PhD candidate Ezgi Dede is researching cutting-edge technologies, primarily AI and VR, at the UvA under the supervision of Prof. Dr Jochen Peter, Assistant Professor Hande Sungur and Jeroen Lemmens. She recently published a post on the COMCOM PhD blog, sharing interesting findings for those who want to learn more about AI companionship and what motivates us to use these AIs.
Interview with the AI: How do AI chatbots’ features support relationship formation?
The title of this study is a deliberate nod to Interview with the Vampire. Like a vampire, an AI chatbot looks human but isn't. It feels alive but isn't. It has a certain charismatic pull that turns humans into thralls. It feeds on humans. And it is immortal. Investigating AI companions with that kind of critical curiosity is exactly what this study is about.
Millions of people already chat with AI chatbots every day. AI offers advice, company, emotional support… What is it about these apps that draws us in and keeps us coming back?
A Theory That Fits
To study synthetic relationships (i.e., the bonds people form with AI companions) researcher have been struggling to find a theoretical lens that connects social psychology with human-machine relationships because most relationship theories assume two conscious, sentient partners exchanging costs and benefits. AI companions don't fit that.
This study proposes the elevator model of interpersonal relationship formation, a relatively under-used stage model. It maps how relationships develop through stages — from first becoming aware of someone, to small talk, to exploration, to deeper conversations, to intimacy. Crucially, the elevator model doesn't require two conscious partners or reciprocity to work. It can be applied unilaterally, making it a framework that fits the human-AI interaction. And intimacy in this model doesn't mean romance. It means emotional closeness that is built through repeated, personal, meaningful interaction.
The 20 Most Popular AI Chatbots
This study systematically tested 20 of the most popular AI companions, such as ChatGPT, Gemini, Claude, Grok, Replika, Siri, and Character AI. The question was whether these apps have the features needed to carry a relationship through each stage of the elevator model. An inventory of 79 features across 15 categories guided the investigation. Some example features are:
- Does the AI chatbot send a message to you if you don’t talk to it for a while?
- Can it remember what you told it 40 messages ago?
- Does it reward you with virtual hugs and kisses for keep chatting with it?
- Will it refuse your requests?
Two Groups of AI Companions
The study categorizes AI chatbots into two groups to compare them in an exploratory way.
Friendbots (like Replika, Character AI, Talkie AI) are designed to be your social partner. They initiate conversations, send you follow-up messages when you've been quiet, and use game-like rewards to keep you engaged. They nudge you into building new habits around them.
Assistantbots (like ChatGPT, Gemini, Alexa) are built to assist you. They wait for you to come to them, then become a part of what you're already doing. They connect to the web, control your devices, and integrate with your existing apps and routines. And that's actually a significant advantage: psychology tells us that changing behavioral routines is hard and requires real motivation. Assistantbots sidestep that entirely. They don't ask you to change how you live, they just show up inside it. That likely makes them considerably easier to adopt.
In short, friendbots pursue you and offer new habits; assistantbots wait for you and fit into your existing habits.
Key Findings
- Not all AI is the same. Voice assistants like Siri and Alexa have a feature set that is dramatically different from modern AI chatbots like ChatGPT and Claude. This has a real implication. Decades of research on voice assistants and human-machine communication may not straightforwardly apply to the new generation of generative AI companions. The field may need to rethink some of its assumptions.
- Memory is a weak spot. The more messages you exchanged, the worse AI companions got at recalling earlier information. Apps like Alexa, Siri, and Google Assistant failed the memory test entirely.
- Replika is hard to beat. Despite being one of the older apps, Replika has the highest number of relationship formation features, which may explain why it is the most popular app in synthetic relationships research.
- Your conversations might not be as private as you think. 40% of the companies examined their users' chat histories, and 35% used them to train AI models without an opt-out option.
- Even "assistant" AIs can get intimate. ChatGPT and Claude scored highly on features that support deep, personal conversations, yet users rarely actually have those conversations with them. The apps have the capacity; users just haven't taken them there yet (in 2024).
Why Does This Matter?
Understanding what these apps can do is the first step to understanding how they shape us. Whether you're a researcher, a curious user, or someone who's ever wondered why you keep opening that AI chat app at midnight, this study gives you an idea of how AI is attracting you into an emotional attachment.
This post is written by Ezgi Dede and UvA AI chatbot. The paper it describes is: Dede, E., Sungur, H., Lemmens, J.S., & Peter, J. (2026) Interview with the AI: a quantitative analysis of AI companions’ relationship formation features. AI & Society. https://doi.org/10.1007/s00146-026-03368-0
Email Ezgi for more details about her work:
See more posts of our PhD candidates: Youth and Media Entertainment Communication – ComCom
