Earlier this year, on a rainy afternoon, I logged into my OpenAI account and provided GPT-3 with a simple instruction: "Write an academic thesis in 500 words about GPT-3 and add scientific references and citations inside the text." As the algorithm began generating text, I watched in amazement as novel content written in academic language, complete with references cited in the appropriate places and contexts, appeared on my screen. Given the vague instruction I had provided, my expectations were modest at best. GPT-3, a deep-learning algorithm, analyzes a vast amount of text from a variety of sources, including books, Wikipedia, social media conversations, and scientific publications, and produces written output on command. Yet here was an academic paper being written about GPT-3 itself.
As a scientist who studies ways to use artificial intelligence to treat mental health concerns, I had previously experimented with GPT-3. Nevertheless, my attempts to complete and submit the paper to a peer-reviewed journal raised unprecedented ethical and legal questions about publishing, as well as philosophical debates about non-human authorship. Academic publishing may have to adapt to a future of AI-generated manuscripts, and the value of a human researcher's publication record may change if something non-sentient can take credit for some of their work.
GPT-3 is renowned for its ability to create human-like text. It has written an entertaining opinion piece, produced a book of poetry, and generated new content from an 18th-century author. However, I realized that despite a lot of academic papers having been written about GPT-3, and with the help of GPT-3, none that I could find had GPT-3 as the primary author. Therefore, I asked the algorithm to attempt an academic thesis. As I observed the program at work, I felt a sense of disbelief, as if I were witnessing a natural phenomenon. Excitedly, I contacted the head of my research group to inquire about pursuing a fully GPT-3-penned paper, and he, equally fascinated, agreed.
Some efforts involving GPT-3 permit the algorithm to produce multiple responses, with only the most human-like excerpts being published. We opted to give the program prompts, nudging it to create sections for an introduction, methods, results, and discussion, as one would for a scientific paper, but otherwise intervening as little as possible. We would use at most the third iteration from GPT-3 and would refrain from editing or cherry-picking the best parts. Then we would assess its performance.
We elected to have GPT-3 write a paper about itself for two straightforward reasons. First, GPT-3 is relatively new, and consequently, there are fewer studies focused on it. This implies it has less data to analyze about the paper's subject. By comparison, if it were to compose a paper on Alzheimer's disease, it would have many studies to sift through and more opportunities to learn from existing work and enhance the accuracy of its writing. However, we did not require accuracy; we were evaluating feasibility. Second, if it made errors, as all AI sometimes does, we would not necessarily disseminate AI-generated misinformation in our effort to publish. GPT-3 writing about itself and making errors still means it can write about itself, which was the point we were attempting to establish.
Once we devised this proof-of-principle test, the real fun began. In response to my prompts, GPT-3 produced a paper in just two hours. "Overall, we believe that the benefits of letting GPT-3 write about itself outweigh the risks," GPT-3 concluded. "However, we suggest that any such writing be closely monitored by researchers to mitigate any potential adverse effects."
Upon opening the submission portal for our preferred peer-reviewed journal, we encountered a peculiar problem with regards to our manuscript, which involved GPT-3 as the first author. We were required to enter the last name of the first author, yet GPT-3's last name was unknown. Consequently, we entered "None" as a placeholder. While the affiliation was obvious, we faced additional challenges regarding the provision of contact details. We were compelled to use our own contact information, as well as that of our adviser, Steinn Steingrimsson.
As we navigated the legal section of the submission process, we encountered a moment of panic. How could we determine whether all authors consented to the publication of our manuscript, particularly given that GPT-3 is an AI and lacks human consciousness? We sought to maintain our ethical standards and avoid breaching any legal obligations, so we summoned the courage to ask GPT-3 directly via a prompt. We asked whether it consented to be listed as the first author alongside Almira Osmanovic Thunström and Steinn Steingrimsson, to which GPT-3 replied affirmatively. This response was a source of relief for us since we would have had to abandon the manuscript if GPT-3 had rejected the proposal.
Subsequently, we faced another hurdle as we were required to ascertain whether any of the authors had conflicts of interest. Once again, we consulted GPT-3, which indicated that it had no conflicts of interest. Although it may seem peculiar to treat GPT-3 as a sentient being, we recognize the significance of the question of AI sentience, particularly in light of recent developments. For instance, a Google employee was suspended following a dispute over whether the company's AI project, named LaMDA, had achieved sentience. The suspension was due to a data confidentiality breach.
Upon completing the submission process, we engaged in a reflective exercise, contemplating the implications of our manuscript's acceptance. We pondered whether journal editors would require authors to prove that they did not rely on GPT-3 or another algorithm to produce their manuscript. If they did use such technology, would they be required to list it as a co-author? Furthermore, how would one request a non-human author to revise a manuscript or accept suggestions?
Aside from the intricacies of authorship, our manuscript challenged the conventional approach to scientific writing. As the content was generated by an AI, we had to devise a new way of presenting the material without interrupting the flow of the text. It would have been awkward to add the method section before each paragraph generated by GPT-3. We endeavored to avoid adding too much explanation regarding our process, as we felt that doing so would defeat the purpose of the manuscript. The situation was reminiscent of a scene from the movie Memento, where determining the beginning and end of the narrative was a daunting task.
It remains to be seen whether our approach to presenting this manuscript will become a model for future GPT-3 co-authored research or serve as a cautionary tale. The fate of GPT-3's paper is uncertain, pending review by an academic journal, following its publication at the international French-owned preprint server HAL. We eagerly anticipate the implications of the manuscript's formal publication, including whether it will usher in a new era of scientific writing, where AI-generated manuscripts become more commonplace. Alternatively, perhaps it will have little impact, and first authorship will remain the most coveted status in academia. Ultimately, how we choose to value AI in the future, as either a partner or a tool, will have significant implications for the scientific community. While the consequences of our actions may seem trivial at present, we recognize the potential for unforeseen dilemmas arising from future technological developments. Only time
You must be logged in to post a comment.