AI Placebo Effects: Expectancy Effects on Emotional Benefit from AI Interaction

The aim of this study is to test how the framing of an AI system's design influences emotional outcomes, and whether this effect is moderated by individual differences in attitudes toward AI companionship. Previous research has found that people find empathetic AI to be effective in reducing negative mood and reducing the effects of feeling socially excluded (de Gennaro et al., 2020). This effect has also been consistent in an AI framing context, where AI labelled as "caring" was much more likely to be rated as more empathetic and effective compared to a neutral framing or manipulative framing group (Pataranutaporn et al., 2023). An important moderator in the relationship between framing type and perceived empathy of the AI was attitudes towards AI (Pataranutaporn et al., 2023). People who had a more positive attitude rated AI higher on empathy, trustworthiness and effectiveness (Pataranutaporn et al., 2023). Another predictor of how people engage with AI is an individual's tendency to anthropomorphise it (Pentina et al., 2023). Anthropomorphising AI also affects how emotionally attached a person feels to AI. Attributing human qualities to AI has been found to be a predictor of participants engaging in more frequent self-disclosure, in which people benefit from self-disclosure to an AI system just as much as to a human, producing the same emotional, relational and psychological effects (Pentina et al., 2023; Ho et al., 2018). In subjective evaluations of jokes, people tend to attribute high‑quality jokes to humans and low‑quality jokes to AI when the source is unknown (Bower et al., 2021). However, when the source is given, and high-quality jokes are attributed to AI, people are more likely to rate these jokes highly (Bower et al., 2021). This illustrates how the framing of AI can change the rating of subjective measures. The expectancy effect in this study indicates that clearly labelling the source made participants respond more positively to higher‑quality jokes compared to when it was unlabelled (Bower et al., 2021). The present study will aim to build on this finding by investigating whether being transparent about an AI system's benefits and limitations positively influences participants' emotional outcomes. Baseline measures will be collected, including current mood, stress levels, demographic information, and frequency of prior AI use. The moderator will be measured by an AI companion motives scale. Participants will be randomly assigned to three different conditions, where the framing of the AI design is either a General-Purpose tool (control), a Caring AI or an Open-Label expectancy condition (transparent). Participants will then engage in a text‑based conversation with the AI system, in which the AI asks about their daily experiences, feelings, and stressors. There is no required disclosure from the participant. This session will end after 15 minutes. After this interaction, participants will be asked to report their perceived caring and level of trust in AI, as well as their intentions to use AI in the future. The baseline measures of current mood and stress will be reassessed. The expected outcomes are that participants who score highly in AI companion seeking motives will have a more positive outcome from the Caring AI framing condition compared to the General-Purpose tool. Higher companion seeking motives are also expected to strengthen the positive outcome of the Open-Label placebo. In this condition, rate it as more caring and have greater trust in it compared to participants with lower companion seeking motives.