AI Deactivation Reduces Loneliness and Increases Social Interaction
Dora Zhao, Grace Wang, Yutong Zhang, Lujain Ibrahim, Robert E. Kraut, Diyi Yang
cs.HC, cs.CY
2026-11-19
In 625 active ChatGPT users, two weeks without AI lowered loneliness by 0.10 SD versus a one-day pause, and raised in-person contact. The gap was still there two weeks later.
General-purpose assistants are mostly used to get things done. A coding of more than 1.1 million ChatGPT conversations put practical guidance at 28.8%, information seeking at 24.4%, and writing at 23.9%. Relationships or personal reflection were 1.9%. Well-being research has still clustered on companionship and personal advice. Whether email, lookup, and how-to questions replace asking another person had no causal test.
Observational data cannot separate use from selection: people who already see fewer others may be the ones handing tasks to an assistant. Trials that give AI to new users only see early adoption. For people already using it, the experiment is to turn the tools off. Stanford, Oxford, and Carnegie Mellon ran that test, preregistered, on active ChatGPT users, and measured loneliness and in-person contact.
Screening 7,900 U.S.-based ChatGPT users on Prolific produced a sample of 625. Entry required ChatGPT as the main assistant, at least 15 minutes a day, an account at least a month old, Chrome as the browser, and willingness to take $40 a week to stop. Blocked on intensity and account age, 306 people were assigned 14 days off and 319 a single day off. The short arm comes from social-media deactivation trials. Both groups are asked to abstain, so the demand effect is less concentrated on the long arm.
A Chrome extension blocked 36 assistant domains and their subpaths. Violations were read from recent conversations on the ChatGPT account, including messages sent on other devices. Five survey waves plus daily experience sampling recorded how many people someone saw, time socializing, and time working, with a follow-up two weeks after access returned. The estimate is the condition-by-period interaction from a linear mixed model, in baseline SD units, fit on assigned group. Loneliness and emotional state keep raw p-values, as preregistered directional tests. Other social outcomes also get Benjamini-Hochberg Q-values.
Use was instrumental. Productivity was the top purpose for 84.4%, and 52.8% used ChatGPT for productivity at least daily. Relationships or personal reflection were 3.2% of collected conversations. Another 20.2% said they had never used AI socially.
Against one day off, two weeks off moved social outcomes. Emotional state and broader psychological well-being did not change by a detectable amount. Satisfaction with social interaction was β = 0.137 (p = 0.053, Q = 0.066), short of the line.
| Outcome | Treatment vs control | Test |
| Loneliness, during abstinence | −0.099 SD | p = 0.028 |
| In-person interaction | 0.183 SD | p = 0.019, Q = 0.037 |
| Social support | 0.104 SD | p = 0.015, Q = 0.037 |
| Daily social contacts | 0.153 SD | p = 0.007 |
| Daily time working | −0.045 SD | p = 0.576 |
| Loneliness, two weeks later | −0.164 SD | p = 0.016, Q = 0.047 |
The gap in daily contacts widened over the 14 days (β = 0.014 per day, p = 0.004). Time spent socializing was not significantly higher as a level shift (β = 0.130, p = 0.077), though the gap between arms grew (β = 0.024, p < 0.001). The abstract calls that time increase significant. The level test does not clear 0.05. Work time was flat. Extra contact did not come out of work hours.
Exploratory mediation linked day-7 social behavior to day-14 loneliness. The indirect path through time socializing was −0.025 (95% CI [−0.057, −0.001]); through social satisfaction, −0.054 (95% CI [−0.104, −0.010]). The upper bounds stop just short of zero. Open comments from the treatment arm split 41.5% improved, 42.3% little or no change, and 8.8% worse or mixed. More active outreach was 18.8%. Less phone distraction was 9.2%. Only 7.7% said they asked friends, family, or coworkers what they would have asked an assistant.
Daily time on AI did not moderate loneliness (β = 0.002, p = 0.965). Account tenure did (β = −0.109, p = 0.021). People who had used ChatGPT longer fell further when it was removed.
Follow-up reached 442 people (233 treatment, 209 control). The loneliness gap remained. Emotional state still did not differ. Competence and autonomy, each β = 0.181, were not significant when abstinence ended and showed up only later (p = 0.004, Q = 0.022; p = 0.025, Q = 0.050). Self-reported time on AI (β = −0.434) and interaction intensity (β = −0.276) stayed lower. Logs were weaker. Versus their own pre-period, treated users sent 3.7±4.6 fewer messages, opened 1.2±0.4 fewer sessions, and spent 67.6±43.1 fewer minutes, none of it reliably different from control (p = 0.855, 0.800, 0.686). People felt they were still using less. The usage record did not separate the groups.
A tenth of a standard deviation is small. The trial was powered for 0.3. The shift shows up among people whose main use was productivity (84.4%), with relationship talk at 3.2% of conversations and no detectable change in time working. The fit is displacement of small practical exchanges: what to buy, how to do a step, how to revise a draft, now ending at the assistant instead of landing on another person.
That is a weak case for quitting. The paper treats full abstinence as impractical and undesirable, and task-level gains from assistants sit in other evidence. The design point is narrower. An assistant scored only on whether one person finishes the task can also delete the thin contact that used to come attached to the task.
Attrition was uneven. Of 625 people randomized, 500 (80.0%) finished the endline: 85.9% in treatment, 74.3% in control, 11.7 points apart. Compliant treatment participants received an $80 bonus. Control pay for one day was $5.71, the $40 week prorated. Who stays is tied to the bonus. Missingness is treated as random. The main model also leaves the preregistration, swapping an endline ANCOVA for a mixed model on every wave. Weighted checks in the supplement point the same way. They do not remove the chance that the loneliness gap partly reflects which control participants still answered.
Social outcomes are self-report. Logged use at follow-up does not separate the arms. Mediation was exploratory, and the indirect intervals sit on zero. Better and unchanged comments are nearly tied, and the theme of asking a person instead is 7.7%, too thin to carry the average effect. Methods says 433 endline completers did the follow-up. Results analyze 442. Those counts do not match. The block runs in Chrome. Other assistants opened on a phone depend mostly on self-report. ChatGPT is checked at the account.
The sample is U.S. heavy users who would take money to stop, at least 15 minutes a day, on Chrome. Full compliance among treated endline completers was 81.36%. Two weeks is short. Competence and autonomy appeared only at follow-up, and anything later went unmeasured.