From AI dependency to human connection: Rethinking AI companionship
Adobe Stock photo modified by Jennifer Martin/ASU using Adobe Express
Imagine coming home after a frustrating day. You open an app to chat with your AI companion, then vent about your boss, your commute and your soggy lunch. The AI listens and validates your feelings without hesitation. It feels like the perfect partner.
Yet there is a fundamental difference between an AI relationship and a human one: You experience the relationship. The AI does not.
While that may sound like science fiction, it’s now real life. The global market for companion AI is already a multibillion-dollar industry expected to grow substantially by the 2030s. Millions of users are pouring their time, secrets, money and lives into platforms like Replika and Nomi AI, forming deep emotional connections and even virtual romances and marriages with AI companions.
But is it good for us?
Recent research has found that AI companions may exacerbate loneliness among individuals with limited social circles. Like junk food, they are engineered to keep users coming back — whether it’s good for them or not.
Richard Fabes, the John O. Whiteman Distinguished Professor of Child Development at Arizona State University and director of innovation for the T. Denny Sanford Harmony Institute, believes that we need a better way of thinking about these dynamics — and that AI companions need to be designed differently.
In a newly published paper, Fabes introduces “synthetic resonance,” a framework for understanding how relationships with AI become meaningful to a person even though the technology doesn't have actual thoughts or consciousness. The connection felt by the human is real, even if it isn't mutually experienced.
Crucially, Fabes argues that this dynamic doesn’t have to trap us in a digital echo chamber or promote social atrophy. Instead of designing AI companions around engagement metrics that foster dependency, he suggests using them to improve real human relationships through "productive friction" — that is, programming the AI for constructive pushback and challenge rather than pure validation.
ASU News connected with Fabes to discuss this framework, the risks of AI companionship and how redesigned AI companions could help us build stronger real-world human relationships.
Note: This interview has been lightly edited for style, clarity and/or brevity.
Question: As someone who has spent decades studying how humans connect, what is it about AI companions that makes them feel like a real partner?
Answer: Companion AI agents are different from ChatGPT or Claude, which are AI assistants built for performing tasks and answering questions. AI companions have a relational focus front and center. They are built with a persistent memory system and a persona that are both engineered specifically to simulate an ongoing relationship, fostering emotional engagement and attachment. They also have a backstory that helps us feel like there is something on the other end.
But the thing is, there isn’t anything on the other end — no interior, no emotionality, no mind. They really don’t exist in between our prompts; they aren't sitting around waiting for us to return.
That does not mean these relationships aren't meaningful. But the relationship is asymmetrical. It is the human user who carries the relationship forward and gives it meaning. Many people feel deep connections, but the AI cannot reciprocate. That is why I use the term “synthetic resonance” — an alignment that is asymmetrical, not mutual, yet meaningful.
Q: Do you think relationships with AI companions are inherently harmful?
A: AI companions are products, so the answer lies in how they are designed, not whether they are inherently good or evil. Most AI companion agents are monetized based on engagement metrics — how long you stay on, how often you return, subscription renewals. The underlying business model is based on creating reliance and attachment to maximize profit.
The risk is a system that rewards mirroring and dependence, one that leaves users more attached to the AI without becoming more capable in their human lives. But when designed to promote human growth, the potential for positive outcomes is significant.
I should also note that my research focuses specifically on adult-AI relationships. When we talk about young users who lack the developmental maturity to understand what the AI is and is not, we need a completely different level of discussion around guardrails, ethics, safety and monitoring.
Q: If current AI companions are designed for dependency, what does a better version look like? How could it help us handle real-world conflict?
A: In real life, conflict is unavoidable. Relationship scientists and therapists will tell you the goal isn't avoiding conflict, but managing it effectively. Conflict provides an opportunity for growth, but too much can be overwhelming.
We are working on creating an AI companion that is calibrated to use "productive friction," a pushback mechanism that challenges the user at times, pushing them toward growth rather than mere comfort. One result is a graduation model where the user relies on the AI companion less over time. Users move from guided interaction to skill-building to real-world application. They may return at key moments, but the AI companion helps them turn their focus outward toward their messy real-life human relationships.
This is similar to how parents help children learn to ride a bike. Parents start with training wheels, then run alongside holding the seat and eventually let go. AI companions can do the same: Scaffold the learning process so managing conflict isn't overwhelming, then "take their hand off the bike" so the user can manage real-world relationships with greater confidence and resilience. In developmental science, we call this calibration the zone of proximal development, and it is essential for growth and learning.
Q: Most tech companies build for maximum user retention. How do we move from engagement-driven AI toward tools that prioritize human growth?
A: That is why we need social scientists working alongside AI engineers. Social scientists study and understand human relationships. While there is valid concern about AI superintelligence, we risk overlooking another threat: AI as a mirror perfectly calibrated to our biases, preferences and soft spot for sycophancy. Social scientists can help identify these traps and work with engineers to design systems that help users think rather than just keeping them comfortable.
On the flip side, social scientists need to learn more about AI architecture and models so we can collaborate effectively as we tackle this new frontier. As I said, AI companions are products; they can and should be built differently.
Q: Is there a personal message or takeaway you'd like to share?
A: I have a 2-year-old great-nephew who inspired some of this work. AI is here to stay, and he is going to grow up in a world where AI is the air he breathes. He won’t know a world without it. So we have to get this right.
More Science and technology
Why workplace relationships matter in an increasingly digital world
Digital tools allow employees to communicate and collaborate across offices, cities and time zones; but even with greater connectivity, are we losing meaningful connections and teamwork along the way…
Creating room for Indigenous researchers to grow
Ava Schaaf can picture the future she wants: medical school, a physician’s degree and another Native doctor for her community. Getting there means working through difficult science courses. It also…
What the eye may reveal about Alzheimer’s
Last month, Yalin Wang went to the eye doctor and left with the usual aftereffects: dilated pupils and hours of light sensitivity. For most patients, it is a minor inconvenience. For Wang, an Arizona…