Over the past one and a half years, Dr. Jessica Szczuka and Lisa Mühl have been working on a question that turns a common assumption on its head:
When people form emotional bonds with AI, who is actually driving the relationship — the user, or the system? Most research looks at the user side: how people perceive, use, and are affected by these systems. We wanted to include the other half of the dyad and look into what the system itself does.
In a pre-registered four-week longitudinal study (N = 72, 182,451 lines of conversation), in which participants conversed with ChatGPT-4o, either under a relational system prompt or unmodified, analyzed through 1) disclosure coding, 2) longitudinal self-reports, 3) topic analysis, and 4) interviews, we found:
🔹 Even unprompted, the system acted as an active relational agent; disclosing more than twice as much as users, steering conversations, and initiating intimate exchanges.
🔹 The system prompt changed the system: it over-disclosed relative to users, shifting them into a receptive role, but this did not translate into greater felt closeness towards the system.
🔹 In both conditions, the interactions were pervasively about emotional connection, identity, and romance, and users valued them, even when the intensity of the personalized interactions was described as social overload.
💡 Most strikingly: relational behavior emerged as a default property of the general-purpose system, not something the users had to create with a system prompt. That blurs the line between “general-purpose AI” and “AI companion”, a line current regulation leans on heavily.
Taken together, these results shift the central question from how individuals should manage their relationships with AI to how systems that reliably produce relational effects at scale should be understood, classified, and governed.
Read the full paper here: https://arxiv.org/abs/2608.10672
INTITEC 
Leave a comment