Skip to content

Two weeks without ChatGPT lowered loneliness a little in a 625-user Stanford study

A Stanford, Oxford and CMU preprint found 14 days without ChatGPT cut loneliness by about 0.1 standard deviations in 625 U.S. users. Here is how solid that is.

ChatGPT Break Lowers Loneliness in 625-User Study

People who stayed away from ChatGPT for 14 days reported less loneliness than a control group that skipped it for one day, according to a new arXiv preprint from researchers at Stanford, Oxford and Carnegie Mellon. This ChatGPT loneliness study followed 625 U.S. users, and most of them used the chatbot for work rather than company.

The effect is small, about a tenth of a standard deviation, and it rests on surveys the participants filled out themselves. It still stands out, because loneliness moved even though most people were using ChatGPT mainly for productivity.

TL;DR: Researchers randomly assigned 625 ChatGPT users to 14 days or one day without AI. The two-week group reported modestly lower loneliness, and the gap was still there two weeks after they came back. The effect is about a third of what the study was sized to detect, so it reads as a signal rather than a cure.

How the 14-day ChatGPT break was run

The team recruited U.S. ChatGPT users through the Prolific platform and ran the study between May 6 and July 7, 2026. Participants had to use ChatGPT at least 15 minutes a day, have held an account for a month or more, and use Chrome.

A Chrome extension blocked 36 AI assistant sites and logged any violation. Treatment participants went 14 days without AI and earned an $80 bonus for complying, while the control group was blocked for one day, a design meant to limit demand effects.

What happened to their mood is where the numbers turn modest.

Loneliness fell 0.10 standard deviations

Two weeks without AI lowered loneliness by 0.10 standard deviations against the control group (p = 0.03). In a follow-up survey two weeks after people resumed use, the gap was 0.16 (p = 0.02).

Treatment participants also reported more social support (0.104) and more in-person interactions (0.183). Open-ended answers describe people asking others for help with tasks they would normally have handed to ChatGPT, so a coworker answered the question instead of a prompt box.

Account tenure mattered more than daily hours. Longtime users saw bigger drops, while time spent per day did not change the effect.

Why the 2025 MIT and OpenAI trial reads differently

An earlier randomized trial from MIT Media Lab and OpenAI researchers, posted to arXiv on March 21, 2025, followed 981 participants for four weeks. Its assigned conditions had no significant effect on loneliness or three other outcomes, though people who chose to use the chatbot more showed worse results regardless of condition.

That study varied how people talked to a chatbot, while the Stanford team took the chatbot away. The distinction is the detail most summaries will flatten, and it makes the new result a different test rather than a contradiction.

Where the ChatGPT loneliness study is thin

The researchers recruited 625 people to detect an effect of at least 0.3, assuming 20% attrition. The loneliness effect they found is about a third of that size.

Attrition did land at 20.0%, but it was uneven: 85.9% of the treatment group finished the endline survey against 74.3% of controls. Of the treatment participants who finished, 81.36% fully complied with the break.

The loneliness result was preregistered as a directional hypothesis and is reported without adjustment for multiple tests. OpenAI funded participant payments through its Mental Health Grant, and the paper says the company had no role in the design, data collection or analysis.

What a two-week pause does and does not show

For this group of paid volunteers, removing a work tool nudged some time toward other people. It does not show that ChatGPT causes loneliness in general, that quitting for good would help, or that the result would repeat outside a paid study.

The paper is version 1 of a preprint. A tenth of a standard deviation is small, but the direction is the part worth watching as follow-up work arrives.

Leave a Reply

Your email address will not be published. Required fields are marked *