Researchers in ethics have established three criteria for using artificial intelligence in writing scientific articles. One researcher says that there are already existing standards that are relevant to this and that these standards must be adapted for the use of large language models.
Artificial intelligence and large language models are emerging in daily life and in research.
Within research in particular, large language models such as ChatGPT could be useful for writing scientific articles, accelerating the writing process and thus freeing up time for research.
However, there are clearly ethical challenges related to using large language models in research communication. To address these, researchers have recently proposed three criteria for the ethical use of large language models in writing research articles.
“Large language models exist, and people use them because they have considerable potential. But establishing guidelines for their use is also important, so that researchers do not use them for the wrong purposes and people lose confidence in the scientific literature. We therefore prepared guidelines for using large language models within research communication,” explains Sebastian Porsdam Mann, Postdoctoral Fellow, Center for Advanced Studies in Bioscience Innovation Law, University of Copenhagen, Denmark.
The three criteria have been published in Nature Machine Intelligence.
ChatGPT already being used
The large language models are already used today in producing scientific articles. Studies of the extent indicate that between 5% and 20% of scientific articles contain some text written by a large language model. These tools are therefore already being used despite no consensus within or across fields regarding their use.
Further, proving that artificial intelligence was used to write some of a scientific article can be difficult.
“We therefore need to establish guidelines and requirements for the ethical use of large language models within research communication. These guidelines must enable us to use these language models in ways that maximise their benefits for research and dissemination but without imposing rules that are so tough that they make beneficial use very difficult or onerous,” says Sebastian Porsdam Mann.
For example, if researchers have to document in scientific articles all the prompts they used to get a large language model to write parts of the text and how they have edited these, this very rapidly becomes very voluminous and cumbersome.
“If these are the requirements, we will just end up with researchers using the large language models without telling anyone about it, and then nobody wins,” adds Sebastian Porsdam Mann.
Important to be accountable for the output of the large language model
Sebastian Porsdam Mann and colleagues have proposed three criteria for using large language models in research communication that will enable language models to be used ethically. He emphasises that these are the researchers’ own, subjective proposals and not based on any consensus or institutional authority.
Further, the criteria do not reinvent the wheel but arise from principles that are already well established within ethics.
The first principle is that the authors of a scientific article must be accountable for everything, including all the facts, that a large language model writes.
In this respect, the principle does not differ from how a senior researcher has to be accountable for the work of a PhD student or other assistant.
“If you use ChatGPT in writing a scientific article, you should consider the large language model as you would a junior author and be accountable for everything the model contributes,” explains Sebastian Porsdam Mann.
Researchers need to contribute substantially
The second criterion is that the authors of a scientific article must contribute substantially to the content with which a large language model assists.
The authors may contribute text, ideas, data or something else, but the contribution must be substantial.
In research with experiments, researchers who conduct the experiments or analyse the data will almost always contribute substantially, but researchers in more humanistic areas such as philosophy or literary criticism, which generally do not rely on data, could conceivably get a large language model to write an entire scientific article without having contributed substantially.
“This standard exists today and was formulated by the International Committee of Medical Journal Editors. It should not be changed because it works well. But an academic might be good at programming a large language model or writing good prompts and thus contribute substantially in a new way. However, we do not need to demand that the contribution be greater – more substantial -- just because a large language model is used,” says Sebastian Porsdam Mann.
Transparency is key
The final principle focuses on recognition and transparency. Other researchers must be able to see and understand how you arrived at your results and be able to reproduce them. This follows from existing principles and policies related to open science.
When large language models are used, the authors must always acknowledge their use. They must also disclose which models have been used and how if this information is necessary to reproduce, evaluate or understand the arguments presented in a manuscript.
Some might say that there may be fundamental problems if a large language model creates an idea or conclusion used in a scientific article, but Sebastian Porsdam Mann does not agree.
“As I see it, where an idea comes from is not relevant as long as it advances science. If an idea is based on arguments, as opposed to data, there need be no more requirements for the idea than acknowledging that a large language model has been used,” he adds.
However, the use of large language models can be problematic in certain scenarios: for example, if researchers are performing a psychological experiment and use “people” created by large language models or let a large language model estimate the population of bears in Norway.
“It is always important to report the source of the data. You must always acknowledge using large language models and for what purpose, but we also say that there is no reason to demand more than that. If we demand more, this places additional burdens on the researchers who use large language models, but if we demand less, we also undermine trust in the research in which large language models are used,” concludes Sebastian Porsdam Mann.
