How long can digital companionship help? From a moment to daily life

Feeling less alone after a chat—and in everyday life

When no one is available, a response may ease loneliness in the moment. Digital wellbeing research can take that experience seriously while asking how it fits into life: does a change remain before the next conversation, and what happens when use stops?

All these questions concern companionship, but they call for different evidence. A responsive conversation may offer a meaningful moment of support. Understanding sustained support also requires observing life between sessions and after use ends.

AI Companions Reduce Loneliness reports five studies, including app-review analysis and experimental comparisons. Here we focus on Study 3. Its week of repeated observations lets us distinguish feelings after a conversation from the state in which people begin the next one.

Who did what during the week?

The researchers recruited 1,088 people and excluded 16 who failed a comprehension check, leaving 1,072 initially. The preregistered main analysis used people who completed all seven days in the two longitudinal groups. The final sample is often summarized as 922, but this includes the prediction group, which did not take part in seven days of follow-up.

There were 676 seven-day completers, rather than 922. The prediction group helped examine what people expected from AI companionship; it was not another group receiving a week of intervention. Questions about persistence below primarily concern the experience and no-chat control groups.

The companion was built by the researchers using GPT-4, specifically gpt-4-0125-preview. It retained summaries to support memory across days and combined caring response instructions with features such as proactive check-ins. The study tested this package. It did not separately manipulate memory, wording or check-ins, so the result cannot be attributed to one feature.

Controls completed a daily loneliness questionnaire without a conversation or another activity of equal duration. This provided a no-chat comparison, while leaving time, attention and participation in an activity incompletely matched.

When did the average score fall?

Pooling the repeated daily observations, the experience group's mean loneliness fell from 36.64 before chats to 30.74 afterwards. The paper reports an effect size of d = 0.20. These numbers describe the repeated before-and-after difference. They are neither a cumulative change from day one to day seven nor a percentage of people who responded to treatment.

The first row shows that immediate relief could recur across the week. The second indicates that daily comparisons with the control were less consistent. The third is particularly informative: the experience group did not show lower loneliness than controls before its daily conversations.

Both groups' scores gradually declined over the week, so the shared trend cannot be credited to AI. Conversely, a nonsignificant difference does not establish that the groups were identical or that no AI companion could ever have a lasting benefit. This study did not establish that sustained advantage.

Three questions about how long support lasts

Does the feeling change by the end of an exchange?

This is where the evidence is most direct. Under this study's conditions, participants reported lower current loneliness after talking with this system. A brief experience can have value when someone temporarily has nobody available to talk to. The study did not establish a clinical threshold for that value; the scores should be interpreted within the scope of what they measured.

Does the change remain before the next exchange?

Daily pre-chat measurements provide a view of the time between sessions. They did not show a continuing advantage over controls. “People felt better after chatting on each of seven days” and “people became progressively less lonely during the week” therefore require different statements. Repeated immediate changes do not automatically accumulate into a different everyday baseline.

Does life change after use ends?

The study did not follow participants for weeks or months after the seven-day task ended. It did not establish changes in social networks, clinical symptoms or remission rates. Persistence after stopping, greater willingness to contact others and dependence on the system all require additional observation points and appropriate measures.

What limits the conclusion?

The control activity was not matched to chatting. The experience group spent substantially more time interacting. Activities with comparable duration, attention and effort would help distinguish the specific contribution of AI conversation. Other studies in the paper include different randomized comparisons, but they mainly examine short encounters and do not supply the missing long-term follow-up.

Completers may differ from natural users. About 23% of the experience group and 14% of controls did not complete the week. Missing a day ended invitations to subsequent sessions. The main analysis focused on seven-day completers, and completing the week brought an additional $15 bonus. Payment, task requirements and unequal attrition may influence who stays and how they participate. The authors report supplementary robustness analyses; we did not download and rerun the underlying data or reproduce those analyses.

The outcome was a self-reported current feeling. Participants knew their activity and repeatedly answered loneliness questions. Expectations, repeated measurement and guesses about the research purpose could influence ratings. This was not a treatment trial in a clinically diagnosed sample, and there was no blinded assessment of clinical outcomes.

The population and system were specific. One week with a research-built GPT-4 companion among paid US online adults does not represent Chinese-language use, all commercial products or established human relationships. The paper did not test either the Luoboli game or Luoboli Sleep, nor compare long-term AI use with maintaining human ties.

How this distinction could guide product evaluation

The following are proposals for testing, informed by the research. They are neither features shown effective by this paper nor findings from hands-on tests of our products.

Specify the change and its timing in advance

For brief companionship, a study might preregister state changes before and after an exchange. For sustained support, it could add measurements before the next day's use, on non-use days and after use ends. Defining the primary outcome and timing first reduces the temptation to select the most favorable observation. Reporting means, effect sizes and uncertainty also fits the evidence better than converting a scale change into a “cure rate.”

Make leaving an understandable choice

A product should accurately disclose its AI identity and allow people to pause or leave. Evaluation could ask why someone stops: did they get what they needed, lose interest, feel uncomfortable or find it unhelpful? Retention and time spent cannot distinguish these possibilities on their own. Natural-use research also needs the experiences of people who leave, rather than only those who return every day.

Reconnect the exchange with everyday relationships

Beyond asking how a chat felt, research could examine actions users choose outside it: expressing a need to someone they know, for example, or attending an activity they had wanted to join. Whether arrangements that support stepping away from the screen and reconnecting help people is a separate question to test. A lower post-chat score does not answer it.

Momentary relief and lasting change both deserve study. Keeping them distinct preserves the meaning of an immediate experience while making clear what has been observed and what still needs follow-up.

Make the timescale of support explicit

Digital wellbeing does not require every reassuring moment to become a claim of lasting change. Separating session, next-day, and post-use outcomes describes support more accurately. Future designs can test a specific aim, such as expressing a need or preparing for an offline conversation; this study did not establish those effects.