Across 12 months and more than 2,000 adults, heavier use of social chatbots predicted greater emotional isolation later on, while feeling disconnected also pushed people toward the bots — suggesting AI companionship can become a loop in which the symptom and the attempted cure quietly reinforce each other

An AI companion is easiest to reach at the moment another person is hardest to reach. It answers immediately and can keep talking for as long as the user wants. That availability may ease a lonely evening. It also creates a harder question: what if feeling disconnected increases the use, while the use is followed by greater emotional isolation months later?

That two-way pattern appears in a peer-reviewed Psychological Science paper by Dunigan Folk and Elizabeth Dunn. They followed 2,149 adults for 12 months and found that an increase in social-chatbot use predicted a small increase in emotional isolation four months later. Feeling more isolated also predicted more use at the next survey. The results are associations across time, not proof that either change caused the other.

This is not treatment advice. I am reading one study carefully and asking what it does, and does not, show. This is one study, not settled consensus.

Four surveys captured a year of change

The researchers recruited Prolific users living in the United Kingdom, United States, Canada and Australia. Their average age was 39.99; 49 per cent were men. Each participant could complete four surveys, separated by about 121 or 122 days, from November 2023 through February 2025. Of the full sample, 979 completed all four waves, while the statistical models used every available response.

Social use was defined broadly. Participants were asked how often, during the previous four months, they had used a chatbot to seek advice about life decisions, hold regular conversations, talk about their day, ask the bot about itself or receive companionship. The six response options ran from never to daily. This measured recalled frequency, not message logs, minutes spent or the content of a conversation.

Between 26.38 and 29.68 per cent reported some social-chatbot use at each wave. Among users, the most common answer was only a few times in the preceding months. Roughly 1 per cent reported daily use. These figures are useful context for a title containing “heavier use”: the study was not dominated by people living inside companion apps.

“Heavier” meant more than a person’s own usual level

The analysis used a random-intercept cross-lagged panel model. Behind that long name is an important distinction. The researchers separated stable differences between people from fluctuations within the same person. Their central test was not whether habitual heavy users were lonelier than habitual non-users.

It asked whether a person who used social chatbots more than their own typical level during one four-month period reported a change in isolation during the next, and whether the reverse sequence also appeared. Stable individual tendencies were placed into separate components. The models also accounted for four changing social events: a breakup, relocation, starting a steady relationship and becoming a parent.

This is a stronger design than a one-off survey because time has an order and each person acts partly as their own reference point. It still cannot erase an unmeasured event that affects both variables. Someone might ask a bot for advice about a new job and then feel isolated after taking that job, for example. The model could assign the later isolation to chatbot use even if the move or workplace change did more of the work.

Emotional isolation moved in both directions

The emotional-isolation outcome came from one direct question. Participants rated how lonely they had felt over the past four months, with loneliness defined for them as feeling emotionally isolated from other people. Responses ran from one to four.

When a person’s social-chatbot use rose above their usual level, their emotional isolation tended to be higher than usual four months later. The standardised coefficients were about .07 to .08 across the three intervals, with p = .006 for the shared unstandardised path. The other direction also appeared: above-usual isolation predicted above-usual chatbot use at the next wave, with a standardised coefficient of .06 and p = .023.

Those are modest associations, not a dramatic transformation. The paper’s robustness checks support caution as well as interest. Chatbot use predicted later isolation in five of six alternative analyses, with the sixth just outside the conventional threshold at p = .069. The path from isolation back to use was less consistent.

The loop is therefore a reasonable description of the observed sequence, but “pushed” should be read as predicted, not caused. Nor did the researchers establish that companionship was an attempted cure in every case. Their use question covered advice and ordinary conversation as well as companionship.

The wider connection measure broke the neat loop

The researchers repeated the analysis with a 20-item social-connectedness scale. It asked about a person’s place in a larger social world, including connection to other people, comfort in new situations and a sense of being in tune with the world. Scores were highly stable: enduring differences between people explained 85 to 89 per cent of the variation at any wave.

On this measure, lower-than-usual connection predicted more social-chatbot use four months later. The standardised paths ranged from -.07 to -.09, with p = .004. More chatbot use did not significantly predict a later fall in broad connection; those paths were only -.02 to -.03, with p = .369.

That asymmetry matters. A direct, single question detected a small later rise in isolation, while a broader and more stable scale did not detect a comparable fall in connection. The authors suggest technology use might shift a person’s immediate sense of isolation without quickly altering their social identity, but they label that explanation as speculation. A single item can also be noisy, while a 20-item trait-like measure may be slow to move.

Additional models found no evidence that chatbot use predicted a later change in perceived social support or number of close friends. Both were measured with one item, so that null result does not show that human contact stayed unchanged.

Immediate comfort and longer change can coexist

The study does not cancel experimental evidence that a responsive chatbot can make someone feel less lonely immediately after a conversation. A series of experiments on AI companions reported short-term reductions. A few minutes of relief and a small association four months later answer different questions.

I saw the same distinction in Silicon Canals’ earlier look at a two-week randomised study. A deliberately supportive AI friend reduced some negative mood, but only students paired with a real person were less lonely at the end. That study used 296 first-year students and a short intervention; the new paper observes adults over a year. Neither design completes the other’s missing evidence, but together they make “did the bot feel helpful?” look too narrow.

There is a social reason these systems can feel helpful. They offer attention without scheduling, embarrassment or the risk of being rebuffed. Earlier Silicon Canals coverage of attachment to AI companions examined how consistently feeling heard can make the relationship matter to a user. That experience should not be mocked. It also should not be treated as evidence that the person’s wider social world improved.

A four-week MIT Media Lab and OpenAI study covered here previously likewise linked heavier daily use with worse psychosocial outcomes in some analyses. Its authors also resisted a simple causal reading. The repeated difficulty across this work is selection: people who need companionship may use the systems differently from people who open them occasionally.

What the study still cannot tell us

Folk and Dunn are unusually direct about the weaknesses. The analyses were not preregistered. The dataset contained other possible outcomes, creating many analytical choices, and the authors say several p values would not survive correction for multiple comparisons. They describe the findings as exploratory. The main article’s computations were independently reproduced, and the materials, data and analysis scripts are public, but openness does not turn an exploratory association into a causal estimate.

Attrition left fewer than half of participants completing all four surveys. Prolific users may be more comfortable with technology than the wider population. Nearly 30 per cent reported social use at a typical wave, which the authors note is unusually high. The frequency scale also grouped very different products and relationships together. Two people who both selected “daily” might have spent different amounts of time, discussed different subjects and received very different responses.

Most of all, the study did not observe the proposed mechanism. It did not show that chatbot conversations replaced calls, messages or time with friends. It did not test whether certain bot behaviours are safer than others, or whether a user who treats AI as an occasional supplement differs from one who treats it as a primary relationship. Longer preregistered studies with usage logs, daily measures and experiments are needed to separate those possibilities.

If loneliness has become persistent, severe or disruptive to daily life, a GP or licensed therapist can offer support that an article or chatbot cannot.

For now, the evidence supports watching both sides: whether the chatbot conversation feels comforting, and what happens to the person’s human contact in the days and months around it.