Life after digital companionship: lessons from a four-week experiment

Look beyond a single conversation

Digital wellbeing research concerns both immediate feelings and everyday life: handling difficulties, staying connected, and choosing when an interaction ends.

Fang and colleagues followed chatbot use over four weeks. Their experiment compared different arrangements for chatting, rather than AI use against a no-use condition. That distinction determines what its findings can establish.

What was compared?

The 981 participants were assigned to nine combinations of text, neutral voice, or expressive voice and open-ended, non-personal, or personal topics. They were asked to use the chatbot for at least five minutes daily. Surveys tracked loneliness, offline social interaction, emotional dependence, and problematic use.

The assigned conditions did not produce significant differences in these outcomes. This does not establish that all designs are equivalent; it describes what was detected under this study’s arrangements and measures.

Voluntary use is a different comparison

Participants who chose to use the chatbot more reported worse outcomes, including greater loneliness and dependence and less offline social interaction. Usage quantity was not randomly assigned, so these associations do not directly show that longer conversations caused deterioration.

For digital wellbeing, experimentally varied features and naturally developing habits provide different kinds of evidence. Both can matter, but they should be labeled clearly.

Evaluate a sustained support experience

Future studies can specify whether an experience aims to support emotional expression, help people organize a problem, or change a sustained outcome. Measures, comparisons, follow-up, and reporting of withdrawals can then match that question.

For an emotion-reflection feature, proposed outcomes might include clarity about feelings, a personally meaningful later action, and whether any change persists. These are research proposals, not effects established by this experiment.

Measure engagement and support separately

Duration and retention describe participation. Feeling understood describes an immediate experience. Offline connection and daily functioning concern wider outcomes. Considering them together may reveal increasing engagement alongside a growing burden.

A useful evaluation also allows for no improvement, no change, and experiences that prove unsuitable. The aims and reporting approach should be clear before results are known.

Scope and next questions

The study used a particular ChatGPT system, had no no-AI comparison group, and covered four weeks. This article uses the October 2025 revised preprint. OpenAI funding and author affiliations remain disclosed in the sources.

The next question is what forms of use can offer support, for whom, and in which circumstances. Greater willingness to chat should be examined alongside those outcomes.