Small Effects, Wide Reach, and High Attrition in the Body Image Chatbot

The study was remote, randomized, and preregistered

1,715 Brazilian adolescents were assigned to a chatbot or a survey-only control. Of the total sample, 901 were girls. Participants came from different regions of Brazil.

Measurements were taken at baseline, immediately after the intervention period, and at one week and one month. Primary outcomes concerned momentary and more general body image.

Secondary outcomes included momentary and longer-term affect, as well as body-image-related self-efficacy. The preregistration strengthens transparency about which outcomes were originally the focus.

A fully digital implementation facilitates reach but reduces personal support. Technical and motivational hurdles thus have a more direct impact on participation.

Micro-interventions were offered over 72 hours

The chatbot contained brief, standalone techniques. Micro-interventions are intended to provide immediate relief or a small improvement, rather than fully replicating long-term therapy.

Of the 327 people who entered the chatbot, 258 completed at least one technique. This corresponds to 78.9 percent of this actual user group.

On average, five techniques were completed within the 72-hour period. The format was thus deliberately compact and did not require months-long commitment.

This limitation is a strength as long as the purpose remains clear. A brief impulse can be useful without claiming to comprehensively treat deeper body image issues or mental health conditions.

The effects were small and statistically significant

For momentary body image, Cohen's d was 0.30. For the more general body image, the study reported small values between 0.10 and 0.26 depending on the time point.

Small significant differences were also found for momentary mood, positive and negative affect, and self-efficacy at certain measurement time points.

Small effects are not meaningless, especially for low-threshold offerings with large potential reach. However, they should not be presented as strong individual change.

Baseline burden moderated the benefit, but gender did not. This suggests that the starting state was more important for the effect than a blanket gender-specific assumption.

61.9 percent attrition changes the assessment

In the intervention arm, 531 of 858 participants were lost. The attrition of 61.9 percent was high and corresponds to a known problem in digital interventions.

An application can work for those who use it while simultaneously losing many of the people originally reached. Both statements must be reported together.

Reasons for dropout are central to further development. Technical issues, lack of time, low relevance, unpleasant content, or already sufficient benefit can be very different causes.

A scalable product should therefore not merely count registrations or starts. The path from initial contact to actual use and to the completion of individual elements must become transparent.

In addition, the analysis should distinguish between participants who never entered the chatbot, dropped out after the first step, or completed multiple techniques. These groups pose different questions regarding outreach, technology, and content fit.

Reminders and notifications can increase engagement, but they can also create pressure. In the context of body image topics, it is particularly important to assess whether frequent prompts unintentionally reinforce preoccupation with one’s own appearance.

Woebot demonstrates the value of process factors

Fitzpatrick and colleagues randomized 70 young adults to either two weeks of Woebot or an informational e-book. The Woebot group used the chatbot on average 12.14 times.

On the PHQ-9, a significant group difference emerged in favor of Woebot. In the analysis of complete cases, both groups showed improvements in anxiety scores.

Participants’ comments suggested that process factors influenced acceptance more strongly than content that mirrored traditional therapy.

For micro-interventions, this means: not only the exercise itself matters. The entry point, tone, length, transitions, and the feeling of not being lectured can determine whether a user even reaches the technique.

Trust can promote use and increase risks

Choudhury and Shamszare found significant associations between trust, intention to use, and actual usage among 607 regular ChatGPT users.

A body-image chatbot needs sufficient trust for adolescents to engage in an exercise. At the same time, its social design must not create excessive authority over statements about appearance, weight, or health.

The system should not derive individual assessments that it cannot professionally substantiate. Positive support must not tip into new normative ideas about what a body should look like.

Transparent role, data protection, and the ability to reject content are therefore part of the intervention, not just the legal framework.

Brazil is not just a place for scaling up

The study addresses gaps in care in low- and middle-income countries and demonstrates a geographically broad implementation in Brazil. This is more than the translation of a Western pilot project.

Local language, platform usage, and ideas about body image influence whether interventions are adaptable. An application should be developed with the target group, not merely rolled out technically.

Scalability must also not mean handing over data and control to a few external providers. Costs, infrastructure, and long-term availability are part of the real benefit.

Subsequent experiments should investigate which groups particularly benefit or drop out, and how the offering can be embedded in local prevention and care.

This should also take into account end devices, data costs, and network quality. A fully remote intervention can be geographically broad and still exclude adolescents whose technical conditions do not allow for regular, private use.

Micro-intervention is a clear and testable role

The study shows small positive effects of a limited body-image chatbot and, at the same time, a high dropout rate. It is precisely this combination that makes it valuable for product development.

A micro-intervention does not promise a fully-fledged conversation partner or therapy. It offers a specific technique at a particular point in time and can therefore be evaluated in a targeted manner.

Generative chats can embed such techniques more naturally, but they increase variability and the risk of inappropriate transitions. Their effect must not simply be assumed to carry over from a structured intervention.

Small effects, broad reach, and high dropout rates belong in the same balance sheet. Anyone who presents only one of these parts describes neither the opportunities nor the limitations of digital support in full.