Year: 2024
Type: dissertation
Abstract: This thesis studies whether neurodivergent rhetorical traits can be repre-sented with a conversational AI system. As the data demands of fine-tuning custom large language model (LLM) solutions have lowered, it is feasible to cater to smaller demographics. In the Tuumailubotti (Pondering bot) project, the FinGPT-3 LLM was trained with a dataset authored by an anonymous group consisting mainly of members of neurominorities. The model was then used to create a demo for a chatbot for discussing performance review-adjacent topics. By applying neuroqueer theory into the design process of a chatbot, the project critiques the neuronormative assumptions that underlie the design of mainstream generative AI assistants and tries to imagine an alternative space for private reflection for neurominority employees. To assess whether the rhetoric of the LLM was neuroqueered via fine-tuning, the thesis investigates which rhetorical traits were present in the text outputs of the chatbot, and how they affected the preferences of users. In a user study, 31 participants had an hour to interact with the chatbot and then fill in a survey. The participants classified quotes from their conversations into five categories: good, bad, weird, toxic and neurodiverse. These quotations were then coded by the thesis author using a list of common neurodivergent rhetorical traits, such as fast switching of topics, repetition and exactness. The results show that many of the coded traits were prevalent in the quotes. However, without a direct point of comparison, such as a chatbot trained with data authored by neurotypicals, it is unclear to what extent rheto-ric factored in the preferences of the testers. From a design perspective, better results might have been achieved with a more participatory design process. While overall impressions of the chatbot were not negative, the coded traits were mostly perceived negatively by the participants. The participants did not identify the traits as neurodivergent, and mostly viewed them as bad design decisions. The responses indicate that the project succeeded in its goal of challenging implicit assumptions the participants had about the rhetoric of chatbots, therefore neuroqueering rhetoric.
Source: Aaltodoc (Aalto University)
Author: Luostarinen, Vertti
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7021178017