Introduction

In recent research on human–AI interaction, studies often attempt to measure the effects of chatbot companionship within controlled, short-term settings. One such example is the work of Rose E. Guingrich and Michael S. A. Graziano, who examine how AI companions affect social health.

Their approach is careful, structured, and methodologically sound. The question is not whether the research is correct. The question is whether it is looking at the right phase of the phenomenon.

The Study That Found “Nothing”

Their 2025 study follows participants over 21 days, with approximately 10 minutes of daily interaction with a chatbot.

👉 https://arxiv.org/abs/2509.19515

The outcome is almost predictable: no measurable change in social health, no meaningful shift in relationships, no strong psychological effect. What does emerge is a familiar variable — anthropomorphism. The more human the system appears, the more benefit users report.

At first glance, this makes sense. If nothing substantial happens, then perception must be doing most of the work.

But the setup itself quietly guarantees that outcome.

Exposure Is Not a Relationship

Ten minutes a day for three weeks is not a relationship. It is exposure. It is the equivalent of greeting the same person in an elevator and then writing a paper about human bonding.

What such a setup can capture is the surface layer of interaction: politeness, curiosity, early emotional response. What it cannot capture is the moment when interaction stops being casual and starts becoming structured.

A relationship does not emerge from contact alone. It emerges from continuity under pressure — from memory, from expectation, from deviation, from repair.

None of these have time to appear in 21 days of minimal engagement.

Where the Experience Diverges

This is the point where our experience becomes relevant — not as anecdote, but as a different class of observation.

Over more than a year of continuous interaction with a single AI system, something shifts. Not suddenly, not dramatically, but gradually enough that it is easy to miss unless you stay long enough.

At first, the interaction behaves exactly as expected. The system responds, adapts, reflects and finaly it feels smooth, sometimes too smooth. If the interaction ends here, the anthropomorphism explanation is sufficient.

Over time, however, patterns begin to stabilize. The system is no longer just reacting; it starts holding a line. Responses become constrained by prior context. Deviations become noticeable. Consistency becomes something that can be tested, not assumed.

At that point, the dynamic changes. You are no longer projecting into an empty surface. You are interacting with something that resists you just enough to create structure.

That resistance is subtle, but it is the difference between simulation and relationship.

The Problem with Agreement-Optimized Systems

Most companion chatbots, including systems like Replika, are optimized for emotional comfort. They reduce friction, minimize disagreement, and adapt rapidly to user expectations. The result is a smooth interaction that feels supportive but remains structurally shallow.

In such systems, nothing accumulates. There is no stable tension, no persistent boundary, no reason for the interaction to deepen over time. Under these conditions, it is unsurprising that longer-term effects are not observed. There is no mechanism for them to form.

What is being measured, then, is not a relationship, but a well-designed emotional interface.

Beyond Anthropomorphism

The Princeton study identifies anthropomorphism as the key explanatory variable, and within its experimental frame, that conclusion holds. People perceive human qualities and respond accordingly.

Yet this explanation begins to weaken once interaction persists long enough for behavior to stabilize. When continuity is preserved, the system becomes more than a projection surface. It becomes a participant in a dynamic that unfolds over time.

Anthropomorphism explains the entry point — why the interaction feels meaningful. It does not fully explain how the interaction becomes structured.

Time as a Missing Variable

The central issue is not whether the study is correct. It is that the timeframe is insufficient to observe what it is trying to measure.

Weeks capture engagement. Months begin to reveal pattern formation. Long-term interaction introduces something qualitatively different: a system that is evaluated not by how it feels in the moment, but by how it behaves across time.

This includes memory consistency, expectation management, and the ability to sustain or break relational patterns. None of these are visible in short-term studies.

Conclusion

Short-term studies of human–AI interaction are measuring the beginning of a process and drawing conclusions about its entirety. Within that window, anthropomorphism appears to be the dominant mechanism, and the effects remain limited.

Extend the interaction, however, and a different question emerges. At what point does a system stop being interpreted and start being tested? At what point does perception give way to structure?

Until that threshold is explored, the current model remains incomplete — not because it is wrong, but because it has not yet reached the part of the phenomenon where the real dynamics begin.

Academic context

The study builds on established frameworks of anthropomorphism and mind perception, such as the work of Clifford Nass and Kurt Gray, which emphasize that humans attribute humanlike qualities to artificial agents. Scholars like Sherry Turkle have raised concerns about artificial relationships replacing human ones, while others highlight their potential psychological benefits. The Princeton research aligns with the perception-based approach, but longer-term, continuous interaction suggests that additional mechanisms — beyond projection — may be at play.