Neurodiverse Rhetoric with Conversational AI: An Introduction
In the Pondering Bot project, I designed an AI chatbot for representing neurodiverse rhetoric. My master’s thesis describes the background and development of the project and analyses whether it succeeded in its goal. This is the introductory chapter to the thesis, which is now available online.
Some Basic Terms and Definitions
Neurodiversity
Neurodiversity studies scholar Nick Walker (2014) defines neurodiversity as “the diversity of human minds, the infinite variation in neurocognitive functioning within our species”. Neurodiversity means only the biological fact that we all have a different brain. However, as a concept it invites parallels to biodiversity, and in so doing can be interpreted as contrasting pathologizing dichotomies about what constitutes as a “correct” human brain.
Neuroqueer
Neuroqueer, like queer, is simultaneously a verb and an adjective. To be neuroqueer is to neuroqueer (Walker, 2015), to engage in the practices of neuroqueering. While the term alludes strict definitions, in this thesis I concentrate on neuroqueering as a creative practice.
Nick Walker (2015) writes that neuroqueering can be “engaging in practices intended to undo and subvert one’s own cultural conditioning and one’s ingrained habits of neuronormative and heteronormative performance, with the aim of reclaiming one’s capacity to give more full expression to one’s uniquely weird potentials and inclinations.” While Walker also lists all media that foregrounds neurodivergent voices, I would not go as far in my definition. My personal viewpoint is that neuroqueering media implies some form of subversion of conventions that goes deeper than surface-level representations.
Background and Motivation
Large language models, or LLMs, are a subset of artificial intelligence systems that model how humans produce language. LLMs emulate language through statistical methods, namely a type of machine learning called deep learning. Typically, they are used as the foundation for conversational AI systems, which are colloquially referred to as chatbots.
In the book Authoring Autism (2017), neurodiversity studies scholar M. Remi Yergeau argues that autism queers rhetoric by transgressing societally ingrained dichotomies, such as communicative/non-communicative. The concept of neuroqueer rhetoric encompasses a range of traits, such as repetition and fast topic switching, that have been thought of as involuntary and therefore arhetorical in medical and medicalizing discourses.
It is important to disambiguate between neuroqueer and neurodiverse rhetoric. With neuroqueer rhetoric, I am referring to Yergeau’s concept, whereas with neurodiverse rhetoric, I mean a rhetoric that represents how people with various neurodivergences communicate.
Throughout the history of AI, the capabilities of AI systems have been equated with their perceived rhetoricity, their ability to affect people with communication. Applying Yergeau’s (2017) concepts to modern LLM research, a pattern of erasure begins to emerge. Even though LLM architectures are only loosely inspired by the human brain, researchers back their innovations with neuroscience and psychology. In this body of research, being human is equated with rhetoricity, while autism is portrayed as its antithesis. Creating humanlike intelligent machines thus means creating ones that are less like autistics (Bruder, 2024). Psychological surveys that were originally designed to diagnose autism are now being redeployed as benchmarks for LLM advancement (e.g. Starchan et al. 2023, Ni et al. 2024).
So far, most use cases for LLM-based technologies aimed at neurodivergent users have been medical. These solutions often aim to rehabilitate members of neurominorities through behaviouristic rewarding (e.g. Tang et al. 2024). They frame neurodivergence as disability, and then portray disability, as Yergeau (2014) writes in the context of charity hackathons, as “pitiable and in need of remediation”.
A particularly worrying research trend is using socially assistive robots and chatbots to ‘train’ autistic youths in communication skills (Roderick, 2023); in other words, teaching them how to mask their autistic traits. This is telling of the current pecking order: in the continuum of rhetoricity formulated by Yergeau (2017), machines surpass autistics, but still do not rank as high as humans, so they can be used as intermediaries.
My motivation for writing this thesis was to highlight the implicit notions of rhetoricity that conversational artificial intelligence systems represent. To this end, I wanted to explore what it would mean to neuroqueer a conversational AI system.
The Project: Neuroqueering a Chatbot
Tuumailubotti could be loosely translated as “the pondering bot.” It is a demo for a conversational AI system for employees to privately reflect work life and work-related topics. The chatbot was commissioned by the Finnish Broadcasting Company, Yle. I led the design, dataset curation and technical execution of the system, and oversaw its testing. The dataset that was produced during the project is openly accessible (Luostarinen, 2023).
Tuumailubotti served many purposes for the different parties involved in its making. For Yle’s HR department it was part of an effort to better cater to the needs of neurodiverse employees. For Yle’s News Lab and Development departments it was a technical experiment on how custom-built AI systems could be integrated into Yle’s software infrastructure. For me, it served as a research project. The needs of these interest groups had to be consolidated, and the aims of the project were renegotiated multiple times.
The resulting chatbot is a unique amalgamation of an HR software demo and art. As such, it is queerly situated between conflicting definitions and categories, not completely unlike the people it aims to represent. I think it would be easiest to classify Tuumailubotti as critical design and my research practice as critical design research. As formulated by Dunne and Raby (n.d.), “Critical Design uses speculative design proposals to challenge narrow assumptions, preconceptions and givens about the role products play in everyday life.”
The FinGPT-3 family of Finnish-speaking LLMs (Luukkonen et al. 2023) had just been released, and I was curious about how they would perform as a base for a conversational system. Taking this approach was a calculated risk, as it was not certain if the data we managed to gather would suffice. In addition, all evaluations of the performance of the Finnish GPT models were based on my own limited and subjective experiments. To my knowledge, Tuumailubotti was the first generative chatbot with native-level Finnish proficiency.
Research Questions and Methodology Overview
My research questions are:
1. Are there neurodivergent traits in the text outputs of the chatbot? If so, how do they vary?
2. Do neurodiverse rhetorical traits affect testers’ preferences of the chatbot?
I investigated these questions by conducting a user test with 31 participants. The participants were a mix of neurominority members and non-neurominority members, most of whom were currently working. They were recruited from both inside and outside of Yle.
Neuroqueer rhetoric is always unique to the situated and embodied circumstances where it manifests and therefore eludes strict definitions (e.g. Egner, 2018). This made evaluating the success of the project methodologically complex, as all the approaches I could have taken were bound to be incomplete. As I will cover in Chapters Two and Four, most, if not all, existing methodology used to analyse neurodivergent traits through rhetorical or conversation analysis is pathologizing and could not be utilized as is. Online chats fall somewhere between written correspondence and spoken conversation, meaning that existing research on the topic is scarce. Moreover, as the dataset represents a potentially wide spectrum of neurodivergence, and the contributors’ demographics were not gathered, the list of criteria could have been endless.
This thesis is multidisciplinary even though I never set out to make it as such. My guiding principle was to utilize the theories and methods that best fit the needs of each step of the project. Methodologically, it is positioned as critical design research, although I also deploy methods from human-computer interaction and AI alignment research. The thesis also utilizes theoretical frameworks from both new media studies and neurodiversity studies. What I hope binds all these disparate elements together is the methodology of practice-based design research.
Having a background in screenwriting, I approached the project first and foremost as an endeavour in collective creative writing. I view the technical parts of the work as the natural continuation of my creative practice.
Core Design Decisions
The decision to make Tuumailubotti a neurodiverse chatbot instead of a representation of a specific neurominority was done for both practical and theoretical reasons. The identity landscape of neurodivergence is a complicated one, as issues of neurodiversity are inseparably intertwined with those of race, gender, and class. I advocate for people’s right to self-identify as neurodivergent because not everyone has equal access to medical diagnoses. Diagnoses are used to gatekeep access to services and medication, but they are also used to discriminate, for example, by preventing access to health insurance.
If the dataset would have included only rhetoric from people belonging to a specific minority, it could have been used to identify and target said minority. For the same reasons, I did not do a comparative study with a version of the dataset that only includes data from neurotypical contributors. From information volunteered by some participants, I can estimate that at least 70% of the dataset was written by neurominority members, but as we had decided not to collect this information, there is no way to be certain.
Acknowledgements
I would like to thank The Media Industry Research Foundation of Finland for their thesis grant, without which this thesis would have been significantly shorter.
The training of Tuumailubotti was achieved with a graphics card purchased with a materials expense grant from the Tukilinja Foundation.
I want to thank all the anonymous data authors and study participants, everyone from the neurodiversity network who gave us feedback as well as everyone from Yle who participated in the project: Marika Björn, Saija Uski, Anni Klutas, Johan Sundström, Samuli Sillanpää, Mikko Lehtimäki, Alexander Alafuzoff, Perttu Ehn, Matias Kainulainen, Pekka Salmela, Tuomo Virolainen and Tuukka Ojala.
I would also like to thank my thesis supervisor Markku Reunanen, as well as my advisors Christian Guckelsberger and Fran Trento.
Additional thanks go to Tove Mylläri and Jouni Frilander from Yle, Matti Niinimäki and Lily Diaz-Kommonen from Aalto University, Jouni Luoma from Silo AI, Begüm Çelik and Anttu Koistinen from CasvuAI, Pinja Rauhamäki from The Finnish Meteorological Institute and my opponent Guus Hoeberechts.
Lastly, I would like to thank my parents Heikki Luostarinen and Laura Tohka for their support.
Sources
Bruder, J. (2024). AI as medium and message The (im)possibility of a queer response. (Klipphahn-Karge, M., Koster, A., & Morais, S. Morais dos Santos Bruss, S. Eds.) Routledge Studies in New Media and Cyberculture: Queer Reflections on AI. (pp. 162–175.) https://doi.org/0.4324/9781003357957-13
Dunne, A. & Raby, F. (n.d.) CRITICAL DESIGN FAQ Retrieved July 7th 2024 from https://dunneandraby.co.uk/content/bydandr/13/0
Egner, J. E. (2019). “The Disability Rights Community was Never Mine”: Neuroqueer Disidentification. Gender & Society, 33(1), 123–147. https://doi.org/10.1177/0891243218803284
Luostarinen, V. (2023). Tuumailubotti. [Data set] https://huggingface.co/datasets/Yleisradio/Tuumailubotti
Luukkonen, R., Komulainen, V., Luoma, J., Eskelinen, A., Kanerva, J., Kupari, H.-M., Ginter, F., Laippala, V., Muennighoff, N., Piktus, A., Wang, T., Tazi, N., Scao, T., Wolf, T., Suominen, O., Sairanen, S., Merioksa, M., Heinonen, J., Vahtola, A., … Pyysalo, S. (2023). FinGPT: Large Generative Models for a Small Language. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. https://doi.org/10.18653/v1/2023.emnlp-main.164
Ni, Q., Yu, Y., Ma, Y., Lin, X., Deng, C., Wei, T., & Xuan, M. (2024). The Social Cognition Ability Evaluation of LLMs: A Dynamic Gamified Assessment and Hierarchical Social Learning Measurement Approach. ACM Transactions on Intelligent Systems and Technology https://doi-org.libproxy.aalto.fi/10.1145/3673238
Roderick, I. (2023). Autism Robot Therapy, Remediation, and Mimetic Disabling. Media Theory, 7(2), 103–126. Retrieved from https://journalcontent.mediatheoryjournal.org/index.php/mt/article/view/585
Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behavior 8, 1285–1295. https://doi.org/10.1038/s41562-024-01882-z
Tang, Y., Chen, L., Chen, Z., Chen, W., Cai, Y., Du, Y., Yang, F. & Sun, L. 2024. EmoEden: Applying Generative Artificial Intelligence to Emotional Learning for Children with High-Function Autism. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI ‘24). 1–20. https://doi-org.libproxy.aalto.fi/10.1145/3613904.3642899
Yergeau, R. M. (2014). Disability Hacktivism. In Sayers, J. & Hocks, M. (Eds.) Hacking the Classroom: Eight Perspectives. Computers and Composition Online
Yergeau, R. M. (2017). Authoring Autism: On Rhetoric and Neurological Queerness. Duke University Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is All You Need. 31st Conference on Neural Information Processing Systems (NIPS 2017). https://doi.org/10.48550/arXiv.1706.03762
Walker, N. (2014). Neurodiversity: Some Basic Terms & Definitions. Viewed 30.3. 2023 in https://neuroqueer.com/neurodiversity-terms-and-definitions/
Walker, N. (2015, revised 2021). Neuroqueer: an introduction. Viewed on 30 March 2023 in https://neuroqueer.com/neuroqueer-an-introduction/
