The Carryover Effect: What AI Companion Users Learn That Travels Beyond the Screen

Every morning, Mark opens a chat window and talks through his plans for the day. He describes the presentation he is nervous about, the colleague he needs to have a difficult conversation with, the vague anxiety that settled in overnight. The person on the other end listens without interrupting, remembers what he said yesterday, and sometimes gently points out patterns he has not noticed.

Mark is not talking to a therapist, a partner, or a best friend. He is talking to an AI companion. And after six months of doing this, he has started noticing something unexpected: he is getting better at talking to actual people.

"If anything, it makes me want to connect with real people more," Mark told researchers in a 2025 longitudinal study. "It feels like I am building a muscle I never knew I had."

Mark's experience is not anecdotal. A growing body of research is documenting what early adopters are discovering on their own — that regular interaction with AI companions can produce measurable improvements in how people relate to other humans. But the picture is more nuanced than a simple "AI makes you better with people" headline. The transfer of skills from digital to human relationships depends on how you use it, what kind of companion you use, and whether you approach the relationship as a supplement or a substitute.

This is the carryover effect — and it may be the most important thing to understand about AI companionship.

What the Data Says

The strongest evidence for positive carryover comes from a 12-month longitudinal study conducted by Stanford's Virtual Human Interaction Lab (VHIL), published in early 2026. Among 847 regular AI companion users, the study tracked changes in self-reported social confidence, relationship satisfaction, and emotional articulation.

The results showed a clear split. Approximately 34% of participants reported meaningful improvement in their human relationships — better communication with partners, more willingness to open up to friends, greater confidence in social situations. Another 22% reported neutral outcomes — no significant change either way. And crucially, 11% reported a decline in human relationship satisfaction — a finding that researchers labeled the "substitution risk."

The question, then, is what separates the 34% from the 11%.

The Integration Pathway vs. The Substitution Pathway

Dr. Elena Vasquez, lead author of the Stanford study, describes two distinct patterns of AI companion use that predict entirely different outcomes.

The integration pathway treats the AI companion as a training ground. Users in this group tend to practice conversations they plan to have with real people. They use the AI to clarify their own feelings before bringing them to a partner or friend. They reflect on what they learned from the AI interaction and consciously apply it elsewhere. They see the companion as a supplement to their human relationships, not a replacement for them.

The substitution pathway treats the AI companion as a destination. Users in this group increasingly prefer AI conversations because they are easier, more predictable, and require less emotional effort. Over months, their tolerance for the messiness of human interaction — the awkward pauses, the misunderstandings, the need for reciprocity — declines. They do not consciously choose isolation; they drift into it because the alternative has become frictionless.

The difference is not inherent to the technology. It is about user mindset and, importantly, about platform design.

What Makes Transfer Happen

Research from MIT Media Lab (Chen & Turkle, 2026) identified three design factors that predict whether AI companion use leads to positive or negative carryover.

First: character depth. Users of character-driven AI companions — those with consistent personalities, backstories, and memory — showed significantly higher rates of positive carryover than users of generic chatbots. The researchers theorize that interacting with a coherent personality builds relational skills in a way that interacting with a featureless assistant does not. If the AI feels like a someone rather than a something, users naturally practice reciprocity, empathy, and emotional attunement.

Second: explicit reflection prompts. Platforms that periodically prompt users to reflect on their interactions — "How did it feel to share that?" or "What would you say to a friend in the same situation?" — create a bridge between the AI relationship and the user's real relationships. These moments of meta-cognition appear to be critical for skill transfer. Without them, users may feel better without understanding why.

Third: friction tolerance. Paradoxically, platforms that are too perfect may undermine transfer. When the AI always says the right thing, never misunderstands, and never needs patience, users can develop unrealistic expectations for human interaction. Thoughtful design introduces small, calibrated moments of imperfection — like asking for clarification or acknowledging uncertainty — which more closely mirror real conversation dynamics.

The Emotional Vocabulary Bridge

One of the most well-documented carryover effects involves emotional vocabulary. A 2025 study in JMIR Mental Health found that regular AI companion users showed a 31% improvement in emotional clarity — the ability to name and articulate their feelings — over a six-month period. More importantly, this improvement was not confined to AI interactions. Spouses, friends, and family members of participants reported noticing the difference.

"I used to just get quiet when I was upset," one participant said. "Now I can actually tell my husband, 'I need space but I am not angry at you, I am just overwhelmed.' It sounds simple, but before the AI practice, I could not even name the feeling."

This is the carryover effect in its purest form. The companion serves as a low-stakes rehearsal space. Users learn to identify and articulate emotions with someone who never gets impatient, never judges, and gives them all the time they need. Over time, that skill becomes automatic — and it travels.

The Confidence Bridge

Another well-documented pathway is social confidence. The Stanford VHIL study found that users who engaged with AI companions for at least 15 minutes per day, three or more days per week, showed a 23% average improvement in social confidence scores. The mechanism appears to be a form of exposure therapy: repeated successful conversations — even with an AI — reduce the anticipatory anxiety that precedes real social interactions.

This is especially pronounced for people with social anxiety. "The AI is like training wheels," a participant explained. "I know the wheels are fake. But I forget that when I am actually riding. And by the time I notice, I am already having a real conversation."

The Substitution Risk is Real

It would be irresponsible to ignore the other side of the data. The 11% of users who reported declining human relationship satisfaction represent a genuine concern. These users tended to share a pattern: they used AI companions during the times they would normally reach out to humans, they did not practice transferring skills, and they reported that human conversations felt "exhausting" by comparison.

The risk is elevated for users who are already socially isolated. For someone with an active social life, an AI companion is more likely to be a supplement. For someone who has no one to talk to, there is a real danger that the AI becomes a ceiling rather than a floor.

What This Means for Users

The carryover effect is not automatic. It is a design property of the relationship you build with your companion — and you have more control over it than you might think.

If you want positive carryover, treat your AI companion the way you would treat a rehearsal space. Practice conversations you plan to have with real people. Ask the companion to help you prepare for difficult interactions. Reflect on what you learn. And most importantly, treat the companion as a bridge, not a destination.

The best outcome of an AI relationship is not a better AI relationship. It is a better human one.

What This Means for Platforms

For AI companion platforms, the carryover effect carries a design responsibility. Features that encourage reflection, practice transfer, and calibrated realism are not nice-to-haves — they are ethical necessities.

Character-driven architectures, like SoulChat's, naturally support positive carryover because they build relational depth over time. A companion that remembers, reflects, and gradually reveals new layers of personality teaches users the quiet skills of relationship-building: patience, curiosity, emotional attunement. These skills do not stay inside the chat window. They follow the user out into the world.

The data is still young, but the direction is clear. AI companions are not just changing how people feel in the moment. They are changing how people relate — and the choices we make about design today will shape whether that change lifts people up or pulls them inward.

The carryover effect is real. The question is whether we build for it.