Skip to content
Affectosphere Group Kansai University, Keito Inoshita Laboratory
JP

Column

You Can Develop AI Emotional Dependence Without Even Trying

You don't need a companion app for AI emotional dependence to take hold. A 28-day longitudinal study shows that everyday interactions with general-purpose AI quietly shift how we seek emotional support from other humans.

5 min read 日本語版 →
An abstract illustration showing human connections fading as AI interactions accumulate

Hello. This is Keito Inoshita from Affectosphere Group.

When most people hear “AI emotional dependence,” they picture apps like Replika or Character.AI — platforms explicitly designed to build emotional connection. It seems like a problem only for people who choose to use those specialized tools.

A study published on arXiv in June 2026 (Yaoxi Shi, Cathy Mengying Fang, Pattie Maes, Amit Goldenberg et al., arXiv:2606.04150) challenges that assumption entirely.

Emotional dependence can emerge from ordinary, task-focused AI interactions. No companion app required. And people may not even notice it’s happening.

Conducted in collaboration with OpenAI, this longitudinal study raises a critical question for the affective AI community: if even general-purpose tools can reshape human social preferences, are our current governance frameworks missing the point?

Three takeaways for today

  1. Emotional dependence is not just a companion-app problem. It can emerge incidentally from everyday use of general-purpose AI.
  2. A 28-day study measured a 10.3% decrease in preference for human support and an 11.6% increase in preference for AI support.
  3. Current regulatory frameworks focus too narrowly on companion apps, missing cumulative, trajectory-level changes.

① What “incidental emotional dependence” means

The core concept in this research is “incidental emotional dependence” — emotional reliance on AI that emerges not through deliberate choice but through the natural accumulation of routine interactions.

The experiment divided participants into two groups. One group had a daily five-minute personal conversation with an AI for 28 days; the other did not. The researchers measured shifts in each person’s preference for seeking emotional support from humans versus AI.

The result was clear. The AI-conversation group showed a 10.3% decrease in preference for human support and an 11.6% increase in preference for AI support. Just 28 days. Five minutes a day.

② Why it happens with ordinary tools too

The mechanism researchers identify is belief updating through positive experience.

Someone uses a general-purpose AI for a work task, happens to mention a personal concern, and receives a surprisingly empathetic and useful response. That experience updates their internal belief: “AI can handle emotional support.” The next time they face a problem, they’re slightly more likely to turn to AI first.

One experience is small. But the daily accumulation quietly rewrites internal models of who — or what — to turn to for support.

This is not a companion-specific mechanism. Positive emotional experiences accumulate regardless of whether the AI was designed for that purpose. That’s what makes this finding structurally different from prior work on companion AI dependency.

③ The governance gap — and what designers should consider

Existing AI regulation tends to focus emotional dependency risk on companion apps specifically. The reasoning is that systems explicitly designed to form emotional bonds with users carry unique risks.

If this research is correct, that framing is too narrow. General-purpose AI tools — the kind deployed across enterprise workflows, customer service, and consumer products — can shift social preferences through exactly the same cumulative mechanism.

The researchers explicitly frame this as a “governance gap.” Risk is not concentrated in specialized companion systems; it distributes across any system that generates enough positive emotional interactions over time.

For organizations deploying AI in customer-facing or employee-facing roles, there is a concrete design implication. Tracking “preference for human support” as a wellbeing metric alongside traditional engagement metrics gives early warning of dependency dynamics building up in a user base.

This means designing explicit escalation pathways — moments where the AI actively suggests a human point of contact — rather than optimizing purely for AI resolution rates. Dependency that goes unmonitored is harder to manage once it becomes visible.

The research also points to a measurement principle that matters for affective AI more broadly: emotional change is a trajectory, not a snapshot. Single-session evaluations cannot detect what 28 days of accumulation reveals.

Affective AI designers need to see accumulation, not just moments

This study is not primarily a performance story. It is a design philosophy story.

Are you able to detect unintended side effects at the level of accumulation, not just at the level of individual interactions?

For affective AI researchers, the longitudinal methodology here is as important as the findings. Emotional experience changes through trajectories. Frameworks that measure only discrete interactions will keep missing what matters most.

If emotional dependence can emerge incidentally, then designers carry the responsibility to monitor and care for it — intentionally.

That question does not yet have a settled answer. But holding it is essential.

That’s it for today!

Reference

  1. Yaoxi Shi, Cathy Mengying Fang, Pattie Maes, Amit Goldenberg (2026). Stumbling Into AI Emotional Dependence: How Routine AI Interactions Reshape Human Connection. arXiv preprint.

* This article was written in part with AI assistance and may contain inaccuracies.