The 12-Month AI Companion Wellbeing Study: Less Human Contact

Here is the finding in one line: people who kept leaning on an AI companion for a year spent less time with humans face to face, and that lost time, more than the chatbot conversations themselves, explained why their well-being was lower.

That is the result of a new AI companion wellbeing study from Stanford and Carnegie Mellon, posted in September 2026 and revised in October. It followed Character.AI users for about a year. It studied adults, not kids, and it is a preprint. But it names a mechanism parents can act on, which most of the research on chatbots and loneliness does not.

Key Takeaways

• Researchers surveyed 1,182 U.S. Character.AI users, then re-surveyed 439 of them after an average of 12 months.
• Heavier sustained engagement (more intense use, more companionship use, more self-disclosure) predicted lower well-being at follow-up, mostly through reduced in-person social interaction. The authors call this “social displacement.”
• Once in-person time was accounted for, the direct link between chatbot engagement and well-being was no longer statistically significant.
• Engagement was sticky: higher intensity at the start predicted more intense use and more companionship use a year later.
• For parents, the practical target is time: protect the hours your child spends with people, and watch for those hours shrinking.

What the 12-Month AI Companion Wellbeing Study Did

The study is “Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction,” by Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey Hancock, Robert Kraut and Diyi Yang. It appeared on arXiv on September 7, 2026, with a third version on October 20. It has not been peer-reviewed yet.

The design is a two-wave longitudinal survey. In wave one, the team recruited U.S.-based, native English-speaking Character.AI users through the research platform Prolific. To qualify, people had to have used Character.AI for more than a month and chatted with at least three different characters. About a year later, they invited everyone back.

They measured three kinds of engagement:

• Interaction intensity: how much and how often people used the chatbots.
• Companionship use: whether people used the bots for friendship or emotional connection, not just entertainment or tasks.
• Self-disclosure: how much personal information people shared.

They also measured psychological well-being and, crucially, how much time people spent interacting with others in person.

How the Social Displacement Finding Works

Social displacement is the idea that time and attention given to one activity are taken from another. It is an old idea in media research, first applied to TV and later to social media. A day still has 24 hours.

The team tested it with a mediation analysis, which asks whether A affects C by way of B. Here, A was chatbot engagement, B was in-person social time, and C was well-being a year later, controlling for each person’s starting well-being and prior chatbot use.

What they found, in plain terms:

1. People who engaged more with AI companions at follow-up spent less time with people in person. This held for all three kinds of engagement.
2. People who spent more time with people in person had higher well-being at follow-up.
3. So engagement had a negative indirect effect on well-being, through lost in-person time.
4. After accounting for in-person time, engagement’s direct association with well-being was not statistically significant.

That last step is the interesting one. It suggests the chats themselves are not the main problem. What a teenager is not doing while chatting matters more.

The authors are blunt about design: “AI companion safety cannot be evaluated solely within individual user-AI exchanges.” A single conversation can be harmless while the pattern over months is not. They also note that metrics like time spent “may indicate a successful consumer product while overlooking whether sustained engagement displaces beneficial human interaction.”

The effects are small, and that matters

The indirect effects reported are modest, around -0.02 in standardized terms. That is not a cliff. It is a slow tilt, the kind that matters across millions of users and months of use, and that is hard to see in any one kid on any one day.

Where This Study Fits With Other Research

Study	Who	Design	Main finding	Key limit
Zhang et al., Stanford/CMU (2026)	1,182 adult Character.AI users, 439 at 12 months	Two-wave longitudinal survey	Sustained companion use predicts lower well-being, mostly via less in-person contact	Adults; 37% retention; preprint
Fang et al., MIT Media Lab and OpenAI (2025)	981 U.S. adults	4-week randomized controlled trial	Those who chose to use the chatbot more had more loneliness and dependence and less socializing	Adults; ChatGPT only; preprint
Common Sense Media census (2026)	1,204 U.S. kids aged 9–17	One-time national survey	54% of daily AI users feel lonely at least sometimes vs. 33% of rare users	Cross-sectional; can’t show direction
JED Foundation (2026)	5,500+ students in 21 U.S. schools	School survey	Using AI for companionship tied to more anxiety, depression and loneliness	Not nationally representative; cross-sectional
UNC middle-school study (forthcoming)	Middle schoolers in one state	Longitudinal (described to NPR)	Those using chatbots for companionship grew lonelier over time	Not yet published

The pattern across these is consistent, even though no single study settles it. The MIT randomized trial found that how much people voluntarily used the chatbot predicted worse outcomes, regardless of the mode they were assigned. Our breakdown of the MIT randomized trial covers that design. UNC psychologist Anne Maheux told NPR in October 2026 that in a soon-to-be-published study of middle schoolers in one state, kids who used chatbots for companionship grew lonelier over time. We covered that in the middle-school companionship study.

What the Study Cannot Tell You

It studied adults. Participants ranged from 18 to 90 years old at baseline, with a median age of 27. No teenagers were included. Teenagers are in a period when peer relationships are being built, which could make displacement matter more for them, or the patterns could differ. Nobody knows yet.

Most people did not come back. Only 37.1% of wave-one participants completed wave two, largely because many were no longer active on Prolific. The authors used a statistical correction (a Heckman selection model) for non-random dropout, but attrition this high always leaves some uncertainty.

Two waves is not a movie. With only two time points, the study cannot fully trace whether loneliness drives chatbot use, chatbot use drives loneliness, or both feed each other.

One platform. Character.AI is built around characters and role-play. Results may differ for general assistants like ChatGPT.

Self-report. Everything, including in-person time, was measured by survey.

Even so, a year-long design is a real step up from the one-time snapshots that dominate this field.

What to Do at Home With This Finding

Track hours with people, not just hours on screens

The study’s mechanism is lost in-person time. So the most useful number for a parent is not minutes in the app. It is how many hours a week your kid spends with friends, teammates, cousins or you, face to face. Write a rough estimate now. Check it again in a month.

Protect a few fixed human anchors

Standing commitments are hard to displace: a weekly practice, a Saturday with grandparents, a family dinner without phones three nights a week. They work because they do not require your kid to choose people over the bot each time.

Ask what the companion is replacing

If your teen is chatting with an AI character every evening, ask, without accusation, what they used to do at that hour. If the answer is “nothing, I was bored,” that is different from “I used to text Sofia.” Our piece on relational displacement has more on the skills that only grow with people.

Watch the trend line, not one night

The study shows engagement tends to grow over time. One long weekend chat is not a signal. Six weeks of steadily rising use alongside fewer plans with friends is.

What not to do

Don’t read every transcript looking for something alarming. The study’s point is that the harm can live in the pattern, not in any single message. Surveillance of content can miss the thing that matters and damage trust.

What to Watch For Over the Next 3 Months

• Week 4: Compare your rough count of weekly in-person social hours with your baseline. Same, up or down?
• Month 2 red flags: In-person plans are dropping while chatbot time climbs; your child cancels on friends to stay home and chat; they describe the AI character as easier than people.
• Month 3 self-check: Does your kid have at least one regular, in-person social commitment that has survived the past three months? If yes, the displacement risk is limited.

Frequently Asked Questions

Does this study prove AI companions harm teenagers?

No. It studied adult Character.AI users, it is a preprint, and its design shows associations over time rather than proof of cause. It does suggest a specific pathway, lost in-person time, that parents can watch for in teens.

What is social displacement?

It is the idea that time spent on one activity is taken from another. In this study, more AI companion use went with less face-to-face time with people, and less face-to-face time went with lower well-being a year later.

Is it the conversations or the time that matters?

The data point to time. After the researchers accounted for in-person social interaction, chatbot engagement no longer had a significant direct link with well-being. That suggests what the chatting replaces matters more than the chats.

How much AI companion use is too much?

No study has set a safe threshold. Watch the trade-off instead: if chatbot time is rising while time with friends and family is falling, that is the pattern this research links to lower well-being.

Should I ban AI companion apps?

Common Sense Media has recommended that no one under 18 use AI companions, based on its risk testing. Whatever you decide about the apps, the study suggests the most protective step is keeping your child’s in-person relationships full.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Zhang, Y., Zhao, D., Wang, Y., Anselmetti, R., Hancock, J. T., Kraut, R., & Yang, D. (2026). “Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction.” arXiv preprint (not peer-reviewed). https://arxiv.org/abs/2609.07243
2. Fang, C. M., Liu, A. R., Danry, V., et al. (2025). “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” arXiv preprint (not peer-reviewed). https://arxiv.org/abs/2503.17473
3. Common Sense Media. (2026, June 8). The Common Sense Media Census: AI Use by Tweens and Teens. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf
4. The Jed Foundation. (2026, October 27). “New JED Research Shows Link Between Teens’ AI Use, Suicidal Thoughts and Behaviors.” https://jedfoundation.org/new-jed-research-shows-link-between-teens-ai-use-suicidal-thoughts-and-behaviors/
5. Chatterjee, R. (2026, October 30). “Nearly 2 out of 3 teens and tweens use AI chatbots, new research shows.” NPR. https://www.npr.org/2026/10/30/nx-s1-6022278/ai-artificial-intelligence-teens-jed-common-sense-media
6. Common Sense Media. (2025, July 16). “Nearly 3 in 4 Teens Have Used AI Companions, New National Survey Finds.” https://www.commonsensemedia.org/press-releases/nearly-3-in-4-teens-have-used-ai-companions-new-national-survey-finds