Ruth Hagengruber Seminar Winter 22_23 Mensch_Maschine_Musse:
A commentary on ChatGPT: A revolutionary tool for writing or the death of scientific papers? by Jan Hendrik Hellmiß
With 2022 coming to an end, OpenAI, the San Francisco A.I. company responsible for releasing different A.I.-based language processing models, revealed its next artificial intelligence system, ChatGPT. At its core, ChatGPT is not different from any other Chatbot released in the last decade, except that it is, put simply, “the best artificial intelligence chatbot ever released to the general public” (Rosse 2022: A1). Additionally, it has generated over 1 Million users within five days, turning it into one of the most successful single application launches in human history, toppling even those of Instagram and Spotify, which required months to generate these user numbers (ibid.). Shortly after its release, numerous statements were made regarding the future of creative writing in academic and literary fields. One side perceives it as the end of academic papers as it is known to modern society, while others consider ChatGPT as only another chatbot, “genius in one moment and absolutely dumb in the next” (Kühl 2022). However, recent discussions about the new chatbot have failed to emphasize the possible implications that it could hold regarding how humans perceive creativity and creative thinking machines. Therefore, this paper will address these implications by exploring what ChatGPT does and discuss its ability to produce meaning in light of current research. Firstly, it should be addressed what ChatGPT is and how it operates. As mentioned above, first and foremost, it is an AI-based chatbot solely operating in text-generating form. Therefore, it can be asked any question and will generate answers as precisely as possible according to its training data. The data on which large language model chatbots refer back to are usually based on next-token-prediction and masked-language-modeling. In the first method, chatbots are trained by feeding them sequences of words to predict the following words in the sentence. In the second, they are fed sentences in which singular tokens are left out for the a.i to fill in. While these methods help the program learn how to produce fluent
and natural-sounding sentences, they do not teach the AI how to differentiate between minor and significant semantic mistakes (Ramponi 2022). Ramponi exemplifies this with the following sentence: “The Roman Empire ‘MASK’ with the reign of Augustus” (ibid.). The
algorithm could fill in the gap with both the words ‘began’ or ‘ended’; however, while both would be historically correct, they imply different meanings indistinguishable for the a.i. To minimize these misalignments, ChatGPT was trained using a method called ‘Reinforcement Learning from Human Feedback (RLHF)’. In this method, human AI trainers provide conversations in which they play both sides; the user and an AI assistant. Additionally, ChatGPT is equipped with large amounts of raw text data created by humans. These texts can derive from different kinds of sources like social networks, news reports, online panels, and books from academic online libraries. These sources are disconnected from the Internet; ChatGPT, therefore, does not crawl the internet for information like many others of its kind but instead refers back to a fixed amount of data exclusive to the year 2022.
Through training with various kinds of data sources and human input to evaluate its authenticity and lifelikeness, ChatGPT has become increasingly assertive in answering analytical questions frequently found in school assignments and academic research papers. It has also proven its ability to produce comedic and literate output, as shown by a Twitter user who asked ChatGPT to produce “a biblical verse in the style of the King James Bible explaining how to remove a peanut butter sandwich from a VCR” (Roose 2020). Most importantly, though, ChatGPT is not stateless, which means that other than most of its fellow chatbots, it can actually remember and learn from previous conversations. Some even have already gone to such lengths as calling ChatGPT “the end of term papers” (Bach & Weßels: A1). Therefore, ChatGPTs’ wide-ranging capabilities of producing various kinds of literary output inevitably lead to the question of the possible existence of creative capabilities within Open’s new AI. There is a problematic nature to the way OpenAI’s new tool has been evaluated in the past months, as the creativity which ChatGPT has often been adjudicated with is based mainly on the final products it creates. This paper argues that ChatGPTs’ creativity should be judged according to its production process and overall human reception of it instead of mainly looking at its produced items. Furthermore, as Pereira argues in his chapter on creative systems, a computational system such as ChatGPT has to fulfill the following conditions to be called creative:
– It should aim to produce solutions that are not replications of previous solutions (known to it).
– It should aim to produce solutions that are acceptable for the task it proposes. (2007: S. 36)
On one side, ChatGPT has shown its capability to produce new and meaningful products by utilizing past knowledge. However, on the other side, it is undeniable that ChatGPT does so only through the computation and reevaluation of already existing data. Therefore, its actual products are not the result of what Pereira would refer to as “divergent and convergent processes” (ibid.: 29) but of an algorithm reproducing already existing knowledge. Its algorithm is also still prone to the same misalignments it shares with already existing chatbots, as it can often not differentiate between similar semantic tokens. These hallucinations, as they are referred to, often result in outrageous misinformation, challenging its ability to produce acceptable solutions for given tasks. Therefore, the first of Pereiras’ conditions is problematic as ChatGPT seems trapped in what Turing calls the “paradox of creativity” (Hagengruber 2017: 337). As a learning machine, ChatGPT can only produce something new based on what has already been disclosed and is hence trapped in a cycle of producing knowledge only based on the already
known. From this spectrum, it seems that ChatGPT is an uncreative entity without any capacity to produce new knowledge. However, as Hagengruber continues to point out, human communities are themselves trapped in a similar paradox as they can only integrate something new into their corpus of knowledge based on the already known (vgl. ibid.). Furthermore, humans have added information produced by artificial intelligence into their common knowledge framework in the past before. In fact, so much so that Google’s search engine has become the first choice for answering questions in most daily situations for many people. Digital watches can govern mental and physical health states and offer perspective into them, one into which many humans put their trust. Social Networks have created digital bubbles in which entire societies can form, with many silently accepting the nature of the digital reality they operate in. Or, as Hagengruber puts it, “the concepts produced by our digital machines provide a stable part of our world interpretation. We live in a world which is, to a great extent, still structured by artificial concepts which we are free to accept and give “meaning” to (ibid.: 341). From this perspective, it is irrelevant if ChatGPTs’ inner workings fulfill the definition of what humans consider to be a creative process because, as a global society, humans have already accepted ChatGPT into their knowledge framework. May it be to produce comedic short stories, have regular ‘human-like’ conversations with, or have it produce entire academic papers. ChatGPT has already broken records and has been downloaded over a million times worldwide, raising OpenAIs’ net worth to around 29 billion dollars (vgl. Varanasi 2023). To conclude, only months after its release, ChatGTP has been titled the literal end for academic writing as known to modern human society and “is already being compared with to the iPhone in terms of its potential impact on society” (Roose 2022). Considering its still many lackluster features, such as its constant hallucinations and uneven accuracy, those comments might arguably turn out to be overstatements. However, it needs to be emphasized
that ChatGTP is only another prototype, and OpenAI has already announced another upgraded version to release this year. If taking into account how profound of an impact ChatGTP has had in only a short time, it is difficult to imagine how the future will unfold with even better and improved versions of it. However, two things seem to have become more certain: humanity’s openness to accept AIs’ into their communities and their willingness to accept their produced items into knowledge corpora.
Bach,Susanne & Doris Weßel (2022): Das Ende der Hausarbeit, In: Frankfurter Allgemeine, 21.12.2022, S. A1.
Hagengruber, Ruth (2017): Creative Algorithms and the Construction of Meaning, In: W. Pietsch et al. (Hrsg.), Berechenbarkeit der Welt?, Wiesbaden, Deutschland: Springer.
Metz, Cade (2022): The New Chatbots Could Change the World. Can You Trust Them?, in: The New York Times, 10.12.2022 [online]
https://www.nytimes.com/2022/12/10/technology/ai-chat-bot-chatgpt.html [last viewed on 12.01.2023].
Kühl, Eike (2022): Gut erfunden ist halb geglaubt?, in: Zeit Online, 06.12.2022 [Online] https://www.zeit.de/digital/internet/2022-12/chatgpt-kuenstliche-intelligenz-openaichatbot [last viewed 12.01.2022].
Pereira, Francisco Camara (2007): Creativity and Artificial Intelligence. A Conceptual Blending Approach, Berlin, New York, Deutschland: Mounton de Gruyter.
Ramponi, Marco (2022): How ChatGPT actually works, AssemblyAI [Online] https://www.assemblyai.com/blog/how-chatgpt-actually-works/ [last viewed 12.01.2023]
Roose, Kevin (2022): The Brilliance and Weirdnees of ChatGPT, in: The New York Times, 02.12.2022 [online] https://www.nytimes.com/2022/12/05/technology/chatgptaitwitter.html?searchResultPosition=1 [last viewed on 12.01.2023].
Schwabe & Co (1976): Kreativität, in: Historisches Wörterbuch der Philosophie, Basel, Stuttgart, Deutschland: Schwabe & Co., S. 1193 [Wörterbucheintrag]
Varanasi, Lakshimi (2023): ChatGPT creator OpenAI is in talks to sell shares in a tender offer that would double the startup’s valuation to $29 billion, in: Business Insider [online] https://www.businessinsider.com/chatgpt-creator-openai-talks-for-tender
offer-at-29-billion-2023-1 [last viewed on 13.01.2022].
You cannot copy content of this page