Saved in this browser. Sign in to sync across devices.
The distinction: Helping someone locate information, prepare questions or organize a plan is not the same as clinically demonstrating improvements in anxiety or depression. This study tested mental-health outcomes, not whether people liked the chatbot or found it convenient.
What was tested?
The US trial enrolled 400 participants across 49 states and Washington, DC. Participants were pregnant or caring for an infant younger than six months, identified with racial or ethnic minoritized groups, and were eligible from age 14 (mean age 32.6). Half received access to an AI maternal-health chatbot with information, reminders and a human-reviewed safety alert pathway. The comparison group received monthly children's books.
The researchers looked at anxiety (GAD-7) and depression (PHQ-9) at 12 months. Outcomes were analyzed for 361 and 362 participants, respectively, due to incomplete responses. Baseline anxiety and depression scores were generally low, so these findings do not directly test chatbot treatment for severe clinical disorders.
What did the trial find?
ANXIETY · 12 MONTHS
+0.10 points
Adjusted chatbot-versus-control GAD-7 score difference (95% CI −0.77 to +0.96; n = 361). Not statistically significant.
DEPRESSION · 12 MONTHS
−0.23 points
Adjusted PHQ-9 score difference (95% CI −1.05 to +0.59; n = 362). Not statistically significant.
Most measured health-care and sleep outcomes also showed no reliable advantage. One secondary pregnancy-complication analysis generated a possible safety signal, but the longitudinal analyses did not confirm it, so it should not be interpreted as proof the chatbot caused harm. More interactions with the chatbot were not randomly assigned, so associations between usage and outcomes cannot establish causation.
How far can we generalize?
Strongest available classification here: randomized evidence of no demonstrated benefit for the measured endpoints of one particular chatbot. It was not designed to compare every AI model, mental-health app or clinician-supported digital therapy. It also does not establish whether AI information tools improve appointment readiness, care navigation or caregiver workload. Those claims would need separate testing.
The study enrolled people as young as 14, but its average age was above 30; it does not prove safety or clinical effectiveness for children. For autism, ADHD, depression, anxiety, OCD and tic disorders, AI should not independently diagnose, prescribe or replace licensed care. Some digital psychological interventions have their own controlled evidence; they should not be lumped together with unrestricted chatbots.
A practical AI mental-health safety checklist
- Know its job: Is it helping with learning, checklists, questions and scheduling—or promising to be a therapist?
- Look for evidence: Was this exact product studied for the condition and age group, with meaningful outcomes and an appropriate comparison?
- Protect privacy: What personal or child information is saved, shared, used for training or accessible to others?
- Keep a human in the loop: Who can interpret the situation, review risks and adjust care when symptoms change?
- Plan for urgent needs: What happens if a user mentions self-harm, psychosis, danger, or a medical emergency? Do not depend on a chatbot as an emergency service.
- Measure actual value: Is daily life easier, are appointments better prepared, and do caregivers feel more supported? Do not confuse conversational fluency with clinical effectiveness.
How NeuroCoach uses this distinction
NeuroCoach offers educational resources, practical planning and nonclinical organization tools. A Care Plan can help collect questions, note preferences and prepare information to share intentionally with qualified professionals. These functions should not be presented as diagnosis, psychotherapy or proof of treatment effectiveness.
If someone is in immediate danger in the United States, call 911. For mental-health crisis support, call or text 988. Chatbots are not emergency services.
Original research
- Norell and colleagues, JAMA Network Open, published November 4, 2026 ↗
- ClinicalTrials.gov registration NCT06053515 ↗
- Understand what the NeuroCoach Care Plan does →
- How to read evidence labels →
Editorial note: Original study published November 4, 2026; NeuroCoach summary added November 5. Independent clinical review pending. Educational material only, not a diagnosis or individualized mental-health advice.