Only Time Will Tell

Dr. Yulu Pi, Research Associate at RC Trust, University of Duisburg-Essen, and Rosco Hunter (University of Warwick) have published a new open-access article in the International Journal of Human–Computer Interaction.

AI companions such as Replika are increasingly embedded in everyday life. As people turn to these systems for emotional support, companionship, and even relationship-like interactions, researchers, policymakers, and developers urgently need to understand their long-term effects. However, most existing research captures only a single snapshot in time, leaving open the question of how these effects – whether beneficial or harmful – unfold and change as people engage with AI companions over longer periods.

To address this gap, Pi and Hunter systematically reviewed 17 longitudinal studies of social AI companionship, comparing how each study was designed, what methods it used, and what it found. Their analysis reveals a striking inconsistency at the heart of this growing field: although most of the reviewed papers describe themselves as “longitudinal,” none actually define the term, and few track how user behaviour or wellbeing changes over the course of the study rather than simply comparing a before-and-after snapshot.

The review also surfaces a set of critical questions that remain largely unexplored, including:

Sycophancy – whether prolonged exposure to overly agreeable AI companions can leave users’ beliefs unchallenged and their resilience to real-world friction gradually eroded

Vulnerable groups – how children, older adults, or individuals with mental health conditions may experience distinct risks from sustained AI companionship

Complementarity vs. substitution – whether AI companions support existing human relationships over time, or gradually come to replace them

Based on their synthesis, the authors propose a concrete roadmap for future research, including clearer criteria for what qualifies as longitudinal research, pre-registration of study protocols, more diverse participant samples, and stronger ethical safeguards for studies that may touch on emotionally sensitive territory.

The article is open access and available now via the International Journal of Human–Computer Interaction.

Read the full article: https://doi.org/10.1080/10447318.2026.2670529