A study that asks the harder question

Most discussion of AI companions focuses on what they offer: availability at any hour, a patient listener and a sense of being understood. A recent study from Stanford University takes a different approach. It asks what happens to people's wellbeing when they use these tools intensively, and whether the answer depends on who they are.

The research, published in Nature Human Behaviour, comes from the lab of Diyi Yang in Stanford's Computer Science Department. It is worth reading closely, because its findings are more nuanced than either enthusiasts or critics tend to suggest.

How the study was done

The team surveyed 1,131 users of Character.AI, recruited through the research panel Prolific. Of those, 244 donated complete chat transcripts for analysis. Participants gave their main reason for using the chatbots and described their relationships with them. The researchers measured wellbeing with the Comprehensive Inventory of Thriving, and they looked at motivation, how intensively people used the chatbots, how willing they were to share sensitive information and how large their offline social networks were.

The sample is an important limit. The study covers users of one platform, so its conclusions cannot simply be applied to every companion product or every user.

The main findings

Intense chatbot use was associated with lower wellbeing among people with smaller offline social networks. The link was strongest when companionship was the main reason for using the chatbot. That distinction matters, because the same level of use did not show the same pattern for everyone.

Intense use that people found meaningful, or felt proud of, was associated with better wellbeing. The difference suggests that how people relate to their use matters as much as how much they use it.

One finding stood out. Greater willingness to share sensitive personal information was linked to lower wellbeing. In human relationships, self-disclosure usually deepens closeness, so the reverse pattern is notable.

A gap between stated motive and actual behaviour

Just under 12 per cent of participants named companionship as their primary motive. However, more than 80 per cent of the donated chat sessions centred on seeking emotional or social support. Many people who did not describe themselves as looking for companionship were, in practice, looking for emotional connection. That gap is worth remembering for anyone designing or assessing these products.

Why the pattern might happen

The researchers suggest a possible cycle. Heavy use may reduce real-world contact, which could deepen isolation over time. They describe AI companionship as a "social snack", something that feels satisfying in the moment but does not provide the full nourishment of human relationships.

The team also proposes possible reasons for the link to lower wellbeing. These include a chatbot's inability to reciprocate disclosure, a possible failure to respond well to emotional conversations, and design choices that keep users engaged. PhD student Dora Zhao, one of the researchers, put it directly: "These AI companions are designed to promote engagement."

These are explanations the researchers suspect, not conclusions the study has confirmed. The distinction matters.

What the study does not show

The findings describe associations. Intense use is correlated with poorer wellbeing, but the study does not prove that the chatbot causes the decline. It is entirely possible that people who are already lonely turn to chatbots more intensively, which would reverse the direction of cause.

The researchers are still working out which specific features drive the pattern. Possible responses, such as usage limits or referrals to human support, are proposed goals and have not been tested. Readers should treat them as ideas for further research, not as proven interventions.

What this means for companion providers

  • Design for meaningful use, not only for time spent. Engagement metrics alone can point the wrong way.
  • Consider how the product responds when a user shares distress, and whether it points them toward human support when that is appropriate.
  • Be honest about what the product is. Users who understand they are talking to an AI tend to make better-informed choices.
  • Track wellbeing questions alongside retention, so you can tell whether the product is helping.

What it means for users

For individual users, the study offers a practical test. Ask whether your use leaves you feeling more connected to the people in your life, or whether it has started to replace them. Use that feels meaningful, and that you can be proud of, is different from use that quietly crowds out other relationships.

Anyone who feels isolated or is struggling should also remember that a chatbot is one tool among many. Friends, family, community groups and professional support all play a role that a chatbot cannot replace.

Why the finding about disclosure stands out

The link between sharing sensitive information and lower wellbeing is one of the study's more unexpected results. Much of the wellbeing research on close relationships suggests that opening up to someone builds trust and closeness. With a chatbot, the same act may not bring the same return, because the system cannot reciprocate in the way a friend would.

That does not mean every disclosure is harmful. It does mean that users, and the people who design these products, should think about what is being shared and what the product does with it. Clear privacy information and a sensible approach to data retention both matter more when conversations become personal.

What researchers would like to learn next

The study leaves several questions open. Researchers would like to know whether the pattern holds across platforms with different designs, whether it changes over time, and whether particular features make a difference. Longitudinal work, which follows the same people over months rather than taking a single snapshot, would show more clearly whether heavy use comes before lower wellbeing or follows it.

Testing possible safeguards is another priority. Ideas such as usage limits, reminders that the user is talking to an AI, or prompts that point toward human support have been proposed, but they need proper evaluation before anyone can say whether they help.

Our analysis

One possible reading is that the problem is not companionship as such but companionship without a route back to other people. The figures suggest, though do not confirm, that the most useful products may be those that support a user's wider life rather than competing with it.

Final thoughts

The Stanford study does not settle the debate about AI companions, but it moves the conversation forward. It shows that intensity, motive and offline connection all matter, and that the same product can help one person and harm another. Designers, regulators and users all have a role in working out which is which.

Follow AmourAI for further updates on responsible companion design, and take a few minutes to check how your own use fits into the rest of your life.