Theory of Mind and Self-Presentation in Human-LLM Interactions
by Wester, Joel, Jacobsen, Rune Møberg, de Jong, Sander, Kollerup, Naja Kathrine, Djernæs, Helena Bøjer and van Berkel, Niels
Abstract:
The use of large language models (LLMs), such as ChatGPT, for social support and other activities is growing. LLM-based interactions require users to express themselves through text, a medium in which people's distinct self-presentation styles (SPS) present themselves. While the divergence of people's SPS is well-established, the effect of SPS on users' LLM interactions has not been explored. In this position paper, we point to this gap by drawing on insights from prior work on people's SPS online. Moreover, we discuss how Theory of Mind (ToM) can be used to increase our understanding of the possible effects of SPS on LLM output. Through this exploration, we shed light on how LLM responses are dependent on and sensitive to how people present themselves in their interactions with LLMs. We discuss the broader implications and suggest future research directions for HCI centred around people's SPS in interacting with LLMs—providing concrete suggestions on how effects of SPS on LLM output can be empirically explored.
Reference:
J. Wester, R. M. Jacobsen, S. de Jong, N. K. Kollerup, H. B. Djernæs, N. van Berkel, "Theory of Mind and Self-Presentation in Human-LLM Interactions", in Adjunct Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems (CHI'24 EA), Association for Computing Machinery, 2024, 1–4.
Bibtex Entry:
@inproceedings{Wester2024TheoryOfMindLLM,
	title        = {Theory of Mind and Self-Presentation in Human-LLM Interactions},
	author       = {Wester, Joel and Jacobsen, Rune Møberg and de Jong, Sander and Kollerup, Naja Kathrine and Djernæs, Helena Bøjer and van Berkel, Niels},
	year         = 2024,
	booktitle    = {Adjunct Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems},
	location     = {CHI'24 EA},
	pages        = {1--4},
	publisher    = {Association for Computing Machinery},
	url          = {https://nielsvanberkel.com/files/publications/chi2024b.pdf},
	howpublished = {Workshop paper},
	abstract     = {The use of large language models (LLMs), such as ChatGPT, for social support and other activities is growing. LLM-based interactions require users to express themselves through text, a medium in which people's distinct self-presentation styles (SPS) present themselves. While the divergence of people's SPS is well-established, the effect of SPS on users' LLM interactions has not been explored. In this position paper, we point to this gap by drawing on insights from prior work on people's SPS online. Moreover, we discuss how Theory of Mind (ToM) can be used to increase our understanding of the possible effects of SPS on LLM output. Through this exploration, we shed light on how LLM responses are dependent on and sensitive to how people present themselves in their interactions with LLMs. We discuss the broader implications and suggest future research directions for HCI centred around people's SPS in interacting with LLMs---providing concrete suggestions on how effects of SPS on LLM output can be empirically explored.}
}