More and more disciplines within the humanities are being significantly impacted by AI technologies. One such discipline is AI-generated text production, which already seems to be bringing about changes in how students and professionals write texts and generate content. The way text producers work with AI text generators (ATGs) thus needs to be to be empirically explored to be able to adjust teaching in the humanities. 

The article is based on insights from an empirical study, which was carried out Q1 2021, investigating how students and professionals work with a selected ATG and what they see as the most important strengths and weaknesses of ATGs. The descriptive-analytical study involved a total of 70 test subjects. First, the test subjects were asked to work with a specific ATG and conduct three writing operations. Second, having tested the ATG, the test subjects were asked to participate in an online questionnaire focusing on how they experienced working with the ATG. The quantitative data resulted in five column diagrams about their ATG perception, and the qualitative data were thematically analysed by means of NVIVO resulting in a multitude of quotes and tree structures illustrating how the test subjects worked with the ATG. 

The data seems to suggest that most of the test subjects in fact found that the ATG in question was easy to use when producing texts, but the data also suggest that the test subjects found the quality of the ATG- generated content to be below standard and that they had to perform several editing operations before, during and after the automatic text generation. Based on the insights, the article presents a theoretical framework for facilitating optimum use of ATGs in connection with text production in the humanities.

Biography

External Lecturer, Department of Management, Society and Communication (MSC), CBS, Director Fremdriften, Director of Resources & Projects, SmartLearning (SL).

This paper explores the potential of artificial intelligence (AI) to generate academic writing. We focus specifically on the recently released Generative Pre-trained Transformer 3 (GPT 3) model, which is a state-of-the-art AI model for text generation. We evaluate GPT 3's ability to generate academic writing by having it generate abstracts for research papers in the field of computer science. We find that GPT 3 is able to generate well-formed abstracts that are comparable in quality to those written by humans. However, we also find that GPT 3's abstracts tend to be shorter and simpler than those written by humans. Overall, our results suggest that AI has great potential for generating high- quality academic writing. 

Is this abstract written by an artificial intelligence? Both yes and no. In fact, the lines above are written by GPT 3. It was generated by using the title and the keywords as prompt, and even though I’m neither a ‘we’ nor work within the field of computer science, the abstract comes very close to describing my project. But what does it mean that computer generated text is ‘comparable in quality’ to text produced by humans? I attempt to answer this question by exploring a series of ‘abstracts’ and ‘research papers’ generated by GPT 3. I do so by applying the perspective of functional linguistics focusing especially on the notion of academic voice. Voice is closely related to notions such as register, context and communicative purpose, and I will try to show how the human- like qualities of the AI generated texts can be understood as the perceived presence of an academic voice belonging to a rational sender with a communicative purpose. Keywords: Artificial intelligence, automated writing, GPT 3, voice, academic writing 

Author (human): 

Ulf Dalvad Berthelsen, Associate Professor, Ph.D. 

School of Communication and Culture, Scandinavian Studies 

I work with language, literacy and writing in an educational perspective with a special focus on how these areas are affected and transformed by digital technologies. 

Recent publications: 

Berthelsen, U. D. (2020). Digitale tekster og skriftlig fremstilling i gymnasiet: Et curriculumperspektiv. Tidsskriftet Læring og Medier (LOM), 13(23), Article 23. doi.org/10.7146/lom.v13i23.120963 

Berthelsen, U. D., & Nielsen, C. F. (2021). Democracy and Computation: A Normative Perspective on the Magic of the New Millennium. I Computational Thinking in Education. Routledge. 

Berthelsen, U. D., & Tannert, M. (2019). The Ecology of Analytics in Education: Stakeholder Interests in Data-Rich Educational Systems. International Journal of Learning Analytics and Artificial Intelligence for Education (IJAI), 1(1), 89–101. https://doi.org/10.3991/ijai.v1i1.11023

Tannert, M., Lorentzen, R. F., & Berthelsen, U. D. (2021). Computational Thinking as Subject Matter: As an Independent Subject or Integrated across Subjects? I Computational Thinking in Education. Routledge.

Yadav, A. & Berthelsen, U. D. (2021). Computational Thinking in Education: A Pedagogical Perspective. Routledge.

Public interest in AI and big data has tremendously increased over the recent years due to deep learning—the latest paradigm shift in AI— and its increasing potential to be implemented in everyday technologies. Deep learning’s potential capacity to bring new opportunities in many fields is yet not tainted with challenges. Gender bias and racism are some of these challenges which can be explored through the language produced by AI. Aiming to contribute to the research revolving around this topic, in this presentation we will delve into auto-generated texts and biased content they have the potential to reproduce when prompted with such biased input. More specifically, we will explore the use of language and production of biases such as racism and sexism in GPT-3 generated texts based on the prompts from Daniel Defoe’s famous 18^th-century novel Robinson Crusoe and John Ruskin’s Victorian essay “Of Queen’s Gardens.” Our choice of texts is quite intentional since the former is infused with imperialist trajectories and racism while the latter is reflective of Victorian gender roles constructing women as submissive and passive. Importantly, while generating texts based on biased content retrieved from these texts although some level of auto-generated warnings is raised for mostly explicit sensitive contents, not all bias is detected; especially, the ones with subtle sexism and other forms of implicit bias go unnoticed. Thus, we will analyze such instances of subtle biases to discuss the challenges and opportunities that arise from deep learning.