Can a maternal-health AI chatbot improve anxiety or depression? A 400-person trial found no 12-month benefit
A US randomized trial found no meaningful improvement in anxiety, depression, sleep, postpartum care or emergency visits after a year. The chatbot was used heavily, but it also relied on human safety follow-up and the trial was not designed to rule out small effects.
The short version
- What happened
- The randomized trial assigned 400 pregnant or postpartum people to a generative-AI maternal-health chatbot or an active control that received monthly children's books; 380 participants contributed at least one 12-month measure.
- Who it affects
- Pregnant and postpartum people should not be asked to trade away access to clinicians because a chatbot is fluent or frequently used; the tested system did not improve anxiety or depression after 12 months.
- What remains unproven
- The trial evaluated one English-language system and an online US sample from racially and ethnically minoritized groups; it does not establish performance for other products, languages or health systems.
Keep in mind
The trial evaluated one English-language system and an online US sample from racially and ethnically minoritized groups; it does not establish performance for other products, languages or health systems.
Outcome-specific missingness reached 23.5%, all outcomes were self-reported and the study was powered for moderate rather than very small effects.
Engagement analyses were observational, and incentives may have produced more use than an ordinary service would see.
The isolated pregnancy-complication signal was not consistent across analyses and cannot establish that the chatbot caused harm.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 11 min The evidence, methods, implications and limitations
The direct answer: access to the chatbot did not improve the two primary mental-health outcomes
A year of access to a generative-AI maternal-health chatbot did not improve anxiety or depression in a randomized trial of 400 pregnant or postpartum people in the United States. At 12 months, the adjusted mean difference between the chatbot and control groups was 0.10 points on the seven-item Generalized Anxiety Disorder scale, with a 95% confidence interval from −0.77 to 0.96. On the nine-item Patient Health Questionnaire for depression, the adjusted difference was −0.23 points, with a confidence interval from −1.05 to 0.59. Both intervals include no difference. They also make a moderate benefit unlikely under the study's assumptions, although they do not rule out every small effect.
That result matters because conversational systems are often presented as a way to fill information and access gaps during pregnancy and the postpartum period. The trial is stronger evidence than a satisfaction survey or a before-and-after deployment because random assignment gives a clearer comparison. Yet its conclusion is deliberately narrow. It tested one system, called Rosie, with structured safety monitoring, incentives and a particular US population. It did not test every maternal-health chatbot, establish that no design could help, or show that automated information is harmless when deployed without clinical oversight. The most defensible reading is that access and frequent use were not enough to improve the measured outcomes over 12 months in this trial.[1]
Who took part, what the comparison received and how follow-up was counted
Researchers recruited English-speaking participants aged 14 years or older who were pregnant or parenting an infant younger than six months and who self-identified with a racially or ethnically minoritized group. Recruitment was online and reached 49 states plus Washington, DC. Two hundred people were allocated to the chatbot and 200 to an active control. The control group was not left with nothing: it received monthly children's books, while chatbot participants could ask questions and also received incentives tied to engagement. This active comparison reduces the chance that contact or receiving a useful resource alone explains a difference, but it may also narrow the contrast with usual care.
At least one 12-month measure was available for 380 participants. The outcome-specific analytic samples were smaller because missingness varied from 3.5% to 23.5%: the primary anxiety analysis included 361 people and the depression analysis 362. The authors used adjusted models and sensitivity analyses, but missing data remain relevant because people who are harder to follow may also have different health needs. The trial was designed with 80% power to detect moderate effects at a two-sided 0.05 significance level while allowing for 30% attrition. That makes the null primary result informative about moderate benefit, not proof of exact equivalence or evidence that effects below the trial's detection threshold cannot exist.[1]
The distinction that matters
What the randomized comparison found at 12 months
The trial measured outcomes, not just use. Differences are adjusted chatbot-versus-control estimates, and the confidence intervals include no effect.
Anxiety (GAD-7)
Adjusted mean difference 0.10
95% CI −0.77 to 0.96; no benefit shown
Depression (PHQ-9)
Adjusted mean difference −0.23
95% CI −1.05 to 0.59; no benefit shown
Use
187 of 190 asked at least one question
Median 119 questions; engagement was substantial
Safety layer
262 alerts among 97 participants
Human clinical-social-work follow-up was part of the intervention
Use was substantial, which weakens the simple explanation that nobody tried the system
Among the 190 chatbot participants with engagement records, 187 asked at least one question. The median was 119 questions over the study period, showing that the intervention was not merely offered and ignored. Participants were paid 50 cents for each question up to a total of $45, and the most active fifth received board books. Those incentives make sustained engagement more likely than it might be in an ordinary service. They also mean raw use cannot be assumed to represent demand without incentives. High use is still important: a failure to change the primary outcomes is harder to dismiss as a zero-exposure problem.
The researchers also explored whether more engagement was associated with better outcomes. Higher engagement was associated with about half an hour more sleep, but also with a 1.63-point higher anxiety score. These analyses were not randomized comparisons. People experiencing greater worry may have asked more questions, and people with more time, support or confidence may have used the chatbot differently. The direction of causation therefore cannot be established. Engagement metrics can show what people did with a tool; they do not, by themselves, show that the tool caused their later health state. That distinction is especially important when vendors use message counts as a proxy for clinical value.[1]
Secondary outcomes did not reveal a hidden service or sleep benefit
The trial did not find improvements in emergency-department visits, postpartum care, contraception use or sleep. Those outcomes broaden the assessment beyond symptom questionnaires and make it less likely that a major practical benefit was missed only because the wrong mental-health scale was chosen. However, all outcomes were self-reported rather than confirmed in clinical records. Self-report is appropriate for lived symptoms, but health-service use and pregnancy complications can be misremembered or classified differently by participants. The study also tested multiple secondary outcomes, increasing the possibility that an isolated statistically significant result appears by chance.
One endpoint analysis among participants who were pregnant at baseline found more self-reported pregnancy complications in the chatbot group, with an adjusted risk ratio of 1.40 and a 95% confidence interval from 1.02 to 1.92. The same pattern did not hold in the longitudinal analysis, and the authors treat it as hypothesis-generating rather than evidence that the chatbot caused harm. There are several plausible alternatives, including chance, differences in reporting, or greater recognition and reporting of a condition after asking questions. The signal should be investigated, not converted into a causal claim. A replication using clinical records and prespecified safety outcomes would be needed before concluding that chatbot access changes complication rates.[1]
The tested intervention included automated alerts and human follow-up, not a stand-alone chatbot
Safety infrastructure was part of the intervention. The system generated 262 alerts among 97 chatbot participants, and a clinical social worker followed up when indicated. The paper reports no intervention-related adverse events. That is reassuring within the trial, but it also changes what can be generalized. A public chatbot with no monitoring, no escalation route and no clinician available after a concerning message is not the same product or service. Health organizations evaluating conversational AI should ask who reviews alerts, how quickly they respond, what happens outside service hours, whether users understand the limits, and whether the system can reliably recognize emergencies across language, dialect and literacy differences.
The trial also does not support replacing antenatal, postpartum or mental-health care. A chatbot can be available at any hour and may help people formulate questions, but it cannot perform a physical examination, confirm a diagnosis, monitor a fetus, assess nonverbal cues or guarantee an emergency response. The large number of alerts illustrates the operational workload that can sit behind a seemingly automated service. A deployment that reproduces the trial's safeguards would need trained staff, documented escalation rules, audit trails and capacity planning. A deployment that omits them cannot cite this trial as evidence for safety simply because the conversational interface looks similar.[1]
What the population, funding and disclosures mean for interpretation
The study focused on people from racially and ethnically minoritized groups, populations that often face unequal access, discrimination and gaps in maternal-health information. That focus is a strength for the question the trial asked. Online recruitment across almost the entire United States added geographic reach, but the sample was English-speaking and able to participate through digital channels. Results may differ for people with limited connectivity, people who use other languages, those in acute crisis or those receiving care in health systems with different access and support. Participants as young as 14 were eligible, but the paper's overall result should not be read as a separate powered conclusion for adolescents.
The work was funded by the US National Institute on Minority Health and Health Disparities, and the paper says the funders had no role in the study's design, conduct, analysis or publication decision. One author reported personal fees from Rashi.ai and outside grants from Adobe, Meta, the National Science Foundation and DARPA; the remaining authors reported no disclosures. Disclosure does not invalidate a result, and the randomized design, registration and published confidence intervals make the evidence inspectable. It does mean readers should keep commercial relationships visible, particularly in a field where product design and evaluation can develop together.[1]
The practical standard should be measured benefit, not conversational fluency
For maternity services, the trial argues against treating a natural conversation or high message count as evidence of improved health. A procurement decision should specify the outcome that matters, the denominator, the comparison, the follow-up period and the human resources needed for safety. Before wider use, providers should independently test answer accuracy, escalation sensitivity, subgroup performance and the burden of false alerts. They should also protect access to human care so that a digital option does not become the only route offered to people already facing barriers.
The assessment would change with replicated randomized evidence showing a clinically meaningful improvement, credible subgroup benefit defined in advance, or reduced access gaps without new safety problems. It would also change negatively if larger trials found confirmed harm, unreliable escalation or greater inequality. For now, this is a consequential peer-reviewed study because it replaces an assumption with a controlled test. It is not Breaking: there is no emergency directive, newly identified acute hazard or immediate change in clinical practice. Its proper role is to sharpen the evidence standard for maternal-health AI and remind readers that a well-used tool can still fail to improve the outcomes it was meant to change.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring mental health: compare the finding and its evidence limits.
Do mental-health chatbots help university students?
A preregistered review found small short-term symptom advantages over waitlists or low-intensity materials, based on three depression and four anxiety comparisons. Two active conversational-control trials found no reliable added benefit, and certainty was low to very low.
Read the next reportExplore the full source and evidence record
What this means for people
- Pregnant and postpartum people should not be asked to trade away access to clinicians because a chatbot is fluent or frequently used; the tested system did not improve anxiety or depression after 12 months.
- Health services that deploy conversational AI need staffed escalation, documented response times and subgroup testing because human safety follow-up was part of the intervention evaluated here.
- Commissioners should ask for patient outcomes and access measures, not rely on message counts, satisfaction or demo performance as substitutes for clinical benefit.
Global context
Maternal-health access and mortality differ sharply across countries, and the infrastructure behind a chatbot matters as much as its language model. Systems with fewer clinicians may feel stronger pressure to automate, but this US trial cannot show that a similar interface will improve outcomes where referral, emergency transport, connectivity or language support differ. The transferable lesson is methodological: compare against a credible alternative, measure patient outcomes over time and describe the human safety service bundled with the software.
What the evidence does not yet show
- The trial evaluated one English-language system and an online US sample from racially and ethnically minoritized groups; it does not establish performance for other products, languages or health systems.
- Outcome-specific missingness reached 23.5%, all outcomes were self-reported and the study was powered for moderate rather than very small effects.
- Engagement analyses were observational, and incentives may have produced more use than an ordinary service would see.
- The isolated pregnancy-complication signal was not consistent across analyses and cannot establish that the chatbot caused harm.
What to watch next
- Independent replications with clinical-record outcomes, prespecified safety endpoints and larger samples capable of detecting smaller effects.
- Trials in additional languages and health systems, including direct measures of whether AI reduces or widens access gaps.
- Transparent reporting of alert accuracy, response times, staff workload and outcomes after escalation.
Living evidence record
Impact record IAI-0BY7TPF
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
4 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what JAMA Network Open published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 4 November 2026
Research question, topics and regions
Whether access to a generative-AI maternal-health chatbot improves anxiety, depression and health-service outcomes over 12 months for pregnant and postpartum people from racially and ethnically minoritized groups
Topics: Maternal health · Mental health · Generative AI · Randomized trial · Chatbots · Health equity · Clinical safety
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.