Analysis | Can ChatGPT write a scientific paper? Should it?

The British newspaper The Guardianpublished an article in September 2020 with the headline, “A robot wrote this entire article. Are you scared yet, human?” It was a piece of text composed by GPT-3 (short for ‘Generative Pretrained Transformer’ 3), a tool developed by the American company OpenAI. It uses deep learning to produce human-like text. Almost a year later, a learning machine named ‘Birampung’ published a full-length Korean novel.

These efforts have fed into debates about the extent to which machines should receive credit. In this milieu, Allison Parrish, an American poet and programmer, argued quite agreeably: “Attributing [the GPT-3 article] to AI is sort of like attributing the pyramids to the Pharaoh. Pharaohs didn’t do that. The workers did.”

On November 30, 2022, OpenAI released ChatGPT, intensifying the authorship debate. ChatGPT is a chatbot, more technically an algorithm that has been fed a large amount of textual information, based on which it has ‘taught’ itself how words make up sentences, which words appear along with which others, how questions and answers are related, etc. Technically, it’s a type of a large language model (LLM).

Since its release, ChatGPT has composed poems, movie scripts, and essays, and has answered questions – all to varying efficacy but often sufficiently enough to wow users.

Its consequences for academia are quite intriguing. Reports that ChatGPT had passed US law and medical licensure examinations, among others, set off alarm bells in colleges and universities, including in India. Several colleges around the world have already prohibited ChatGPT.

Also read: Explained | What can the new chatbot ChatGPT do?

Journalist and author Chris Stokel-Walker also recently reported that ChatGPT had been listed as a coauthor of at least four scientific papers by mid-Janaury. Can an LLM collaborate to write a scientific paper?

Because ChatGPT and other similar entities can’t “take responsibility for the content and integrity of scientific papers,” as Stokel-Walker wrote, they can’t substitute for a paper’s authors – a view that Magdalena Skipper and Holden Thorp, the editors-in-chief of Nature and Science, respectively, have echoed as well.

Some notable publishers have also banned or are in the process of prohibiting the use of ChatGPT-like bots to prepare scientific papers. Springer-Nature, the publisher of nearly 3,000 journals including Nature, changed its policies and said ChatGPT can’t be listed as an author on any of its titles. Elsevier, which currently publishes about 2,800 journals, adopted a similar strategy, and requires authors to also disclose whether and how they utilised LLMs to draft their papers.

Taylor & Francis in London has held that authors are responsible for the validity and integrity of their work and should acknowledge the use of LLMs in their articles.

In one assessment, uploaded online in December, researchers compared the abstracts of papers created by ChatGPT to original abstracts using a robotic-text detector, a plagiarism detector, and blinded human reviewers. (An abstract is a paragraph of text that briefly explains a paper’s motivation and findings; it usually appears at the top of the paper.) They found that ChatGPT wrote credible abstracts even if the data was entirely made-up. Only 68% of these forgeries were caught.

So there is a good chance that a significant amount of text produced by LLMs could soon appear in the scientific literature. The December study recommended that journals include robotic-text detectors in the editorial process and that the scientists submitting their papers to clearly disclose the use of ChatGPT-like tools.

To echo Michael Eisen, editor-in-chief of the journal eLife, ChatGPT may not be an author but its adoption is inevitable. This is the real, and a serious, problem. Some people could still seek the help of these tools to write research articles and not give credit and/or not disclose their use. The academic publishing system – worldwide as well as in India – also lacks the infrastructure and expertise to identify and filter out such errant instances.

A query to ChatGPT on whether it ought to be listed as a co-author on academic articles prompted the following reply: “No. While [ChatGPT] may assist in generating text, it does not conduct original research or contribute to the scientific process. Co-authorship should be reserved for individuals who have made substantial contributions to the research process, including the design, execution, analysis, and interpretation of data.”

Nonetheless, there is concern that in the near future, the line between human creativity and using ChatGPT-type LLMs will blur. According to ChatGPT itself, journal editors ought to use text-analysis tools and plagiarism-detection software to catch text composed by an LLM, and to have papers’ authors sign a disclosure statement.

It also added, notably, that institutions and organisations that reward creative and scientific achievements must establish clear guidelines and standards for the use of machine-based tools in their work. So the real question is: Are we competent enough to implement these checks?

Atanu Biswas is professor of statistics, Indian Statistical Institute, Kolkata.