The following uses of LLMs during manuscript preparation require extra caution and disclosure.
1. Drafting manuscript content.
Because LLMs are susceptible to hallucination and unreliable for attribution of ideas and phrases, their use in the generation of text carries significant risks of plagiarism and scientific misrepresentation. It is the authors’ responsibility to ensure that all text included in their manuscript is original and scientifically sound.
Examples of acceptable use of AI in drafting manuscript text:
- An author writes an outline of the scientific argument they intend to make and uses ChatGPT to generate text connecting these arguments together in complete sentences. The author (and coauthors, as applicable) read the generated text and implement any necessary edits to ensure it is scientifically correct and that it conveys the authors’ intended meaning. This use is described in the authors’ AI disclosure statement included at the end of the article.
- An author wishes to include a summary of a specific published article in their manuscript. They prompt ChatGPT to summarize the article in the desired number of words. They then check that the summary is factually correct and that it faithfully conveys the findings of the article. The author and coauthors implement any necessary edits to ensure it conveys the intended meaning. This use is described in the authors’ AI disclosure statement.
Example of unacceptable use of AI in drafting manuscript text:
- An author prompts ChatGPT to write the introduction section of an article on a particular topic. The author copies and pastes the generated text without reading it. This use is unacceptable whether or not the author discloses their LLM use.
2. Using AI to translate manuscript text written in a language other than English for the purposes of publishing.
Example of acceptable use of AI-assisted translation:
- An author writes a paragraph in Spanish and uses ChatGPT to translate it to English. That author or a co-author reads the translated paragraph and edits it, as appropriate, to ensure that it conveys the authors’ intended meaning. At the end of the article, the authors disclose their use of ChatGPT for translation.
Example of unacceptable use of AI-assisted translation:
- An author writes a paragraph in German and uses ChatGPT to translate it to English. The author copies and pastes the translation into the manuscript without reading it or asking anyone else to read it.
3. Generating, refining, correcting, editing, or formatting images, diagrams, or other figures for illustrative purposes only.
It is not possible to trace source material when images are generated by LLMs. Authors are responsible for doing due diligence to ensure images are not copies of others’ work. The use of tools to create illustrations must be disclosed and should include the specific tool used, and even potentially the prompt, in the image caption.
Examples of acceptable use:
- Creating cartoons or diagrams for illustrative or explanatory purposes to explain complex phenomena or methods.
- Creating illustrative images of physical systems to provide context or background to scientific results.
Examples of unacceptable use:
- Generating scientific plots from unknown data provided by the LLM.
- Presenting generated images as research outputs in themselves or as representing research outputs.
4. Generating, refining, correcting, editing, or formatting visualizations of research data or results.
GenAI tools and LLMs may be used to generate visualizations of research data or results, but care should be taken to ensure the accuracy of the result and that the visualization faithfully represents the data. Authors are responsible for ensuring the accuracy and integrity of what is submitted and published.
Examples of acceptable use:
- Using AI tools to generate graphs, tables, or other visualizations of research datasets generated and provided by the author(s). The authors double-check the output to ensure faithfulness to the data, for example by checking the axis limits and checking that no data have been artificially added/removed by the tool. The figure caption should include the specific tool used, and, if possible, the prompt.
- An author creates a plot from a dataset and uses AI tools to make it more accessible (adjusting axes, font sizes, colors, etc.).
Examples of unacceptable use:
- Using AI tools to generate scientific visualizations with no oversight. For example, using an LLM to generate a visualization of the authors’ data, without double checking for accuracy such that the presented results represent a skewed version of the data.
- Using AI tools to generate scientific visualizations without providing source data. For example, the authors ask the tool to create a graph of two data products but do not provide the data.
- Using AI tools to manipulate data values to change the results.
5. Assisting with gathering references.
LLMs have proven to be powerful tools for identifying previously published articles in a given research area, and they are being used by many researchers during the information-gathering phases of their projects. At the same time, numerous experiments have shown that the publication lists generated in this way are often incomplete, and in the worst cases can even include entirely fictitious articles (hallucinated publications). The lists can also be more heavily weighted toward frequently appearing or cited articles, which can introduce a degree of groupthink into the compilations, causing important work by less visible researchers to be overlooked.
With these factors in mind, LLMs can provide a useful tool for supplementing references gathered using more traditional tools, but authors should refrain from relying solely on LLMs for generating the list of articles cited in their manuscripts. Authors should also apply commonsense checks to the LLM-generated references to ensure the legitimacy of the articles and their relevance to the research being cited, for example, by downloading the article to confirm its relevance (good practice regardless of where the reference came from), and even to confirm that the cited article actually exists. Publishing fictitious references is regarded as a serious professional breach, and it can result in the article being retracted by the publisher, with longer-term consequences for the authors. Authors should also bear in mind that they are responsible for citing and crediting previous work that relates directly to their own research, regardless of whether said articles were recovered by an LLM search. The application of an LLM does not reduce the need to also employ more traditional tools, such as the Astrophysics Data System (ADS), to ensure a complete and unbiased check of the literature.
