Does heavy ChatGPT use make people lonelier and more dependent?

OpenAI researchers studying ChatGPT use ran two parallel studies: an on-platform analysis of "over 3 million conversations" plus a survey of "over 4,000 users," and an IRB-approved randomized controlled trial (RCT) of "close to 1,000 participants" over 28 days. Across both, "very high usage correlates with increased self-reported indicators of dependence," and the platform analysis found "a small number of users are responsible for a disproportionate share of the most affective cues" — most conversations carry little affective content, but a minority of power users show elevated activation on nearly every emotional-cue classifier. In the RCT, "participants were both less lonely and socialized less with others at the end of the four-week study period," and those who spent more time using the model were "statistically significantly lonelier and socialized less."

The modality effect is nuanced rather than uniform. Controlling for usage duration, using a voice mode was tied to "statistically significantly less loneliness, less emotional dependence and less problematic use of the model" than text — but "longer usage duration of neutral voice modality had statistically significantly lower socialization and greater problematic usage compared to using the text-based model," so voice's benefit erodes at the high end of usage. Conversation task mattered too: personal conversations produced "more loneliness but also less emotional dependence and problematic usage" than open-ended chat, an effect that became non-significant at longer durations. A user's starting state moderated outcomes most strongly: people who began the trial high in emotional dependence and problematic use saw those measures drop when using the more engaging voice mode rather than text. The authors are explicit about the limits of the on-platform half of the work: "From a purely observational study, we cannot draw direct connections between model behavior and users' usage patterns."

This gives the Character.AI finding that Does sustained engagement with AI companions harm well-being? a second, randomized data point on a general-purpose platform: ChatGPT usage duration also tracks more loneliness and less socializing. But this study complicates rather than confirms that picture — the effect is not fixed, it is offset by voice interaction and shaped by what the conversation is about. It also sharpens How does emotional chatbot use develop from practical use?: where that 18-person study showed socioemotional use emerging gradually from practical use, this platform-scale analysis shows the pattern is concentrated in a small subset of heavy users rather than spread evenly across the user base.

The excerpt does not establish why voice attenuates loneliness relative to text, nor whether the "power users" who generate most affective cues are representative of any broader population — the paper flags that it lacks demographic controls (gender, prior AI familiarity) for its on-platform data and that self-report survey responses "may reflect their self-perception more than their actual behavior." The RCT runs only 28 days, so nothing here speaks to effects beyond a month. The implication to draw is narrow: claims that chatbot use straightforwardly helps or harms emotional well-being should specify modality, task, duration, and the user's starting state, because this study found the sign of the effect flips depending on all four.