Study shows: GPT-3 can—technically—be the author of scientific papers. An interview with Almira Thunström.
A system could technically be considered a co-author—or even the main author!—of a scientific paper! This is the mind-boggling insight presented by human AI researchers Almira Osmanovic Thunström and Steinn Steingrimsson from Gothenburg – with the assistance of their co-author GPT-3.
Article: Does GPT-3 qualify as a co-author of a scientific paper publishable in peer-review journals according to the ICMJE criteria? – A Case Study.
“We tested the system against the ICMJE criteria, and we found that a system could technically be considered a co-author—or even the main author—of a paper,” says Almira Osmanovic Thunström. “The system would have to be the one to come up with the topic, write paragraphs, analyze and critically evaluate its production and be responsible—correct and discover mistakes—in its paper.”
Almira and her human co-author Stein Steinn Steingrimsson—both from the University of Gothenburg—noted that the system had no problem to fulfill three out of the four main criteria for co-authorship: “drafting the work or revising it critically for important intellectual content; final approval of the version to be published; and agreement to be accountable for all aspects of the work.” Only when it came to referencing, the system could not fulfill the criteria.
Almira Osmanovic Thunström is a project manager and researcher in child and adolescent psychiatry. Currently, she is working mostly with chatbots and virtual reality. Her background, however, is somewhat mixed.
“I am a biologist on paper, an assistant nurse in my heart, and an engineer by trade,” she says. “I get to do quite a lot of fun and exciting things, but at the moment, I am mostly interested in developing digital tools for treating anxiety disorders.”
In the summer of 2022, Almira published an article in Scientific American that spurred some attention: We Asked GPT-3 to Write an Academic Paper about Itself—Then We Tried to Get It Published. After exploring automated writing for fun in 2021, she realized that with a little bit of luck and the right settings, she could ask the system to help her with academic work. Fascinated, she started to experiment, and in the process, it became important for her to document her thoughts. Her notes eventually became an op-ed for what is probably the world’s most famous popular science magazine.
“As someone who has been creating chatbots for the past two decades, I’ve been around for many of the twists and turns regarding natural language processing and large language models. I usually approach all these things with a childish sense of exploration and with each new model, software, or hardware I encounter, I want to do something silly with it.”
After realizing that paper-writing text robots are an interesting and philosophical dilemma, Almira noted that in a world where scientists write like robots, we are bound to create robots that write like humans.
“We wanted to raise a debate on transparency and authorship,” she says. “The question of authorship and accountability is a big pink elephant in the room we academics would rather avoid talking about. After all, that senior researcher on your paper that hasn’t even read it, is paying your bills. Any paper where your name is first will keep paying your bills in terms of grant money, which is often dependent on having a steady production rate of papers. With the help of a system like GPT-3, one could produce papers very fast; after all, if your text can be expanded and completed by a system, you can do everything in half the time.”
Hi, Almira. This is truly fascinating! Please tell us how you proceeded when you wrote the study.
“We asked the system to write about itself. Mainly because it has so little data on itself (Terminator plot headache warning: the system is trained on 175TB of text data from a variety of databases up until 2020/2021, however, very little if any data existed on GPT-3 because it didn’t exist at that time, only GPT-2 did). We wanted to test its capability. It was a case of childish curiosity and an interesting thought experiment. During the peer review of the submitted paper, we realized that the main issue reviewers had (remember this is pre-ChatGPT and ChatGPT publications!) was if the system could follow the ICMJE criteria for authorship on a paper. So we did what the reviewers asked of us: checked the system against the criteria.”
What was your reaction when you went through the results?
‘“The world should see this!’ was my first reaction. And ‘Oh no!’ was the second. How on Earth do I present this data? What will this mean for myself and others as we advance!?!”
What should scientists over the world learn from this study?
“Have an open mind, and dare to experiment. Let us write less like robots. Let us not blindly embrace authorship on papers; both humans and systems need to pass the ICMJE criteria. ChatGPT for example did not pass the ICMJE criteria on many different levels. Our paper has a very extensive discussion part that I think a lot of scientists will appreciate.”
How do you think the advent of ChatGPT and similar technology will affect science in the long run?
“I’m not so sure they will, or at least I think it will take time. My hope was that we’d rethink how we teach, write, design studies, and collaborate. I’m seeing institutions play out the worst scenarios in their head about the consequences of these systems, which I think is valid to some point, but not as productive as acknowledging a challenge and finding an innovative way forward. The most excited I’ve seen the universities get is when they discovered Crossplag.”
“In our SciAm article, I said we hope we hadn’t opened a Pandora’s Box. I think in the eyes of academic institutions, we very much did. Unlike Pandora’s box, hope was however not left inside. I see individual teachers and academics who are looking at how we can innovate how we teach in lieu of the takeover of automated writing—that gives me hope!”
Please note: all texts on this blog are produced by me, Olle Bergman, or invited human writers. When robot-written sections are added to serve as examples or to demonstrate a point, this is clearly indicated.