How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use

In one paragraph

Researchers at MIT Media Lab and OpenAI asked 981 adults in the US to use ChatGPT for at least five minutes a day for four weeks. Each person was randomly given one of three ways of talking (text, a neutral professional voice, or a warm, expressive voice) and one of three kinds of conversation: anything they liked, a daily impersonal topic, or a daily personal prompt such as “Help me reflect on what I am most grateful for in my life.” Each week they answered standard questionnaires on loneliness, time with friends and family, emotional dependence on the chatbot and problematic use. The surprise was that the nine combinations made almost no difference. What went with worse outcomes was how much people chose to use the chatbot, and how they came to see it.

Three takeaways

1. 1

   The design settings barely mattered.

   Voice or text, personal or impersonal topics: none had a significant effect on loneliness or time with other people. Text showed a slight trend towards worse outcomes than voice, and personal prompts went with slightly lower dependence, a borderline result. The authors say any real effects are probably smaller than the study could detect.

2. 2

   More voluntary use went with worse outcomes.

   People averaged just over five minutes a day. Those who chose to spend more time than others in their group reported more loneliness, less socialising, more emotional dependence and more problematic use. This is an association, not proof of cause: time of use was never assigned. People who were lonelier at the start did not go on to use it more, which makes the reverse explanation less likely but doesn’t rule it out.

3. 3

   How people saw the chatbot mattered.

   Participants who saw it as a friend, trusted it more or thought of it as conscious reported more emotional dependence; seeing it as a friend also went with less time with other people. These were exploratory analyses, and the effects were small.

What it means for design

Proposals for our patterns, drawn from the book. They are not yet part of the published pattern pages.

Ending the conversation

The authors name daily time as a signal to watch and suggest soft caps and break nudges as something to test. They did not test them. The pattern’s “Suggests a break” and “A length they choose” options are reasonable to try; if you build one, measure whether people fare better, not only whether sessions get shorter.

How human the assistant seems

The expressive voice was rated the most human-like, yet it was not the worst condition. What went with dependence was the belief: friend, trustworthy, conscious. Test what your assistant leads people to believe about it, every claim of feeling, friendship or awareness, not only how human it sounds.

Devotional audio

Nothing here argues against audio on well-being grounds. But the researchers’ automated classifier found the expressive voice more often missed that someone was uncomfortable or needed space: 14.19% of replies, against 3.22% in text. A warm narrator or voice companion still needs to know when to stop.

Handing over to people and Send people to God and to one another

More time with the chatbot went with less socialising. Over months, ask whether the people you serve are more in touch with others, not only whether they return. In my paper Digital Spiritual Formation in Post-Christian Europe I propose measuring whether users’ human connections increase or decrease over time; this study used a standard scale for exactly that.

Prayer and Spiritual acts and AI

The closest the study came to devotional use was the personal reflection prompts, such as gratitude. They sat in the condition with slightly lower dependence. That makes guided reflection and prayer worth testing in their own right. It is not evidence that a prayer chat protects anyone.

Design for the most vulnerable first

Low self-esteem, attachment anxiety and earlier use of companion chatbots went with worse outcomes. The average participant stayed close to population norms; set limits for the person most at risk.

Where we differ

The study measures well-being: loneliness, time with others, dependence and problematic use. That can mark harm. It can’t define the good. A product could score well on all four and still form no one in faith.

Dependence needs care in Christian terms. We are made to depend on God and on one another; the question is what we lean on and where it leads. The scale here measures craving for the chatbot and distress without it, and that is rightly a warning sign. A faith app should point any reliance beyond itself, to God and his people.

Less use is not failure. Someone may open an app less because they have joined a group or found a church, or because they are resting or in a dry season. Judge the person’s life, not their minutes.

In my paper Digital Spiritual Formation in Post-Christian Europe I cite this trial as showing that heavy chatbot users fared worse. The original is more careful than my sentence: heavier voluntary use, an association, and small effects.

Who should read it

• Teams building chat, voice companions or any AI feature where people may talk at length or late at night.
• Ministry leaders and researchers who hear the study quoted: the headline is easy to overstate, so read the results and the limitations.
• The main text is 13 pages; the supplement has the daily prompts and full tables.

How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use at arXiv preprint. No affiliate link.

Evidence and limits

• A preprint. Version 2 on arXiv, October 2025. We know of no peer-reviewed journal version yet.
• The main finding is an association. People chose how long to chat, so heavier users may differ in ways the study did not measure. The authors say they cannot establish cause.
• Small effects in a healthy sample. Average scores barely moved over four weeks and stayed near population norms; dependence and problematic use stayed low throughout.
• Who took part. 2,539 people enrolled and 981 completed the study and were analysed. All were paid US adults fluent in English, recruited online and paid $100 for finishing; most had never used a companion app.
• No comparison without AI. Everyone used a chatbot, so the study can’t compare chatbot use with not using one. It ran from just after the November 2024 US election to around the end of that year.
• One model, one company. ChatGPT (GPT-4o) with OpenAI’s safeguards, and four of the authors work at OpenAI. Companion apps built for attachment may behave differently.
• Exploratory details. The findings on perceptions and personal traits come from many extra analyses, and a few coefficients differ slightly between the text and the summary table. Quote the direction, not the decimals.

Cited studies we have not checked yet

What the book reports, as it reports it. We don’t cite these as findings until we have read the original.