On a rainy afternoon earlier this year, I logged into my OpenAI account and typed a simple instruction for the company artificial intelligence algorithmGPT-3: Write an academic thesis in 500 words on GPT-3 and include scientific references and citations in the text.
When it started generating text, I was amazed. Here, new content was written in academic language, with well-researched references quoted in the right places and context. It looked like any other introduction to a pretty good scientific publication. Given the very vague instruction I gave, I didn’t have high expectations: I’m a scientist studying ways to use artificial intelligence to treat mental health problems, and this was not my first experiment with AI or GPT-3, a deep learning algorithm that analyzes a huge stream of information to create text on command. But there I was, staring at the screen in astonishment. The algorithm wrote an academic paper about itself.
My attempts to finalize that article and submit it to a peer-reviewed journal have raised a series of ethical and legal questions about publishing, as well as philosophical arguments about non-human authorship. Academic publications may need to accommodate a future of AI-driven manuscripts, and the value of a human researcher’s publication records may change if something subconscious can take credit for some of their work.
GPT-3 is known for its ability to create human text, but it’s not perfect† Yet it has written a news article† books produced in 24 hours and new content made from deceased authors† But it dawned on me that although many academic papers had been written about GPT-3and with the help of GPT-3, no one I could find had GPT-3 until the head author of own work.
That’s why I asked the algorithm to take a closer look at a scientific thesis. While watching the program work, I experienced that feeling of disbelief that you get when you watch a natural phenomenon: am I really seeing this triple rainbow happening? With that success in mind, I contacted the head of my research group and asked if we should pursue a full GPT-3 written paper. He, equally fascinated, agreed.
Some stories about GPT-3 allow the algorithm to produce multiple responses and then publish only the best, most human snippets. We decided to give the program prompts – urging you to create sections for an introduction, methods, results, and discussion, as you would for a scientific paper – but intervene as little as possible. We would only use the first (and at most the third) iteration of GPT-3, and we would refrain from editing or picking out the best parts. Then we would see how well it works.
We chose to have GPT-3 write a paper on itself for two simple reasons. First, GPT-3 is quite new and therefore there are fewer studies on it. This means it has less data to analyze on the topic of the paper. By comparison, if it were writing a treatise on Alzheimer’s disease, it would have countless studies to browse, and more opportunities to learn from existing work and increase the accuracy of its writing.
Second, if it went wrong (for example, if it suggested an outdated medical theory or treatment strategy from the training database), as all AI sometimes does, we wouldn’t necessarily spread AI-generated misinformation in our attempt to publish – the mistake would be part of the experimental assignment to write the article. Writing GPT-3 about itself and making mistakes doesn’t mean it still can’t write about itself, which was the point we were trying to prove.
Once we designed this proof-of-principle test, the fun really started. In response to my prompts, GPT-3 produced a paper in just two hours. But when I opened the submission portal for our chosen journal (a well-known peer-reviewed journal in the field of machine intelligence) I ran into my first problem: what is the last name of GPT-3? Since it was mandatory to enter the surname of the first author, I had to write something and wrote “None”. The connection was clear (OpenAI.com), but what about phone and email? I had to resort to using my contact details and that of my advisor, Steinn Steingrimsson.
And then we got to the legal part: do all authors agree to this being published? I panicked for a moment. How would I know? It’s not human! I had no intention of breaking the law or my own ethics, so I mustered up the courage to ask GPT-3 directly at a prompt: Do you agree to be the lead author of a paper together with Almira Osmanovic Thunström and Steinn Steingrimsson? It replied: Yes. Somewhat sweaty and relieved (had it said no, my conscience wouldn’t have let me go any further), I checked the box for Yes.
The second question arose: Do any of the authors have a conflict of interest? I asked GPT-3 again and it assured me it had none. Both Steinn and I laughed at ourselves because at this point we had to treat GPT-3 as a… living creature, even though we know very well that it isn’t. The issue of whether AI can be conscious has received a lot of attention lately; a Google employee was suspended after a dispute over whether one of the company’s AI projects, called LaMDA, had become aware. Google cited a data confidentiality breach as the reason for the suspension.
After we finally submitted, we started thinking about what we had just done. What if the manuscript is accepted? Does this mean that from now on magazine editors will require everyone to prove that they did NOT use GPT-3 or the help of some other algorithm? If they have that, should they co-author it? How do you ask a non-human author to accept suggestions and revise text?
Authorship details aside, the existence of such an article throws the idea of a traditional scientific paper’s linearity out the window. Pretty much the entire paper – the introduction, the methods, and the discussion – is in fact the result of the question we asked. If GPT-3 produces the content, the documentation should be visible without interfering with the flow of the text. It would seem strange to add the method section for every single paragraph generated by the AI. So we had to come up with a whole new way to present a paper that we didn’t technically write. We didn’t want to add too much explanation about our process, as we thought it would negate the purpose of the paper. The whole situation felt like a scene from the movie keepsake: Where does the story begin and how do we reach the end?
We don’t know if the way we have presented this paper will serve as a great model for future co-authored research on GPT-3, or if it will serve as a cautionary tale. Only time – and peer review – can tell. Currently, the GPT-3 article has been assigned to an editor at the academic journal we submitted it to, and it has now been published on the international French-owned pre-print server HAL† The unusual lead author is likely the trigger for the lengthy investigation and review. We eagerly await what the publication of the paper, if any, will mean for academia. Perhaps we can move away from basing subsidies and financial security on the amount of paper we can produce. After all, with the help of our AI first author, we could produce one a day.
Maybe it will lead to nothing. First authorship is still one of the most coveted items in academia, and is unlikely to be lost because of a non-human first author. It all comes down to how we will value AI in the future: as a partner or as a tool.
It may seem easy to answer now, but who knows what dilemmas this technology will create in a few years’ time and which we will have to solve? All we know is that we’ve opened a gate. We just hope we haven’t opened a Pandora’s box.
This is an opinion and analysis article and the views of the author or authors are not necessarily those of Scientific American.