Self-reported emotional change in conversations with an emotional support assistant: an observational analysis in TerapIA
Abstract
People who talk to an emotional support assistant usually want to feel a bit better by the time they stop talking. Measuring whether that happens, and under what conditions, is a different thing from proving clinical efficacy. We analyze the change in a continuous emotional score in [0, 1] that users record at the beginning and at the end of their conversation with Pía, the assistant of the TerapIA app. The analysis is restricted to conversations where both ratings are present, where the slider was moved away from the interface default at both ends, and where the dialogue reached at least ten user messages. Once we discount the automatic greeting that opens the chat, that threshold amounts in practice to a conversation with actual content. Over n = 1935 conversations from 909 people, the mean score went from 0.42 at the beginning to 0.56 at the end (Δ = +0.138; 95% CI by user-clustered bootstrap [+0.121, +0.155]; dz = 0.54; Wilcoxon p < 0.001), and 63.4% of conversations improved. The mean change is zero in very short dialogues and consistently positive from ten user messages onwards, with an association between length and improvement that is statistically significant and robust to clustering by person. The result holds, and is in fact slightly larger, if only the first eligible conversation of each person is used (Δ = +0.147; dz = 0.57). Among the people who contributed at least three such conversations, the end score of the last one minus the start score of the first was +0.195, and 70% of them went up; the emotional starting point also rises with the number of accumulated conversations within each person. The design is observational: there is no control group, filling in the rating is optional, the score is a self-report, and regression to the mean is present. We do not infer causality, and we do not claim equivalence with a clinical outcome. What we do document is a robust association between talking with Pía for enough turns and a better self-reported emotional state at the end, of a magnitude close to the one observed in other deployed conversational agents with a pre/post within-session measure.
Related
- Intelligent-assistant support versus usual self-scheduling for extracurricular physical activity among university students: a comparative evaluation.
- Marital satisfaction and parent-reported emotional competence in Chinese preschoolers: serial indirect associations through mindful parenting and parental involvement
- Perceived Social and Emotional Support from Co-Workers and Supervisors and Occupational Stress in Emergency Physicians: A Pre-Pandemic National Cross-Sectional Study from Slovenia
- Latent profiles of ambivalence over emotional expression and associations with social support among Chinese patients with cervical cancer
- Gender Differences in Self-reported Emotional Sharing in Social Interaction: A Meta Analysis
BibTeX
@misc{serrano2026selfreported,
title = {{Self-reported emotional change in conversations with an emotional support assistant: an observational analysis in TerapIA}},
author = {Alejandro Vaca Serrano and Alberto Vaca Escribano and José María González de Echevarri},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/jgxcy_v1},
url = {https://osf.io/preprints/psyarxiv/jgxcy_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Self-reported emotional change in conversations with an emotional support assistant: an observational analysis in TerapIA
AU - Alejandro Vaca Serrano
AU - Alberto Vaca Escribano
AU - José María González de Echevarri
PY - 2026
DA - 2026/09/16
PB - PsyArXiv
DO - 10.31234/osf.io/jgxcy_v1
UR - https://osf.io/preprints/psyarxiv/jgxcy_v1/
AB - People who talk to an emotional support assistant usually want to feel a bit better by the time they stop talking. Measuring whether that happens, and under what conditions, is a different thing from proving clinical efficacy. We analyze the change in a continuous emotional score in [0, 1] that users record at the beginning and at the end of their conversation with Pía, the assistant of the TerapIA app. The analysis is restricted to conversations where both ratings are present, where the slider was moved away from the interface default at both ends, and where the dialogue reached at least ten user messages. Once we discount the automatic greeting that opens the chat, that threshold amounts in practice to a conversation with actual content. Over n = 1935 conversations from 909 people, the mean score went from 0.42 at the beginning to 0.56 at the end (Δ = +0.138; 95% CI by user-clustered bootstrap [+0.121, +0.155]; dz = 0.54; Wilcoxon p < 0.001), and 63.4% of conversations improved. The mean change is zero in very short dialogues and consistently positive from ten user messages onwards, with an association between length and improvement that is statistically significant and robust to clustering by person. The result holds, and is in fact slightly larger, if only the first eligible conversation of each person is used (Δ = +0.147; dz = 0.57). Among the people who contributed at least three such conversations, the end score of the last one minus the start score of the first was +0.195, and 70% of them went up; the emotional starting point also rises with the number of accumulated conversations within each person. The design is observational: there is no control group, filling in the rating is optional, the score is a self-report, and regression to the mean is present. We do not infer causality, and we do not claim equivalence with a clinical outcome. What we do document is a robust association between talking with Pía for enough turns and a better self-reported emotional state at the end, of a magnitude close to the one observed in other deployed conversational agents with a pre/post within-session measure.
ER -