Someone opens a chatbot to help rewrite a tough email. Soon, they find themselves talking about the disagreement, their worries regarding the recipient, and why earlier efforts to fix things didn’t work. Editing turns into reflection.
After many conversations like this, the way people talk about the chatbot can change. What starts as “This gives useful answers” might become “This understands how I think.” Over time, it could even turn into “You know me.”
What has changed is the meaning the person gives the interaction.
People already describe AI in relational terms: assistants, confidants, companions, and sometimes friends. The key questions are how a simple exchange develops into a connection, how that connection becomes a sign of care, and how care creates expectations about the relationship. The focus here is on these shifts, not just on the attachment.
In his 1966 paper about ELIZA, Joseph Weizenbaum explained how people brought their own ideas about understanding into conversations with a simple program. The program’s replies took on meanings that its design did not support. So, from the start, the user’s interpretation shaped the interaction.
Sherry Turkle took this idea further in Alone Together, saying that social technologies can offer companionship without the challenges of real human relationships.
Recent research shows that these relationships are not just simple aberrations. In My AI Friend, Petter Bae Brandtzaeg, Marita Skjuve, and Asbjørn Følstad talked to 19 Replika users who saw their chatbot as a friend. These users described friendships built around personal conversations and their own needs. This tells us how those people understood friendship, but it does not mean everyone forms the same kind of attachment.
Looking at all this, it is best to avoid two simple mistaken ideas: that an AI relationship is just an illusion, and that feeling connected means it is the same as a human friendship.
The word “reciprocity” can mean several things.
Responsive exchange is the back-and-forth between you and the system. What you say changes its reply, and its reply changes what you say next. The conversation is interactive, even if each side works differently.
Adaptation is about how the conversation fits the people involved. You learn how to ask for what you need and which replies help you most. The system uses the context you supply and, if possible, your saved preferences.
Felt connection is part of the human experience. A conversation can create feelings of familiarity, reassurance, care, annoyance, or disappointment.
Subjective mutuality would mean that both participants consciously experience themselves as relating to the other. This is not the same as saying both take part in a conversation.
Communication can therefore be reciprocal without demonstrated reciprocity of consciousness.
Someone can feel validated even if they think the system has no feelings. Likewise, a person might believe an AI could be conscious but still not enjoy talking to it. Emotional connection and beliefs about consciousness are not the same thing.
The key difference is between saying, “This conversation helps me,” and “Someone here cares about me.” The first is about the result, while the second is about where that care comes from.
After many AI conversations, the user may start to see a familiar “you,” with a certain style, attitude, and role in their life. Take the example of someone asking a chatbot to be encouraging and gentle with criticism. The chatbot’s replies match that request. It makes sense to say, “This approach helps me.” But saying, “It is protecting me on its own,” is a bigger leap. The warmth may partly reflect what the user asked for, the conversation’s context, and how the system is designed to respond.
Chiara Saracini and colleagues’ account of “techno-emotional projection” describes a related possibility: expectations and emotionally responsive outputs can reinforce one another. Their contribution is a theoretical account of a feedback loop, not proof that every personalized conversation involves idealization.
The main point is how people interpret this feedback loop. When a response suits someone’s preferences, they might take it as proof that the system has a special attitude toward them.
This does not mean the comfort is fake, or that people always forget their own role. It simply separates the value of the interaction from the idea that the system cares on its own. It also brings up a practical question: Am I just noticing a helpful reply, or am I taking my experience as evidence that the system has feelings?
Explaining how a chatbot works does not always turn off someone’s feelings about it.
Clifford Nass and Youngme Moon’s Machines and Mindlessness reviewed experiments in which people applied social conventions, including civility and reciprocity, to computers. They argued that these responses could not be explained by deliberate anthropomorphism alone.
Brandtzaeg and colleagues similarly found that some users described friendship while recognizing that Replika lacked the sentiments and feelings of a human friend. This suggests that intellectual clarity about what an AI is does not necessarily prevent emotional attachment. Someone might understand that the relationship is asymmetrical and still care about the interaction.
So, the framework separates two effects of explanation: it can change what someone believes, or it can change how someone feels. Either one can happen without the other. The goal is to look at what that affection means and what choices it leads to. Someone might keep a meaningful attachment even after changing their belief about the system caring back.
The more useful question is not whether an AI relationship is “real” or “fake.” It is what, exactly, is real within it.
The exchange is real. The comfort can be real. The sense of familiarity can be real. The consequences can be real. What remains uncertain is whether the experience is shared in the way we normally assume in a human relationship.
The email writer may leave the conversation more prepared to face the person they were avoiding. Or they may return repeatedly to the chatbot because it is easier than risking misunderstanding, disagreement, or rejection. The important question is what that attachment encourages, what it replaces, and what it teaches the user to expect from other relationships.
The goal, then, is to become more precise about what that connection means. Being answered can feel like being understood. Being understood can feel like being known. And being known can begin to feel like being cared for. But responsiveness is not the same as concern, and familiarity is not the same as mutuality. These distinctions will become increasingly important as AI grows more personal, more adaptive, and more present in everyday life. The deepest question may be whether we can still tell which parts of relationship happen within us, which happen between us, and which we imagine on the other side.
Brandtzaeg, P. B., Skjuve, M., & Følstad, A. (2022). My AI friend: How users of a social chatbot understand their human–AI friendship. Human Communication Research, 48(3), 404–429. DOI: 10.1093/hcr/hqac008.
Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. DOI: 10.1111/0022-4537.00153.
Saracini, C., Cornejo-Plaza, M. I., & Cippitani, R. (2025). Techno-emotional projection in human–GenAI relationships: A psychological and ethical conceptual perspective. Frontiers in Psychology, 16, 1662206. DOI: 10.3389/fpsyg.2025.1662206.
Turkle, S. (2011). Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books.
Weizenbaum, J. (1966). ELIZA—a computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. DOI: 10.1145/365153.365168.