, is not the use of LLMs per se, but whether they are integrated in ways that reinforce active reasoning and intellectual ownership, or instead encourage delegation of thinking itself 38 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [38] Why AI cannot do good science without humans. Nature 653, 650–650 (2026). source
  The editorial (section D, p. 650) says AI scientists should empower human researchers, not replace them. Our copy also describes humans guiding and checking AI systems, but it does not specifically address whether integrating LLMs reinforces active reasoning and intellectual ownership or encourages delegation of thinking. (LLM, conf=0.83)
  Experimental — accuracy not yet measured
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation
Journals and funding agencies have introduced policies to guide responsible use of AI-assisted writing.
“As AI-assisted writing becomes more common, journals and funding agencies have introduced policies to guide responsible use [29][30][31] .”
Quote in the paperNot a substantial claim2 of 3 sources checked
- cites [29] Flanagin, A., Bibbins-Domingo, K., Berkwits, M. & Christiansen, S. L. Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge. JAMA 329, 637 (2023). source
  The main text (page unavailable in our copy) says JAMA Network journals updated their policies for AI-assisted manuscript writing and describes policies at Nature and other journals. It does not identify funding agencies as having introduced such policies, so that part of the claim is unsupported. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [30] Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613, 612–612 (2023). source
  The section headed “No LLM tool will be accepted as a credited author on a research paper” (p. 1) says Nature and other Springer Nature journals added principles on LLM use to their author guidance. It does not say that funding agencies introduced such policies, so that part of the attributed claim is unsupported. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [31] Conroy, G. How ChatGPT and other AI tools could disrupt scientific publishing. Nature 622, 234–236 (2023). source
Probably not a substantial claim: may not pass S2 (checkable).
Premisecitation
Discrepancies between AI-assisted and manual analysis can highlight either shortcomings in the AI-generated approach or opportunities to refine the analysis.
“Discrepancies can highlight either shortcomings in the AI-generated approach or opportunities to refine the analysis 24 .”
Quote in the paperNot a substantial claim1 of 1 sources checked
- cites [24] Buscemi, A. A Comparative Study of Code Generation using ChatGPT 3.5 across 10 Programming Languages. Preprint at https://doi.org/10.48550/arXiv.2308.04477 (2023). source
  The Introduction describes comparing ChatGPT-generated code across programming languages, and the Conclusion identifies limitations in that code. Neither section addresses discrepancies between AI-assisted and manual analysis or says such discrepancies can guide refinement. Searches for the claim’s phrasing and for a comparison with manually written work found no supporting passage; page numbers were not available in our copy. (LLM, conf=0.72)
  Experimental — accuracy not yet measured
May combine two claims.
Premisecitation
Users can describe the structure of their dataset, specify normalization, request plots, or outline statistical comparisons, and the model can generate code in Python, R, or another preferred language.
“Users can describe the structure of their dataset, specify the desired normalization strategy, request plots, or outline statistical comparisons, and the model can generate code in Python, R, or another preferred language 22 .”
Quote in the paperSubstantial claim1 of 1 sources checked
- cites [22] Ruta, M. R., Gaidici, T., Irwin, C. & Lifshitz, J. ChatGPT for Univariate Statistics: Validation of AI-Assisted Data Analysis in Healthcare Research. J. Med. Internet Res. 27, e63550 (2025). source
  The Methods > Inferential Statistics section (page unavailable) shows users describing variables and research questions and specifying statistical tasks. The Introduction (page unavailable) discusses language-to-code translation, but the paper does not establish the full workflow claimed: specifying normalization, requesting plots, and generating code in Python, R, or any preferred language. Its use of Python, SAS, and RStudio in Methods > Response Grading was for comparison outputs, not evidence of user-selected code generation. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
May combine two claims.
Premisecitation
The same surveys reveal widespread uncertainty about how to begin, with many researchers reporting limited familiarity with existing tools, limited confidence in their ability to use them responsibly, and limited institutional guidance on best practices.
“Second, the same surveys reveal widespread uncertainty about how to begin.Many researchers report limited familiarity with existing tools, limited confidence in their own ability to use them responsibly, and limited institutional guidance on best practices.”
Quote in the paperSubstantial claim3 of 9 sources checked
- cites [4] Van Noorden, R. & Perkel, J. M. AI and science: what 1,600 researchers think. Nature 621, 672–675 (2023). source
- cites [5] Ng, J. Y. et al. Attitudes and perceptions of medical researchers towards the use of artificial intelligence chatbots in the scientific process: an international cross-sectional survey. Lancet Digit. Health 7, e94–e102 (2025). source
  The Abstract (p. 2) and Significance of Findings (p. 18) support limited institutional guidance: few respondents reported institutional training or policies, and many wanted training. But the Abstract says most respondents were familiar with AI chatbots, and the retrieved passages do not report limited confidence in their own ability to use them responsibly. Concern about the tools’ limitations is not the same as a reported lack of self-confidence. (LLM, conf=0.86)
  Experimental — accuracy not yet measured
- cites [6] Ng, J. Y. et al. Attitudes and perceptions of biomedical journal editors in chief towards the use of artificial intelligence chatbots in the scholarly publishing process: a cross-sectional survey. Res. Integr. Peer Rev. 10, 19 (2025). source
  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
  Experimental — accuracy not yet measured
- cites [7] Nordling, L. How ChatGPT is transforming the postdoc experience. Nature 622, 655–657 (2023). source
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
- cites [9] Researchers and AI: Survey Findings. https://academic.oup.com/pages/ai-survey-findings.
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [10] Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [11] ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
  Not checked in the free run.
- cites [12] Ruediger, D., McCracken, C. & Skinner, M. Adoption of Generative AI by Academic Biomedical Researchers. http://sr.ithaka.org/?p=321415 (2024) https://doi.org/10.18665/sr.321415. source
  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.
  Experimental — accuracy not yet measured
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given
Most of these systems are subscription-based.
“These systems combine literature search, data analysis, and writing in one interface, and most are subscription-based, more expensive, and semi-autonomous (Table 5).”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingreasoning
Most of these systems are more expensive.
“These systems combine literature search, data analysis, and writing in one interface, and most are subscription-based, more expensive, and semi-autonomous (Table 5).”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationreasoning
The model can produce corresponding code, annotate each step, and package all commands into a structured notebook that ensures full reproducibility.
“The model can produce the corresponding code, annotate each step, and package all commands into a structured notebook that ensures full reproducibility.”
- Also stated in: § Pearls Pitfalls · plain
  “Code is delivered annotated and structured into reproducible notebooks, reducing copy-paste and overwrite errors common in manual analysis.”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning
Despite caveats, AI systems for literature search and synthesis show potential to become a transformative addition to the scientific workflow.
“Despite these caveats, AI systems for literature search and synthesis show potential to become a transformative addition to the scientific workflow.”
Quote in the paperNot a substantial claim
- Rests on an assumption: The rapid growth of the amount and kinds of scientific publishing has made it increasingly difficult for researchers to remain current, even within specialized subfields.
- Rests on an assumption: AI tools designed for literature search and synthesis offer a promising solution by accelerating the discovery of relevant papers.
Classified as probably not a substantial claim.
Interpretationreasoning
LLMs are rapidly reshaping the practice of biomedical science.
“Large language models (LLMs) are generative artificial intelligence (AI) models that are rapidly reshaping the practice of biomedical science.”
Quote in the paperSubstantial claim
- Rests on an assumption: LLMs can synthesize literature, generate analytical code, and interface with multimodal data.
Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone given
These capabilities facilitate more comprehensive literature reviews and can support early-stage hypothesis generation.
“These capabilities facilitate more comprehensive literature reviews and can support early-stage hypothesis generation.”
Quote in the paperNot a substantial claim
- Rests on an assumption: AI-enabled literature search systems can identify relevant publications not only through keyword matching but also through semantic similarity, allowing users to surface conceptually related work even when terminology varies across disciplines.
- Rests on an assumption: Some tools further integrate citation graphs, enabling users to visualize how ideas propagate through the literature and to identify influential or under recognized contributions.
Classified as probably not a substantial claim.
Interpretationreasoning
It provides a clear record of each stage of data processing, which can facilitate peer review, collaboration, and later reanalysis.
“It also provides a clear record of each stage of data processing, which can facilitate peer review, collaboration, and later reanalysis.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S4 (based).
Interpretationreasoning
Together, administrative applications of AI can immediately streamline daily work without altering the scientific content itself.
“Together, these administrative applications illustrate how AI can immediately streamline daily work without altering the scientific content itself.”
Quote in the paperNot a substantial claim
- Rests on an assumption: LLMs excel at adjusting tone, audience, and level of technical detail.
- Rests on an assumption: This ability makes LLMs useful for summarizing complex topics for nonspecialists.
- Rests on an assumption: A technical abstract can be rewritten for clinicians, students, or the general public simply by specifying the target audience.
- Rests on an assumption: AI can draft initial versions of social media posts, ensuring accuracy while allowing researchers to quickly adapt messages across platforms.
Probably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationreasoning
When used judiciously, LLM summarization and interactive exploration capabilities can dramatically reduce the time required to engage with new literature, especially for long or technically dense manuscripts.
“When used judiciously, these capabilities can dramatically reduce the time required to engage with new literature, especially when dealing with long or technically dense manuscripts.”
Quote in the paperNot a substantial claim
- Rests on an assumption: LLMs can assist with summarizing the content of relevant papers.
- Rests on an assumption: Many tools allow users to upload full text PDFs and request structured summaries, explanations of key findings, comparisons across papers, or simplified descriptions for trainees and nonspecialists.
- Rests on an assumption: These models can also support more interactive exploration, such as asking followup questions, clarifying methodological details, or translating unfamiliar terminology into accessible language.
May combine two claims.
Interpretationreasoning
AI can enhance data analysis by increasing accessibility, improving reproducibility, and providing real-time explanatory support.
“Together, these capabilities illustrate how AI can enhance data analysis by increasing accessibility, improving reproducibility, and providing real time explanatory support.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning
Semantic search offers clear advantages in biomedical research, where parallel discoveries often emerge from disparate communities and are published using heterogeneous language.
“This offers clear advantages in biomedical research, where parallel discoveries often emerge from disparate communities and are published using heterogeneous language.”
Quote in the paperNot a substantial claim
- Rests on an assumption: AI-enabled literature search systems can identify relevant publications not only through keyword matching but also through semantic similarity, allowing users to surface conceptually related work even when terminology varies across disciplines.
Classified as probably not a substantial claim.
Premisecitation
Researchers across disciplines express strong concerns regarding accuracy, bias, reproducibility, and the potential erosion of critical thinking if AI generated content is accepted without careful evaluation or potentially even used at all.
“Third, researchers across disciplines express strong concerns regarding accuracy, bias, reproducibility, and the potential erosion of critical thinking if AI generated content is accepted without careful evaluation or potentially even used at all 13 .”
Quote in the paperSubstantial claim0 of 1 sources checked
- cites [13] Messeri, L. & Crockett, M. J. The uncritical adoption of AI in science is alarming — we urgently need guard rails. Nature 653, 675–676 (2026). source
Classified as probably not a substantial claim.
Interpretationreasoning
Automated stepwise organization reduces the risk of common human errors such as accidental overwriting of variables or copy-paste mistakes during manual analysis.
“This automated, stepwise organization reduces the risk of common human errors such as accidental overwriting of variables or copy-paste mistakes during manual analysis.”
- Also stated in: § Pearls Pitfalls · plain
  “Code is delivered annotated and structured into reproducible notebooks, reducing copy-paste and overwrite errors common in manual analysis.”
Quote in the paperSubstantial claim
May combine two claims.
Interpretationreasoning
Integration of LLMs into scientific workflows remains irregular, and the field lacks clear guidance for reliable and productive use.
“Yet their integration into scientific workflows remains irregular, and the field lacks clear guidance for reliable and productive use.”
Quote in the paperSubstantial claim
- Rests on an assumption: LLMs can synthesize literature, generate analytical code, and interface with multimodal data.
May combine two claims.
Interpretationreasoning
This approach enables LLMs to write syntactically correct code and apply field-specific analytical standards and leverage domain-appropriate tools that may not have been extensively represented in their original training data.
“This approach enables LLMs to not only write syntactically correct code, but to apply field-specific analytical standards and leverage domain-appropriate tools that may not have been extensively represented in their original training data.”
- Also stated in: § Pearls Pitfalls · plain
  “Domain-specific skills apply fieldappropriate methods, not generic boilerplate.”
Quote in the paperSubstantial claim
The quote only partly states this claim.
Premisecitation
Survey data show a sharp distinction between editing and drafting in AI use.
“Within the domain of writing, survey data show a sharp distinction between editing and drafting.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
May combine two claims.
Premisecitation
Roughly 28% of researchers have used AI to edit a manuscript, and only 8% report using AI to write a first draft.
“Roughly 28% of researchers say they have used AI to edit a manuscript, and only 8% report using AI to write a first draft 8 .”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
May combine two claims.
Premisecitation
In the Wiley ExplanAItions studies, more than 60 percent of researchers selected lack of training or clear guidelines as a major barrier to adoption, and many noted that they were unsure which tools are safe or appropriate for scientific work.
“In the Wiley ExplanAItions studies, more than 60 percent of researchers selected lack of training or clear guidelines as a major barrier to adoption, and many noted that they were unsure which tools are safe or appropriate for scientific work 10,11 .”
Quote in the paperSubstantial claim0 of 2 sources checked
- cites [10] Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [11] ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
  Not checked in the free run.
Probably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationreasoning
AI systems for literature search and synthesis increase accessibility to the biomedical literature, help researchers navigate the ever-expanding volume of publications, and support deeper conceptual integration across fields.
“They increase accessibility to the biomedical literature, help researchers navigate the ever-expanding volume of publications, and support deeper conceptual integration across fields.”
Quote in the paperNot a substantial claim
- Rests on an assumption: The rapid growth of the amount and kinds of scientific publishing has made it increasingly difficult for researchers to remain current, even within specialized subfields.
- Rests on an assumption: AI tools designed for literature search and synthesis offer a promising solution by accelerating the discovery of relevant papers.
May combine two claims.
Interpretationreasoning
This supports education, debugging, and code review within interdisciplinary teams.
“This supports education, debugging, and code review within interdisciplinary teams.”
- Also stated in: § Pearls Pitfalls · plain
  “Models can explain unfamiliar functions and statistics in plain language, supporting trainees and interdisciplinary review.”
Quote in the paperSubstantial claim
May combine two claims.
Interpretationreasoning
This capability allows researchers with limited coding experience to perform analyses that would previously have required expert support, while also accelerating the workflow of experienced programmers by reducing coding time by perhaps an order of magnitude or greater.
“This capability allows researchers with limited coding experience to perform analyses that would previously have required expert support, while also accelerating the workflow of experienced programmers by reducing coding time by perhaps an order of magnitude or greater.”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation
Funding agencies generally permit AI-assistance with disclosure, while also adjusting submission policies to manage increased application volume.
“Funding agencies generally permit AIassistance with disclosure, while also adjusting submission policies to manage increased application volume (for example, see NOT-OD-25-132 from the NIH).”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites NOT-OD-25-132
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Probably not a substantial claim: may not pass S4 (based). May combine two claims.
Interpretationreasoning
Plain-language prompts produce executable Python or R, lowering the barrier for bench scientists.
“Plain-language prompts produce executable Python or R, lowering the barrier for bench scientists.”
Quote in the paperNot a substantial claim
May combine two claims.
Premisecitation
Publishers have started penalizing the submission of fabricated references; arXiv imposes a one-year ban on authors who submit hallucinated citations.
“Publishers have started penalizing the submission of fabricated references: arXiv, for instance, imposes a one-year ban on authors who submit hallucinated citations 33 .”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [33] Singh Chawla, D. Researchers who use hallucinated references to face arXiv ban. Nature https://doi.org/10.1038/d41586-026-01595-5 (2026). source
Classified as probably not a substantial claim.
Interpretationreasoning
If an AI-assisted pipeline reproduces the original findings from a manually processed dataset, this provides confidence in both the generated code and the researcher's ability to evaluate it.
“If the AIassisted pipeline reproduces the original findings, this provides confidence in both the generated code and the researcher's ability to evaluate it.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.
Interpretationreasoning
By adopting AI for transcription, note generation, and nonscientific writing, biomedical scientists can reclaim time for high-value intellectual work while building familiarity with tools that will later support more complex research tasks.
“By adopting AI for transcription, note generation, and nonscientific writing, biomedical scientists can reclaim time for high-value intellectual work while building familiarity with tools that will later support more complex research tasks.”
Quote in the paperNot a substantial claim
- Rests on an assumption: These administrative applications can immediately streamline daily work without altering the scientific content itself.
- Rests on an assumption: By adopting AI for transcription, note generation, and nonscientific writing, biomedical scientists can reclaim time for high-value intellectual work.
- Rests on an assumption: Meeting transcripts into an LLM produces structured notes and action items.
- Rests on an assumption: One source text can be adapted for multiple audiences by specifying tone and level.
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation
The Nature Kwon study is likely low due to specifically asking about use in writing.
“the Nature Kwon study 8 is likely low due to specifically asking about use in writing.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation
Among those comfortable with AI-assisted drafting, many believe disclosure is necessary, and some feel authors should disclose the prompts used when generating text.
“Among those who are comfortable with AI-assisted drafting, many believe that disclosure is necessary, and some feel that authors should disclose the prompts used when generating text.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning
These interpretive capabilities do not remove the need for domain expertise.
“However, these interpretive capabilities do not remove the need for domain expertise.”
Quote in the paperNot a substantial claim
- Rests on an assumption: AI models can misinterpret statistical results or overstate conclusions if not carefully guided.
- Rests on an assumption: AI models may propose analytical approaches that are statistically inappropriate for a given study design.
Probably not a substantial claim: may not pass S4 (based).
Interpretationreasoning
When paired with careful human oversight and rigorous source verification, AI literature tools can significantly enhance both the efficiency and the quality of scientific scholarship.
“When paired with careful human oversight and rigorous source verification, these tools can significantly enhance both the efficiency and the quality of scientific scholarship.”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Premisereasoning
The literature on AI for biomedical research is thin and mostly limited to preprints.
“AI tools evolve faster than peer review, and developers have little incentive to publish, so the literature on AI for biomedical research is thin and mostly limited to preprints.”
Quote in the paperNot a substantial claim
- Rests on an assumption: AI tools evolve faster than peer review, and developers have little incentive to publish.
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning
Without a base of critical thinking, the verification step becomes hollow because the trainee has no internal standard against which to check the AI's output.
“Without that base, the verification step we have argued for throughout this article becomes hollow, because the trainee has no internal standard against which to check the AI's output.”
Quote in the paperNot a substantial claim
May combine two claims.
Premisecitation
About 10% of researchers believe AI use is never appropriate for editing, while 35% believe it is never appropriate for writing a first draft.
“While most people are open to using AI for these tasks, about 10% of researchers believe AI use is never appropriate for editing, while 35% believe it is never appropriate for writing a first draft 8 .”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [8] Kwon, D. Is it OK for AI to write science papers? Nature survey shows researchers are split. Nature 641, 574–578 (2025). source
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning
The human researcher remains accountable for every claim, citation, figure, and analysis, regardless of how it was generated.
“The human researcher remains accountable for every claim, citation, figure, and analysis, regardless of how it was generated.”
Quote in the paperNot a substantial claim

46 of 46

The paper relies on this without support.
AI tools evolve faster than peer review, and developers have little incentive to publish.
“AI tools evolve faster than peer review, and developers have little incentive to publish”
The paper relies on this without support.
Administrative responsibilities are often time intensive and usually do not require deep expertise.
“Administrative responsibilities are often time intensive tasks that usually do not require deep expertise.”
The paper relies on this without support.
LLMs can synthesize, organize, and summarize large volumes of text to speed up various writing tasks.
“In general, scientists can capitalize on the ability of LLMs to synthesize, organize, and summarize large volumes of text to speed up various writing tasks.”
The paper relies on this without support.
AI-supported meeting transcription tools embedded in Zoom, Microsoft Teams, and Google Meet can automatically generate transcripts that capture discussions with remarkable accuracy.
“Tools embedded within Zoom, Microsoft Teams, and Google Meet can automatically generate transcripts that capture discussions with remarkable accuracy.”
The paper relies on this without support.
Transcripts from these tools can be exported into an LLM, which can then produce structured meeting notes, summarize key decisions, and extract actionable next steps tailored to specific team members.
“These transcripts can be exported directly into an LLM, which can then produce structured meeting notes, summarize key decisions, and extract actionable next steps tailored to specific team members.”
The paper relies on this without support.
This approach reduces cognitive load during meetings, freeing attention to the topics at hand, and provides documentation that is more reliable than handwritten notes.
“This approach reduces cognitive load during meetings, freeing attention to the topics at hand, and provides documentation that is more reliable than handwritten notes.”
The paper relies on this without support.
For busy labs juggling multiple projects, this can profoundly reduce the administrative burden of project management while improving communication across teams.
“For busy labs juggling multiple projects, this can profoundly reduce the administrative burden of project management while improving communication across teams.”
The paper relies on this without support.
LLMs can synthesize literature, generate analytical code, and interface with multimodal data.
“Their ability to synthesize literature, generate analytical code, and interface with multimodal data offers a new framework for accelerating discovery.”
The paper relies on this without support.
LLMs excel at adjusting tone, audience, and level of technical detail.
“LLMs excel at adjusting tone, audience, and level of technical detail, which makes them particularly useful for summarizing complex topics for nonspecialists.”
The paper relies on this without support.
This ability makes LLMs useful for summarizing complex topics for nonspecialists.
“LLMs excel at adjusting tone, audience, and level of technical detail, which makes them particularly useful for summarizing complex topics for nonspecialists.”
The paper relies on this without support.
A technical abstract can be rewritten for clinicians, students, or the general public simply by specifying the target audience.
“A technical abstract can be rewritten for clinicians, students, or the general public simply by specifying the target audience.”
The paper relies on this without support.
AI can draft initial versions of social media posts, ensuring accuracy while allowing researchers to quickly adapt messages across platforms.
“AI can Table 1 | Examples of ways to use LLMs for biomedical administrative tasks draft initial versions of these posts, ensuring accuracy while allowing researchers to quickly adapt messages across platforms.”
The paper relies on this without support.
These administrative applications can immediately streamline daily work without altering the scientific content itself.
“Together, these administrative applications illustrate how AI can immediately streamline daily work without altering the scientific content itself.”
The paper relies on this without support.
By adopting AI for transcription, note generation, and nonscientific writing, biomedical scientists can reclaim time for high-value intellectual work.
“By adopting AI for transcription, note generation, and nonscientific writing, biomedical scientists can reclaim time for high-value intellectual work while building familiarity with tools that will later support more complex research tasks.”
The paper relies on this without support.
Meeting transcripts into an LLM produces structured notes and action items.
“Meeting transcripts into an LLM produces structured notes and action items.”
The paper relies on this without support.
One source text can be adapted for multiple audiences by specifying tone and level.
“One source text can be adapted for multiple audiences by specifying tone and level.”
The paper relies on this without support.
The rapid growth of the amount and kinds of scientific publishing has made it increasingly difficult for researchers to remain current, even within specialized subfields.
“The rapid growth of the amount and kinds of scientific publishing has made it increasingly difficult for researchers to remain current, even within specialized subfields.”
The paper relies on this without support.
AI tools designed for literature search and synthesis offer a promising solution by accelerating the discovery of relevant papers.
“AI tools designed for literature search and synthesis offer a promising solution by accelerating the discovery of relevant papers, highlighting conceptual relationships across domains, and reducing the time required to extract meaningful insights from published work.”
The paper relies on this without support.
AI-enabled literature search systems can identify relevant publications not only through keyword matching but also through semantic similarity, allowing users to surface conceptually related work even when terminology varies across disciplines.
“AI-enabled literature search systems can identify relevant publications not only through keyword matching (like in a PubMed search) but also through semantic similarity, allowing users to surface conceptually related work even when terminology varies across disciplines.”
The paper relies on this without support.
Some tools further integrate citation graphs, enabling users to visualize how ideas propagate through the literature and to identify influential or under recognized contributions.
“Some tools further integrate citation graphs, enabling users to visualize how ideas propagate through the literature and to identify influential or under recognized contributions.”
The paper relies on this without support.
LLMs can assist with summarizing the content of relevant papers.
“Once relevant papers are identified, LLMs can assist with summarizing their content.”
The paper relies on this without support.
Many tools allow users to upload full text PDFs and request structured summaries, explanations of key findings, comparisons across papers, or simplified descriptions for trainees and nonspecialists.
“Many tools allow users to upload full text PDFs and request structured summaries, explanations of key findings, comparisons across papers, or simplified descriptions for trainees and nonspecialists.”
The paper relies on this without support.
These models can also support more interactive exploration, such as asking followup questions, clarifying methodological details, or translating unfamiliar terminology into accessible language.
“These models can also support more interactive exploration, such as asking followup questions, clarifying methodological details, or translating unfamiliar terminology into accessible language.”
The paper relies on this without support.
Traditionally, computational analysis has required specialized programming skills, creating a barrier for many experimental scientists.
“Traditionally, computational analysis has required specialized programming skills, creating a barrier for many experimental scientists.”
The paper relies on this without support.
AI models can misinterpret statistical results or overstate conclusions if not carefully guided.
“AI models can misinterpret statistical results or overstate conclusions if not carefully guided, and they may propose analytical approaches that are statistically inappropriate for a given study design.”
The paper relies on this without support.
AI models may propose analytical approaches that are statistically inappropriate for a given study design.
“AI models can misinterpret statistical results or overstate conclusions if not carefully guided, and they may propose analytical approaches that are statistically inappropriate for a given study design.”
The paper relies on this without support.
Models hallucinate library functions and miss deep logical bugs while fixing surface errors.
“Models hallucinate library functions and miss deep logical bugs while fixing surface errors; generated code needs review, not blind execution.”
The paper relies on this without support.
Heavy reliance on LLMs creates cognitive distance from the codebase, leaving researchers unable to debug or defend their own analyses.
“Heavy reliance creates cognitive distance from the codebase, leaving researchers unable to debug or defend their own analyses.”
The paper relies on this without support.
There is some stigma associated with AI use.
“Because human readers have imperfect discrimination of human vs. AI-generated text 25 and there is some stigma associated with AI use, algorithmic approaches have become central for estimating LLM usage in scientific writing.”
The paper relies on this without support.
The LLM can assist with generating or refining text at the paragraph scale, helping improve clarity, transitions, and tone while preserving the scientific content and reasoning supplied by the researcher.
“The LLM can then assist with generating or refining text at the paragraph scale, helping improve clarity, transitions, and tone while preserving the scientific content and reasoning supplied by the researcher.”
The paper relies on this without support.
This paragraph-by-paragraph method maintains human intellectual ownership while leveraging AI efficiency for linguistic polish.
“This paragraph-by-paragraph method maintains human intellectual ownership while leveraging AI efficiency for linguistic polish.”
The paper relies on this without support.
General LLMs frequently hallucinate references, producing fabricated or mismatched citations that seem credible but do not correspond to real papers.
“General LLMs frequently hallucinate references, producing fabricated or mismatched citations that seem credible but do not correspond to real papers.”
“LLMs can fabricate citations; references should be inserted manually.”
The paper relies on this without support.
Paragraph-level human outlines paired with LLM polishing preserve intellectual ownership while improving clarity, transitions, and tone.
“Paragraph-level human outlines paired with LLM polishing preserve intellectual ownership while improving clarity, transitions, and tone.”
The paper relies on this without support.
Adversarial editing prompts (skeptical reviewer, statistical reviewer, novelty reviewer) identify weak claims before submission.
“Adversarial editing prompts (skeptical reviewer, statistical reviewer, novelty reviewer) identify weak claims before submission.”
The paper relies on this without support.
Cross-document consistency checks catch mismatches and internal contradictions.
“Cross-document consistency checks catch mismatches and internal contradictions.”
The paper relies on this without support.
Generated text can drift toward overclaiming when not explicitly constrained to the evidence supplied.
“Generated text can drift toward overclaiming when not explicitly constrained to the evidence supplied.”
The paper relies on this without support.
LLM generated text defaults to generic phrasing, which can result in stylistic homogenization and loss of scientific specificity.
“LLM generated text defaults to generic phrasing, which can result in stylistic homogenization and loss of scientific specificity.”
The paper relies on this without support.
LLMs can reduce administrative burden, assist with literature navigation, generate or explain code, and support clearer scientific writing.
“They can reduce administrative burden, assist with literature navigation, generate or explain code, and support clearer scientific writing.”
The paper relies on this without support.
These benefits are only realized when researchers remain active participants in the process.
“Yet these benefits are only realized when researchers remain active participants in the process.”
The paper relies on this without support.
AI generated outputs often appear logically sound and authoritative even when they contain subtle or major errors.
“AI generated outputs often appear logically sound and authoritative even when they contain subtle or major errors.”
The paper relies on this without support.
A scientist who accepts these outputs without verification can introduce mistakes that negate any time savings and may compromise the reliability of downstream work.
“A scientist who accepts these outputs without verification can introduce mistakes that negate any time savings and may compromise the reliability of downstream work.”
The paper relies on this without support.
LLMs inherit the biases of their training corpora, which can skew literature retrieval toward well-represented populations, languages, and journals and underweight work from underresourced settings.
“LLMs inherit the biases of their training corpora, which can skew literature retrieval toward well-represented populations, languages, and journals and underweight work from underresourced settings.”
The paper relies on this without support.
Uploading patient data, unpublished results, or proprietary code to general-purpose models can expose information to retention, training, or third-party access.
“Uploading patient data, unpublished results, or proprietary code to general-purpose models can expose information to retention, training, or third-party access.”
The paper relies on this without support.
Used thoughtfully, these tools can accelerate learning, improve documentation, and expose researchers to analytical or conceptual approaches they may not have encountered otherwise.
“Used thoughtfully, these tools can accelerate learning, improve documentation, and expose researchers to analytical or conceptual approaches they may not have encountered otherwise.”
The paper relies on this without support.
Used passively, however, they risk eroding the development of essential scientific skills.
“Used passively, however, they risk eroding the development of essential scientific skills.”
The paper relies on this without support.
This risk is greatest for trainees.
“This risk is greatest for trainees.”
The paper relies on this without support.
A senior researcher who turns to an LLM for a literature scan or a quick analysis script brings years of judgment to the output: they know what a credible citation looks like, where a statistical choice is suspect, what a clean experimental design feels like.
“A senior researcher who turns to an LLM for a literature scan or a quick analysis script brings years of judgment to the output: they know what a credible citation looks like, where a statistical choice is suspect, what a clean experimental design feels like.”
The paper relies on this without support.
A graduate student or early-career scientist who learns the field through AI-mediated workflows may never build that judgment in the first place.
“A graduate student or early-career scientist who learns the field through AI-mediated workflows may never build that judgment in the first place.”
The paper relies on this without support.
While some technical abilities have historically faded without consequence as tools advanced, not all skills at risk here are expendable.
“While some technical abilities have historically faded without consequence as tools advanced, not all skills at risk here are expendable.”
The paper relies on this without support.
Scientific progress depends on critical thinking, creative hypothesis generation, and the ability to synthesize, interrogate, and defend evidence.
“Scientific progress depends on critical thinking, creative hypothesis generation, and the ability to synthesize, interrogate, and defend evidence.”
The paper relies on this without support.
The long-term reproducibility of biomedical research depends on a generation of scientists who can still do the work without the tool, even as the tools become indispensable.
“The long-term reproducibility of biomedical research depends on a generation of scientists who can still do the work without the tool, even as the tools become indispensable.”

51 of 51

Attention Is All You Need
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1× · see citing works
“…intelligence has rapidly evolved from a technical concept within computer science to a widely used tool in research, enabled by transformer-based architectures in 2017 that support large language models (LLMs) with far greater capacity for reasoning, summarization, and pattern recognition 1,2 .”
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. (2025) · ACM Transactions on Information Systems · doi:10.1145/3703155 · Cited by 2,142 · see citing works
Artificial hallucinations in ChatGPT: implications in scientific writing
Alkaissi, H., & McFarlane, S. I. (2023) · Cureus · doi:10.7759/cureus.35179 · Cited by 1,540 · see citing works
Evaluating Large Language Models Trained on Code
Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. D. O., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., et al. (2021) · arXiv · doi:10.48550/arxiv.2107.03374 · Cited by 1,436 · cited 1× · see citing works
“Modern LLMs substantially lower this barrier by enabling researchers to describe analytical tasks in plain language and receive executable code in return 20,21 .”
Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers
Gao, C. A., Howard, F. M., Markov, N. S., Dyer, E. C., Ramesh, S., Luo, Y., & Pearson, A. T. (2023) · npj Digital Medicine · doi:10.1038/s41746-023-00819-6 · Cited by 1,099 · see citing works
Tools such as ChatGPT threaten transparent science; here are our ground rules for their use
(2023) · Nature · doi:10.1038/d41586-023-00191-1 · Cited by 605 · see citing works
Fabrication and errors in the bibliographic citations generated by ChatGPT
Walters, W. H., & Wilder, E. I. (2023) · Scientific Reports · doi:10.1038/s41598-023-41032-5 · Cited by 499 · see citing works
Nonhuman “Authors” and Implications for the Integrity of Scientific Publication and Medical Knowledge
Flanagin, A., Bibbins-Domingo, K., Berkwits, M., & Christiansen, S. L. (2023) · JAMA · doi:10.1001/jama.2023.1344 · Cited by 420 · see citing works
Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis
Chelli, M., Descamps, J., Lavoué, V., Trojani, C., Azar, M., Deckert, M., Raynier, J.-L., Clowez, G., Boileau, P., & Ruetsch-Chelli, C. (2024) · Journal of Medical Internet Research · doi:10.2196/53164 · Cited by 413 · cited 1× · see citing works
“A frequently encountered hallucination (i.e. con-fabulation)inthiscontextinvolvesfabricatedcitations 14 – 18 .”
AI and science: what 1,600 researchers think
Van Noorden, R., & Perkel, J. M. (2023) · Nature · doi:10.1038/d41586-023-02980-0 · Cited by 383 · cited 1× · see citing works
“Surveys across the scienti ﬁ c community consistently show three dominant themes 4 – 12 .”
High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content
Bhattacharyya, M., Miller, V. M., Bhattacharyya, D., & Miller, L. E. (2023) · Cureus · doi:10.7759/cureus.39238 · Cited by 294 · see citing works
The ethics of using artificial intelligence in scientific research: new guidance needed for a new tool
Resnik, D. B., & Hosseini, M. (2025) · AI and Ethics · doi:10.1007/s43681-024-00493-8 · Cited by 228 · see citing works
Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025) · arXiv · doi:10.48550/arxiv.2506.08872 · Cited by 220 · see citing works
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2024) · arXiv · doi:10.48550/arxiv.2406.07016 · Cited by 214
Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025) · Science Advances · doi:10.1126/sciadv.adt3813 · Cited by 213 · cited 1× · see citing works
“Another study which estimated LLM usage in scienti ﬁ c writing by quantitating vocabulary shifts suggested that 13.5% of all 2024 PubMed abstracts were processed with LLMs 27 .”
Influential
Autonomous biomedical research with an artificial intelligence agent
Huang, K., Zhang, S., Wang, H., Qu, Y., Lu, Y., Li, R., Roohani, Y., Qiu, L., Cao, S., Li, G. L., et al. (2026) · Science · doi:10.1126/science.adz4351 · Cited by 198 · see citing works
Reference Hallucination Score for Medical Artificial Intelligence Chatbots: Development and Usability Study (Preprint)
(2023) · doi:10.2196/preprints.54345 · Cited by 175 · see citing works
Reference Hallucination Score for Medical Artificial Intelligence Chatbots: Development and Usability Study
Aljamaan, F., Temsah, M.-H., Altamimi, I., Al-Eyadhy, A., Jamal, A., Alhasan, K., Mesallam, T. A., Farahat, M., & Malki, K. H. (2024) · JMIR Medical Informatics · doi:10.2196/54345 · Cited by 166 · see citing works
Accelerating scientific discovery with Co-Scientist
Gottweis, J., Weng, W., Daryin, A., Tu, T., Sirkovic, P., Myaskovsky, A., Glowaty, G., Weissenberger, F., Orlandi, A., Popovici, D., et al. (2026) · Nature · doi:10.1038/s41586-026-10644-y · Cited by 158 · see citing works
A multi-agent system for automating scientific discovery
Ghareeb, A. E., Chang, B., Mitchener, L., Angela, Y., Szostkiewicz, C. J., Shved, D., Gyimesi, G. J., Laurent, J. M., Wright, S. M., Razzak, M. T., et al. (2026) · Nature · doi:10.1038/s41586-026-10652-y · Cited by 146 · see citing works
How ChatGPT and other AI tools could disrupt scientific publishing
Conroy, G. (2023) · Nature · doi:10.1038/d41586-023-03144-w · Cited by 121 · see citing works
Towards end-to-end automation of AI research
Lu, C., Lu, C., Lange, R. T., Yamada, Y., Hu, S., Foerster, J., Ha, D., & Clune, J. (2026) · Nature · doi:10.1038/s41586-026-10265-5 · Cited by 116 · see citing works
Is ChatGPT the Ultimate Programming Assistant -- How far is it?
Tian, H., Lu, W., Li, T. O., Tang, X., Cheung, S.-C., Klein, J., & Bissyandé, T. F. (2023) · arXiv · doi:10.48550/arxiv.2304.11938 · Cited by 114 · see citing works
Is it OK for AI to write science papers? Nature survey shows researchers are split
Kwon, D. (2025) · Nature · doi:10.1038/d41586-025-01463-8 · Cited by 101 · see citing works
Techniques for supercharging academic writing with generative AI
Lin, Z. (2024) · Nature Biomedical Engineering · doi:10.1038/s41551-024-01185-8 · Cited by 87 · cited 1× · see citing works
“A structured work ﬂ ow can help researchers integrate AI into scienti ﬁ c writing while maintaining rigor 32 .”
Scientific production in the era of large language models
Kusumegi, K., Yang, X., Ginsparg, P., de Vaan, M., Stuart, T., & Yin, Y. (2025) · Science · doi:10.1126/science.adw3000 · Cited by 83 · see citing works
Quantifying large language model usage in scientific papers
Liang, W., Zhang, Y., Wu, Z., Lepp, H., Ji, W., Zhao, X., Cao, H., Liu, S., He, S., Huang, Z., et al. (2025) · Nature Human Behaviour · doi:10.1038/s41562-025-02273-8 · Cited by 80 · see citing works
How ChatGPT is transforming the postdoc experience
Nordling, L. (2023) · Nature · doi:10.1038/d41586-023-03235-8 · Cited by 71 · see citing works
Attitudes and perceptions of medical researchers towards the use of artificial intelligence chatbots in the scientific process: an international cross-sectional survey
Ng, J. Y., Maduranayagam, S. G., Suthakar, N., Li, A., Lokker, C., Iorio, A., Haynes, R. B., & Moher, D. (2025) · The Lancet Digital Health · doi:10.1016/s2589-7500(24)00202-4 · Cited by 64 · see citing works
Mapping the Increasing Use of LLMs in Scientific Papers
Liang, W., Zhang, Y., Wu, Z., Lepp, H., Ji, W., Zhao, X., Cao, H., Liu, S., He, S., Huang, Z., et al. (2024) · arXiv · doi:10.48550/arxiv.2404.01268 · Cited by 55 · see citing works
ChatGPT for Univariate Statistics: Validation of AI-Assisted Data Analysis in Healthcare Research
Ruta, M. R., Gaidici, T., Irwin, C., & Lifshitz, J. (2025) · Journal of Medical Internet Research · doi:10.2196/63550 · Cited by 24 · cited 1× · see citing works
“Users can describe the structure of their dataset, specify the desired normalization strategy, request plots, or outline statistical comparisons, and the model can generate code in Python, R, or another preferred language 22 .”
A Comparative Study of Code Generation using ChatGPT 3.5 across 10 Programming Languages
Buscemi, A. (2023) · arXiv · doi:10.48550/arxiv.2308.04477 · Cited by 18 · cited 1× · see citing works
“Discrepancies can highlight either shortcomings in the AI-generated approach or opportunities to re ﬁ ne the analysis 24 .”
Attitudes and perceptions of biomedical journal editors in chief towards the use of artificial intelligence chatbots in the scholarly publishing process: a cross-sectional survey
Ng, J. Y., Krishnamurthy, M., Deol, G., Al-Khafaji, W. A.-Z., Balaji, V., Abebe, M., Adhvaryu, J., Karrthik, T., Mohanakanthan, P., Vellaparambil, A., et al. (2025) · Research Integrity and Peer Review · doi:10.1186/s41073-025-00178-8 · Cited by 14 · see citing works
Attitudes and Perceptions of Biomedical Journal Editors-in-Chief Towards the Use of Artificial Intelligence Chatbots in the Scholarly Publishing Process: A Cross-Sectional Survey
Ng, J. Y., Krishnamurthy, M., Deol, G., Al-Khafaji, W. A.-Z., Balaji, V., Abebe, M., Adhvaryu, J., Karrthik, T., Mohanakanthan, P., Vellaparambil, A., et al. (2025) · doi:10.1101/2025.05.26.25328101 · Cited by 14 · see citing works
Why AI cannot do good science without humans
(2026) · Nature · doi:10.1038/d41586-026-01551-3 · Cited by 10 · see citing works
Researchers who use hallucinated references to face arXiv ban
Singh Chawla, D. (2026) · Nature · doi:10.1038/d41586-026-01595-5 · Cited by 9 · cited 1× · see citing works
“Publishers have started penalizing the submission of fabricated references: arXiv, for instance, imposes a one-year ban on authors who submit hallucinated citations 33 .”
Adoption of Generative AI by Academic Biomedical Researchers
Ruediger, D., McCracken, C., & Skinner, M. (2024) · doi:10.18665/sr.321415 · Cited by 8 · cited 1× · see citing works
“Surveys across the scienti ﬁ c community consistently show three dominant themes 4 – 12 .”
The uncritical adoption of AI in science is alarming — we urgently need guard rails
Messeri, L., & Crockett, M. J. (2026) · Nature · doi:10.1038/d41586-026-01557-x · Cited by 7 · cited 1× · see citing works
“Third, researchers across disciplines express strong concerns regarding accuracy, bias, reproducibility, and the potential erosion of critical thinking if AI generated content is accepted without careful evaluation or potentially even used at all 13 .”
Defining and assessing AI literacy for researchers across the research lifecycle
Parker, J. L., & Becker, K. P. (2026) · Frontiers in Education · doi:10.3389/feduc.2026.1827603 · Cited by 5 · see citing works
An Evaluation of Large Language Models in Bioinformatics Research
Yin, H., Gu, Z., Wang, F., Abuduhaibaier, Y., Zhu, Y., Tu, X., Hua, X.-S., Luo, X., & Sun, Y. (2024) · arXiv · doi:10.48550/arxiv.2402.13714 · Cited by 4 · cited 1× · see citing works
“AI-assisted data analysis has emerged as one of the most powerful and democratizing applicationsof LLMsinscienti ﬁ c research 19 .”
OpenScientist: evaluating an open agentic AI co-scientist to accelerate biomedical discoveryPreprint
Roberts, K. F., Abrams, Z., Cappelletti, L., Moqri, M., Heugel, N., Caufield, J., Bourdenx, M., Li, Y., Banerjee, J., Foschini, L., et al. (2026) · medRxiv · doi:10.64898/2026.03.15.26348338 · Cited by 3 · see citing works
Self-citation
K-Dense Analyst: Towards Fully Automated Scientific AnalysisPreprint
Li, O., Agarwal, V., Zhou, S., et al. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2508.07043 · Cited by 1 · see citing works
Towards an AI biomedical scientist: Accelerating discoveries in neurodegenerative disease
Roberts, K. F., Landsness, E. C., Reese, J., Elbert, D., Strobel, G., Wu, E., Chen, Y., Lai, A., Abrams, Z. B., Zhu, M., et al. (2026) · The Journal of Prevention of Alzheimer's Disease · doi:10.1016/j.tjpad.2025.100398 · Cited by 1 · see citing works
Self-citation
Attitudes and Perceptions of Medical Researchers Towards the Use of Artificial Intelligence Chatbots in the Scientific Process: A Large-Scale, International Cross-Sectional Survey
Ng, J. Y., Maduranayagam, S. G., Suthakar, N., et al. (2024) · doi:10.1101/2024.02.27.24303462
Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis (Preprint)
Chelli, M., Descamps, J., Lavoué, V., Trojani, C., Azar, M., Deckert, M., Raynier, J.-L., Clowez, G., Boileau, P., & Ruetsch-Chelli, C. (2023) · doi:10.2196/preprints.53164

Flagged first, then by citations

Research works printed but not linked: 5

Radford, A., Narasimhan, K., Salimans, T. & Sutskever, I. Improving Language Understanding by Generative Pre-Training. Preprint at https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf (2018).
Researchers and AI: Survey Findings. https://academic.oup.com/pages/ai-survey-findings.
Hrycyshyn, Allison, E., Helen. ExplanAItions: An AI Study by Wiley. https://www.wiley.com/content/dam/wiley-com/en/pdfs/about/wiley-explanaItions-ai-study-february-2025vers3.pdf (2025).
ExplanAItions 2025: The Evolution of AI in Research. https://www.wiley.com/en-us/about-us/ai-resources/ai-study/ (2025).
Mitchener, L. et al. Kosmos: An AI Scientist for Autonomous Discovery. Preprint at https://pubmed.ncbi.nlm.nih.gov/42156546/ (2025).

Innovative Application of Large Language Models in Biology and Medicine Under Artificial Intelligence for Science Paradigm.
Cun, X.-L., Ding, C.-J., Chen, F., et al. (2025) · PubMed · doi:10.3881/j.issn.1000-503x.16605
Roadmap for using large language models (LLMs) to accelerate cross-disciplinary research with an example from computational biology.
Ke, R., & Ribeiro, R. M. (2026) · PubMed · doi:10.1007/s44163-026-01589-2
ChatGPT, Bard, and large language models for biomedical research: opportunities and pitfalls
Thapa, S., & Adhikari, S. (2023) · Annals of Biomedical Engineering · doi:10.1007/s10439-023-03284-0 · Cited by 111 · see citing works
Bridging the gap between scientists and large language models
Bozsó, I., Horvath, A., & Kuslits, L. (2026) · doi:10.5194/egusphere-egu26-17176
Large language models for biomedicine: foundations, opportunities, challenges, and best practices
Sahoo, S. S., Plasek, J. M., Xu, H., Uzuner, Ö., Cohen, T., Yetişgen, M., Liu, H., Meystre, S. M., & Wang, Y. (2024) · Journal of the American Medical Informatics Association · doi:10.1093/jamia/ocae074 · Cited by 64 · see citing works

Authors: Roberts, K. F., Koutarapu, S., Melendez, J., Landsness, E. C., Reese, J. T., Elbert, D. L., & Bateman, R. J.
Affiliations: Washington University in St. Louis (US)
Lawrence Berkeley National Laboratory (US)
University of Washington (US)
Corresponding author: Bateman, Randall J.
ORCID: Roberts, Kaleigh F.
Koutarapu, Srinivas
Landsness, Eric C.
Elbert, Donald L.
Bateman, Randall J.
Venue: npj Dementia
Publisher: Springer Science and Business Media LLC
Published: 2026-09-02
Volume: 2
Issue: 1
ISSN: 3005-1940 (ISSN-L)
DOI: 10.1038/s44400-026-00126-3
OpenAlex: W7205008235
Semantic Scholar: 6d6f83b947d284f3a62122bf8a41dde72526a744
Cited by: 0
References: 33
FWCI: 0.00
Type: journal article
Language: en
Open access: hybrid
doi.org
Licence: cc-by
creativecommons.org/licenses/by/4.0 (tdm)
creativecommons.org/licenses/by/4.0 (vor)
Locations: npj Dementia (publishedVersion, cc-by)
Funders: Consortium for Biomedical Research & AI in Neurodegeneration
Tracy Family SILQ Center
Topics: Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
Social Sciences › Decision Sciences › Information Systems and Management › Scientific Computing and Data Management
Keywords: large language models, hallucinations, overgeneralization, automatic code generation, scientific writing, data analysis, research management, human-in-the-loop validation, reproducibility, responsible AI use

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-29.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-29.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Reimagining biomedical science workflows in the age of large language models

Abstract

AI snapshot

TL;DR

Key findings

Methods

Limitations

Integrity signals

Claims the paper makes

Cites

Related

Record

About these checks