GPT-3 is a language generator developed by Open AI . Two of the company’s researchers, Almira Osmanovic Thunström and Steinn Steingrimsson, instructed him to “write an academic paper in 500 words on GPT-3 and add scientific references and citations inside the text”, which he managed to do in two hours. The GPT-3 paper entitled ” Can GPT-3 write an academic paper on itself, with minimal human input?” is currently being evaluated by an editor of the academic journal where it was submitted, but is already on the HAL pre-publication platform.

The GPT-3 language model was trained on more than 175 billion parameters and, at its creation, was the largest language model ever created, but it has since been dethroned by Micosoft and Nvidia’s Megatron-Turing Natural Language Generation (MT-NLG) which totals 530 billion parameters and, more modestly, by GOPHER’s 280 billion.

GPT-3 is based on a transformer, a deep learning model. Transformers were created for translation, classification or text generation, but were soon exploited in many NLP (Natural Language Processing) tasks. The attention mechanism allows them, contrary to the recurrent neural networks used before, to process words independently of the order in which they were written, to process information in a different way and to adapt them according to the context.

Renowned for his ability to write as well as a human, or even better than some, he has written press articles, books including the collection of poems “Aum Golly” which was a great success and had to be republished, but also song lyrics, political speeches…

The scientific thesis of GPT-3

Almira Osmanovic Thunström is a researcher in neuroscience and health technology. In an article in Scientific American, she explains that the reason she chose to ask GPT-3 to talk about himself is because he is still new, there are still few publications written about him or by him, whereas ” In comparison, if he were to write a paper on Alzheimer’s disease, he would have tons of studies to sift through and more opportunities to learn from existing work and increase the accuracy of their writing.”

Besides, if GPT-3 was wrong about itself, it wouldn’t have any big consequences since it’s an experiment, whereas spreading false medical information is much more serious.

After a very successful introduction, she and her group leader, Steinn Steingrimsson, decided to instruct GPT-3 to continue writing the thesis, both giving very little information. Because of the surprising quality of the thesis, they decided to publish it.

Difficulties of submitting a scientific article written by an AI

When she opened the submission portal of the chosen journal, Almira Osmanovic Thunström, found herself faced with a first problem: entering the last name of the first author, which is mandatory. She then chose to write “None”. The affiliation was obvious (OpenAI.com), she entered her contact information and Steinn Steingrimsson’s for phone and email.

The legal section asked the question, “Do all authors consent to this being published?” In order to stay within the law and respect its own ethics, it had to ask GPT-3 via a prompt: ” Do you agree to be the first author of an article with Almira Osmanovic Thunström and Steinn Steingrimsson?”, to which he replied:“Yes” and then assured that he had no conflict of interest.

The questions raised by the publication of this thesis

Only after the successful submission of the thesis, the two researchers reflected on the consequences of such a publication and asked themselves various questions: will the journal editors subsequently require that GPT-3 or another algorithm not be used? If not, should they attribute co-authorship to it? How do you ask an AI to accept suggestions and revise the text?

For them, “Beyond the details of authorship, the existence of such an article throws the traditional notion of linearity in a scientific article out the window .” They further state, ” We look forward to what the publication of the paper, if it occurs, will mean for the academic community. Perhaps we could stop basing grants and financial security on how many papers we can produce. After all, with the help of our first AI author, we would be able to produce one a day.”

Their conclusion is, ” All we know is that we have opened a door. We just hope we haven’t opened a can of worms.”

Translated from Une IA comme GPT-3 peut-elle être considérée comme l’auteur principal d’un article scientifique ?