Maternal health AI chatbot Rosie shows no significant mental health gain in trial
In brief
- Rosie AI chatbot showed no statistically significant anxiety or depression benefit at 12 months versus a book program.
- JAMA Network Open published the trial Nov. 4, with 200 participants assigned to each group.
- Both confidence intervals included zero, showing neither a clear benefit nor equivalence.
- Low baseline symptoms, incentives and social-worker support limit how far the findings translate.
What the trial tested
Researchers assigned 200 participants to each group. The US study enrolled English-speaking people aged 14 or older who were pregnant or parenting an infant under six months old and who identified as members of racial or ethnic minoritized groups. Enrollment ran from October 2023 through December 2024 (follow-up wrapped up in December 2025).
Rosie was designed to answer health questions, offer personalized guidance, and provide reminders and educational support. It wasn't just software, though. A clinical social worker reviewed distress alerts, contacted participants and provided resources. The control group got one children's board book per month for a year, and both groups could communicate with study staff.
The numbers
The primary measures were self-reported anxiety on the GAD-7 questionnaire and depression on the PHQ-9. The 12-month analyses included 361 participants for anxiety and 362 for depression.
The adjusted anxiety difference was 0.10 points (95% confidence interval of -0.77 to 0.96). For depression, it was -0.23 points, with an interval of -1.05 to 0.59.
Both intervals include zero.
That's the detail that matters here. The trial found no statistically significant difference, and it established neither a clear benefit nor equivalence between the two approaches, so the size and direction of any underlying effect are still uncertain.
Limits on what it means
Most participants started with low symptom scores. Among those with baseline responses, the median anxiety score was 2 and the median depression score was 3. Only about 11.1% met the study's moderate-to-severe threshold for anxiety, and 9.5% did for depression.
Participants also received small payments for asking questions, and the most active users got book rewards. CryptoSlate's analysis said those conditions limit how directly the findings translate to routine chatbot use. Outcomes were self-reported, and the authors urged caution about secondary and exploratory findings.
The trial covers one chatbot, with human support attached, in one specific US population. Whether other AI interventions or different clinical populations would produce different results remains an open question.
Frequently asked questions
Did the Rosie chatbot trial show the AI had no effect?
No. The trial found no statistically significant difference in anxiety or depression at 12 months. Both confidence intervals included zero, so it established neither a clear benefit nor equivalence with the children's book program.
Why are the Rosie trial results hard to apply to everyday chatbot use?
Most participants had low starting symptom scores, outcomes were self-reported, and participants got small payments and book rewards. A clinical social worker also handled distress alerts. CryptoSlate's analysis said these conditions limit how directly the findings translate to routine chatbot use.
Who took part in the Rosie maternal health chatbot study?
The US study enrolled English-speaking people aged 14 or older who were pregnant or parenting an infant under six months old and who identified as members of racial or ethnic minoritized groups. Researchers assigned 200 participants to each of the two groups.
