(2/6) We ran two pre-registered studies (N=470, 604) where participants watched short videos framing LLMs as machines, tools, or companions. In both, participants rated their beliefs about LLM mental capacities. In the second, they also used LLM outputs to answer factual questions.
(5/6) 📌This matters because it shows that even brief messages about LLMs can shape how people understand, trust, and rely on them. As these systems become embedded in everyday life, the way we talk about them—across research, media, and product design—has real consequences.
(3/6) 🔑1️⃣: Participants who watched LLMs-as-companions (vs. no video) believed LLMs were more capable of many cognitive and emotional capacities (see fig). However, watching LLMs-as-tools or machines (vs no video) could alter other beliefs about LLMs (see paper!).
How should we talk about LLMs? Does it matter if we frame them as a machines 📠, tools ⚒️, or companions 👥? Our #CHI2026 paper shows that these framings can alter what people believe about LLMs and how they use them. See 🧵!
Paper: arxiv.org/abs/2510.18039
Presentation: Friday 11:15am in P1-128
(4/6) 🔑2️⃣: In Study 2, we observed that the videos had a nuanced effect: watching the LLMs-as-machines video did lead to participants submitting answers that agreed with the LLM responses less, but only when the LLM responses were inconsistent.
(1/6) People talk about AI systems (including large language models, or LLMs) in different ways. We first ask how this might shape what people believe about the technology–primarily, what mental capacities (e.g., the ability to have intentions) people attribute to LLMs.
(6/6) This work was done in collaboration with @sunniesuhyoung.bsky.social , Angel Franyutti, Amaya Dharmasiri, @kushinm.bsky.social, Olga Russakovsky, and @judithfan.bsky.social. I’ll be presenting this work at the Modeling Minds and Mentalities Session (Friday April 17 @11:15am–Barcelona time)
