AI in Scientific Writing: Uses, Limitations, and Ethical Implications

/in , /by

Introduction

Artificial Intelligence has long been widely used in biological research, with huge contributions in structural biology, image analysis, and genomics. Generative AI, however, has only recently become a major player in scientific writing and communication. In this blog, we discuss the different uses of AI in writing, how they can be useful, where they present problems, and ethical considerations.

Generative AI can be used in a variety of ways, it can be used to write text, code, draft emails, analyse and debug code, as well as to create images and figures. We come across generative AI frequently on a day-to-day basis. Large language models (LLMs) are a type of generative AI, which is trained on language and its main output is text. This includes ChatGPT, Gemini, DeepSeek, and LLaMA. To generate images and figures, image-based models, which are trained on images are used. Examples of these include DALL.E, Imagen and Adobe Firefly.

In this blog, we discuss the use of generative AI in scientific writing and publishing: where we stand, what publishers say, and the implications.

 

Plant Science-specific AIs

There are several examples of popular LLMs being asked about plant science, and often getting things wrong (Farmer et al., 2025; Agathokleous et al., 2024; Salt, 2023). One recent example is this preprint by Burda et al. which asks LLMs including ChatGPT and Gemini questions to assess their plant knowledge, and created this set of questions for future benchmarking. While the Ais scored 75% on multiple choice questions (significantly better than random chance), the authors found important information was frequently omitted, and poor acknowledgment of self-limitations.

PlantScience.ai is an LLM lead by the University of Exeter. It continually learns from research, can follow instructions from the user, and has “citation awareness”. This preprint  from Yu et al. explains how it can continually gather information, and therefore regularly benchmarking may be appropriate. Outputs can be traced back to original sources, allowing the user to verify the validity of information, an important step in the prevention of hallucinations and errors. Another plant-specific LLM is PLLaMa, which evolved from Meta AI’s LLaMa-2 (Yang et al., 2024). The most significant difference is that PLLaMa’s 1.5 million scholarly articles are not continuously updated. While obtaining new information may give PlantScience.ai longevity and time to improve, it also increases the risk of copyright infringements and inclusion of low-quality or incorrect data. There is also a rice-specific and an Arabidopsis-specific LLM (Yang et al., 2025; Zhang et al., 2025), which raises the question of how niche will (and should) LLMs become?

 

AI in images and figures

Image-based generative AI tools include DALL.E, stability.ai, Canva AI tools, Imagen and Adobe Firefly. They are often trained on images coupled with descriptive text to respond to a user’s instructions. Figures can be convincingly generated using generative AI, leading to misleading or fraudulent data which escape the notice of the average reader. This is risky in scientific research and leads to incorrect conclusions drawn from the article. Articles which have these AI generated, misleading images frequently slip through peer-review. A famous example of a (since retracted) article used generative AI to summarize findings in diagrams (Guo et al., 2024). The diagrams featured labels with nonsensical text as well as presenting inaccurate scientific information. It is important to stay vigilant for AI generated images. Checking whether the labels are clear, that the language is consistent, as well as keeping an eye on online forums such as PubPeer, can help avoid the propagation of misinformation.

 

AI in Peer Review

The use of LLMs in the peer review process is growing rapidly. This happens frequently, and often quietly. A recent Frontiers survey presents the startling statistic that more than 50% of surveyed researchers (amounting to 1,600) have used AI tools when peer reviewing manuscripts (Frontiers, 2025). In many ways, LLMs could be considered useful tools to note inconsistencies in the text or the data, to flag unclear logic, as well as to summarise the manuscript. However, this is often in direct violation of the publisher guidelines. Many publishers, including Frontiers and Nature, forbid the upload of an unpublished manuscript to chatbot websites, such as ChatGPT, for confidentiality reasons. In response, Frontiers has launched its own, in-house AI platform to allow for responsible use of AI in peer-review with clear guidelines and appropriate training.

Generative AI is often limited in its actual usefulness when it comes to peer-review. A recent Nature article summarises a YouTube video by Mim Rahimi, who tested whether GPT-5 could generate a review report comparable to real peer-review reports based on his own article (Naddaf, 2025). He found that GPT-5 could mimic the structure of the report but, critically, failed to provide constructive feedback.

The productivity and accuracy of LLMs is developing rapidly, however, there are still flaws in the information they provide, Therefore, the use of generative AI in the peer-review of manuscripts should be limited to drafting and clarifying a reviewer’s comments, including rewording for clarity.

 

Ethical considerations

AI massively impacts the environment, and not in a positive way. The UN Environment Programme publication (2025) outlines some major damage being done, and how to ‘rein in the fallout’. Tomlinson et al. (2024) reported that training models such as GPT-3 already consumes around 1,287 megawatt-hours of electricity, equivalent to powering 120 U.S households for a year, and generates hundreds of tons of carbon dioxide. According to The Guardian, recent estimates suggest that AI systems could be responsible for 32.6 – 79.7 million tons of carbon dioxide emissions yearly, comparable to New York City’s total annual output. Meanwhile, water consumption for data centre cooling may reach 765 billion litres, exceeding the global demand for bottled water (Booth, 2025). These impacts, often powered by fossil fuels and concentrated in water-scarce areas, highlight the need for mitigation through efficient algorithms, renewable energy integration, strategic siting, and transparent reporting. In plant science, where AI aids research, balancing benefits against planetary costs is essential for truly sustainable progress.

 

There are also concerns around copyright and generative AI. First, is AI stealing and ‘taking credit’ for the work of others? In short, yes. This is particularly evident in art and image generation, as described by the Guardian (Milmo, 2025). Even text has been taken from the New York Times and output by ChatGPT, resulting in a legal action. As well as regurgitating copyrighted work, AIs frequently use these works in their training datasets, relying on human-generated data. The Washington Monthly summarises this nicely in the article ‘AI Needs Us More Than We Need It’, by saying “Without a constant stream of high-quality, human-made information, artificial intelligence models become useless.” (Radsch, 2024).

A major consideration is how the use of AI will impact future research. LLMs may present incorrect or low-quality information to scientists while they are developing their hypotheses, leading to improperly designed experiments. Possibly the most important consideration is that reliance on LLMs will lead to a generation of scientists lacking critical writing and reading comprehension skills, a concern backed by neuroscience researchers (Kosmyna et al., 2025). Furthermore, the belief that AI can somehow replace human skills is being used to justify the elimination of scientific jobs, as highlighted in a recent Plant Science Research Weekly.

 

Statements from Publishers

Most journals and/or their publishers have provided guidance on the use of generative AI and LLMs in writing for their authors. These statements from several of the most popular plant science journals/publishers are listed here:

Publisher/Journal Author Guidelines
Oxford Academic https://academic.oup.com/jpepsy/pages/ai-policy
Wiley https://authorservices.wiley.com/ethics-guidelines/index.html#22
Nature Plants Artificial Intelligence (AI) | Nature Plants
PLOS Ethical Publishing Practice | PLOS One
Cell Journal policies: Structure

 

In summary, AI can be used in the creation of scientific articles, but cannot be credited as an author. The use of AI is generally required to be disclosed in the methods section of the paper.

Many people use AI tools to improve the grammar, language structure, and tone of their writing. These tools include Grammarly, Microsoft Copilot, and ChatGPT. These can be particularly helpful for non-native speakers, and therefore can make scientific writing more accessible. Journals and publishers therefore typically don’t require authors to declare the use of AI for this reason.

 

Conclusions

Generative AI is a powerful influence in scientific writing, with both positives and negatives.

Some key considerations in effectively using AI for scientific writing:

  • Don’t rely on generative AI, build your own reading and writing skills.
  • When using generative AI, be precise and detailed when writing prompts.
  • Understand and acknowledge limitations like hallucinations and information omissions.
  • Always use primary sources of information, and preferably use AI tools that cite them.
  • Follow the author guidelines when disclosing the use of AI.
  • Do not use AI for hypotheses, interpretations, results, or conclusions in your work.

 

References

  • Agathokleous E, Rillig MC, Peñuelas J, Yu Z (2024) ‘One hundred important questions facing plant science derived using a large language model’ Trends in Plant Science https://doi.org/10.1016/j.tplants.2023.06.008
  • Booth R (2025) ‘AI boom has caused same CO2 emissions in 2025 as New York City, report claims’ The Guardian https://www.theguardian.com/technology/2025/dec/18/2025-ai-boom-huge-co2-emissions-use-water-research-finds
  • Farmer EE, Brown D, Gore MA, Tufan HA (2025) ‘Applying large language models to extract information from crop trait prioritization studies’ Plants People Planet https://doi.org/10.1002/ppp3.70075
  • Fernandez Burda M, Ferrero L,  Gaggion N,  Fonouni-Farde C, The MoBiPlant Consortium,  Crespi M,  Ariel F,  Ferrante E (2025) ‘What Large Language Models Know About Plant Molecular Biology’ BioRxiv https://doi.org/10.1101/2025.08.31.672925
  • Frontiers (2025) ‘Unlocking AI’s untapped potential – responsible innovation in research and publishing’ Frontiers https://www.frontiersin.org/documents/unlocking-ai-potential.pdf
  • Kosmyna N, Hauptmann E, Yuan YT, Situ J, Liao XH, Beresnitzky AV, Braunstein I, Maes P (2025) ‘Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task’ arXiv https://doi.org/10.48550/arXiv.2506.08872
  • Milmo D (2025) ‘The platform exposing exactly how much copyrighted art is used by AI tools’ The Guardian https://www.theguardian.com/technology/2025/oct/18/the-platform-exposing-exactly-how-much-copyrighted-art-is-used-by-ai-tools
  • Naddaf M (2025) ‘More than half of researchers now use AI for peer review — often against guidance’ Nature https://doi.org/10.1038/d41586-025-04066-5
  • Radsch CC (2024) ‘AI Needs Us More Than We Need It’ The Washington Post https://washingtonmonthly.com/2024/10/29/ai-needs-us-more-than-we-need-it/
  • Salt A and A.I. Assistant (2023) ‘ChatGPT may know a lot, but when it comes to botany basics, this AI still has a lot to learn according to researchers who gave it a “plant awareness” quiz.’ Botany One https://botany.one/2023/09/much-plant-life-is-invisible-to-ai/
  • Tomlinson B, Black RW, Patterson DJ, Torrance AW (2024) ‘The carbon emissions of writing and illustrating are lower for AI than for humans’ Scientific Reports https://doi.org/10.1038/s41598-024-54271-x
  • UN Environment Programme (2025) ‘AI has an environmental problem. Here’s what the world can do about that.’ https://www.unep.org/news-and-stories/story/ai-has-environmental-problem-heres-what-world-can-do-about
  • Williams M (2025) ‘Perspective: How should the advancement of large language models affect the practice of science?’ Plant Science Research Weekly https://plantae.org/perspective-how-should-the-advancement-of-large-language-models-affect-the-practice-of-science/
  • Xinyu Guo X, Liang Dong L, Dingjun Hao D (2024) ‘RETRACTED: Cellular functions of spermatogonial stem cells in relation to JAK/STAT signaling pathway’ Frontiers in Cell and Developmental Biology https://doi.org/10.3389/fcell.2023.1339390
  • Yang F, Kong H, Ying J, Chen Z, Luo T, Jiang W, Yuan Z, Wang Z, Ma Z, Wang S, Ma W, Wang X, Li X, Hu Z, Ma X, Liu M, Wang M, Chen F, Dong N (2025) ‘SeedLLM·Rice: A large language model integrated with rice biological knowledge graph’ Molecular Plant https://doi.org/10.1016/j.molp.2025.05.013
  • Yang X, Gao J, Xue W, Alexandersson E (2024) ‘PLLaMa: An Open-source Large Language Model for Plant Science’ arXiv https://doi.org/10.48550/arXiv.2401.01600
  • Yu H, Zhou S, Huang M, Ding L, Chen Y, Wang Y, Ren Y, Cheng N, Wang X, Liang J, The John Innes Centre and The Sainsbury Laboratory Collaboration, Zhang H, Ding Y, Li K (2025) ‘PlantScience.ai: An LLM-Powered Virtual Scientist for Plant Science’ BioRxiv https://doi.org/10.1101/2025.10.24.684337
  • Zhang R, Wang Y, Yang W, Wen J, Liu W, Zhi S, Li G, Chai N, Huang J, Xie Y, Xie X, Chen L, Gu M, Liu YG, Zhu Q (2025) ‘PlantGPT: An Arabidopsis-Based Intelligent Agent that Answers Questions about Plant Functional Genomics’ Advanced Science https://doi.org/10.1002/advs.202503926

 

______________________________________________

About the Authors

Ciara O’Brien

Ciara is a postdoctoral researcher in the John Innes Centre (UK) and a 2025 Plantae Fellows.  SOriginally from Ireland, Ciara moved to the UK to study postharvest ripening at Cranfield University for their PhD, where they fell in love with science communication- from teaching to outreach!  You can find her on Bluesky: @ciara-obrien.bsky.social.

Kestrel Maio

Kes is a second-year PhD candidate at the John Innes Centre and a 2025 Plantae Fellows. She is studying the molecular mechanisms underlying how flowers develop their complex shapes, using the model system Arabidopsis thaliana. Her research integrates computational biology with molecular genetics to uncover conserved laws underlying morphogenesis. You can find her on Bluesky: @kesmaio.bsky.social.

Fatai Ayomide Akande

Fatai is a research assistant at the International Institute of Tropical Agriculture Headquarters in Nigeria and a 2026 Plantae Fellow, specializing in advancing crop resilience to biotic and abiotic stresses through molecular breeding, genomics, and bioinformatics. His research focuses on understanding and improving Striga resistance and drought adaptation in maize using gene expression studies, marker discovery, and validation. You can find him on X: @AkandeAyomide20.