Can AI companionship ease loneliness without replacing connection?

There are at least two reasonable questions behind the hope that AI companionship might help loneliness. Does talking to an artificial companion make a person feel less alone in the moment? And does relying on it leave that person better connected to other people over time? An encouraging answer to one does not settle the other.

A service could provide immediate comfort, become emotionally important and still have uncertain consequences for the user’s wider relationships. Equally, people already feeling isolated might use it more, creating an association that is easy to misread as cause. Separating those possibilities makes the evidence more useful than a simple contest between “AI friends are good” and “AI friends are bad”.

A helpful conversation is a real outcome

The final Journal of Consumer Research article AI Companions Reduce Loneliness reports five studies. Its evidence includes experiments finding momentary reductions in loneliness and a week-long study in which participants experienced reductions after using a companion repeatedly. The authors identify feeling heard as an important mechanism, distinguishing it from distraction alone. The paper first appeared online in June 2025 and belongs to the April 2026 issue.[1]

Feeling heard can matter to someone even when a study does not establish a permanent improvement. But a measurement immediately after conversation is not a direct count of friendships, shared activities or future resilience. The scope of the result is part of its meaning, not a reason to dismiss it.[1]

A year-long study asks a different question

Folk and Dunn’s 2026 study followed surveys across twelve months in the UK, US, Canada and Australia. Of 2,149 people who completed at least one wave, 979 completed all four. The researchers distinguished emotional isolation from a broader social-connection measure. Increased social chatbot use predicted later increased emotional isolation; the equivalent effect on broad social connection was not significant. Lower connection predicted more subsequent use.[2]

The research was observational, exploratory and not preregistered. Its modelling separated stable differences between people from changes within people, but could not eliminate every alternative explanation. Some results varied across robustness checks. It is evidence worth taking seriously, not proof that an assigned dose of chatbot conversation caused worsening loneliness.[2]

Consider an explicitly hypothetical person who begins using a chatbot after moving to a new town. Their isolation may rise because they have lost nearby routines and contacts. Another person might substitute chatbot conversations for invitations they previously accepted. The outward correlation could look similar, while the causal stories differ. A useful study has to distinguish them; these examples are not reported participants.

What was actually randomised?

A four-week study by Fang and colleagues involved 981 participants and more than 300,000 messages. It varied text or voice interaction and conversation type. The revised October 2025 manuscript reports no significant effects detected from those experimental conditions on its psychosocial outcomes. Participants who voluntarily used the chatbot more nevertheless showed consistently worse outcomes.[3]

The second result does not inherit the causal strength of the first design. Random assignment concerned conversation conditions, not how much people chose to use the chatbot. Nor does comparing different chatbot conditions establish the effect of using a chatbot rather than not using one. The manuscript is a preprint; its revised findings should not be replaced by a more dramatic description from an earlier version.[3]

Attachment is not the same as wellbeing

Another exploratory study, hosted by Microsoft Research in April 2025, tracked 149 people over five weeks. Eighty-nine were encouraged to try AI tools for social and emotional interaction, while sixty continued using the internet as usual. The researchers reported increases in emotional attachment and in perceptions of AI as empathetic or entertaining.[4]

Those findings describe how a relationship with technology can develop. They do not, by themselves, establish that loneliness improved or that human relationships were displaced. An entertaining system can be emotionally important without being a tested therapeutic intervention. Attitudinal outcomes deserve to remain visible rather than being relabelled as a general improvement in mental health.[4]

What a more useful evaluation would follow

The next question is not simply whether users like the companion. A stronger evaluation would follow what role it plays. Is it a rehearsal space before calling a friend, a temporary source of company during an unusual week, or the place where most social needs are increasingly directed? These are proposed distinctions for assessment, not proven categories of users.

It should record changes outside the conversation: contact with others, activities, ability to stop using the service and the person’s sense of control. It should distinguish an invitation to use a tool from a measured amount of use, and compare relevant alternatives. Otherwise “companionship” can slide between satisfaction, symptom relief and dependence without readers noticing.

There is no need to deny that a responsive conversation can help someone through a lonely evening. The unresolved issue is whether that help supports or replaces the connections they want beyond the screen. The most credible promise is not that AI can abolish loneliness, but that its benefits and costs can be studied without pretending they are the same outcome.

Report an error or suggest a correction