In 2026, Nature Medicine, NEJM AI, and JAMA Network Open successively published three AI mental health studies whose conclusions appear contradictory: the first states that an AI psychotherapy system 'consistently outperformed human clinicians' in expert blind evaluations; the second reports that an AI well-being app significantly improved resilience and sense of belonging among college students; and the third—based on a 12-month follow-up randomized trial—finds that an AI chatbot delivered no benefit for anxiety or depression. All three are correct—and none is fully correct. Placing them side by side yields this year’s clearest—and most easily misinterpreted—evidence map for the field. For anyone who has used, or is considering using, AI to discuss personal concerns, this map is worth examining.
## First Threshold: Does the Conversation Resemble Therapy?
The study published online in March in Nature Medicine and included in the May print issue was led by the Limbic team, with co-authors from the Max Planck–UCL Centre for Computational Psychiatry and Ageing Research and the University of Tübingen. The team augmented a general-purpose large language model (LLM) with a 'cognitive layer' architecture, embedding specialized psychotherapeutic reasoning capabilities. Evaluation proceeded in two steps. First, a randomized double-blind experiment: 227 human participants engaged in conversations with different therapeutic agents, generating naturalistic mental health dialogue transcripts, which were then blindly scored by a consortium of 22 clinical experts. The conclusion: the LLM enhanced with the cognitive layer consistently surpassed both the strongest standalone LLMs and human clinicians on key clinical competencies required for delivering high-quality cognitive behavioral therapy (CBT). Second, moving toward real-world application: the team analyzed 19,674 dialogue transcripts (from 8,920 users seeking psychological well-being support) drawn from a large-scale deployment and found that higher activation of the cognitive layer correlated with greater symptom improvement and higher likelihood of long-term clinical recovery at approximately 10 weeks.
Three points must be clearly seen. First, the dialogue transcripts scored in the evaluation were simulated by research participants—not generated during actual patient treatment. Second, the real-world analysis was observational and correlational—not a randomized trial—and cannot rule out reverse causation (e.g., people already feeling better may engage more deeply). Third, of the paper’s nine authors, eight are current or former employees of Limbic—the technology developer being evaluated—and hold equity stakes; the remaining author is a paid consultant to Limbic and also holds equity. Among the evidence covered here, this is the most impressive study on 'conversational competence'—and also the one with the deepest commercial interest alignment.
## Second Threshold: Can It Make People Feel Better?
The Flourish trial, published online in NEJM AI in July, is described by its authors as the first multi-institutional, longitudinal, pre-registered randomized controlled trial (RCT) of a generative AI well-being application. From autumn 2024, over six weeks, 486 undergraduate students across three U.S. institutions were randomly assigned to either the app group (instructed to use the app twice weekly) or a usual-care control group. Results were half-positive: the app group showed significant improvements in psychological resilience and social well-being—increased sense of belonging and community connectedness, and reduced loneliness—and buffered declines in positive affect, mindfulness, and overall well-being. Results were half-negative: in this general-population sample, no significant group-by-time effects emerged for depression, anxiety, or stress. The authors’ conclusion is cautiously worded: under purposeful, ethically aligned design, generative AI can deliver 'proactive, population-level well-being interventions' yielding measurable benefits. Note precisely what it promises—well-being benefits, not symptom benefits. Limitations are equally apparent: only six weeks long; limited to college students in a general-population sample; actual depth of usage remains unknown; it measures 'short-term benefit', not 'capacity for sustained personal change'. Another noteworthy detail: the first and second authors’ institutional affiliations listed on the paper are Flourish Science—the developer of the app.
## Third Threshold: Can It Treat Illness?
The maternal-infant trial published online in JAMA Network Open on November 2, raises the bar highest: primary endpoints were clinical symptom scales—not well-being metrics—and follow-up lasted 12 months—not six weeks. Four hundred participants—pregnant or parenting (infants <6 months old)—all self-identified as racial/ethnic minorities (47% Black, 20% Asian, 16% Hispanic, 14% multiracial, 2% Indigenous), were recruited online from all 49 U.S. states and Washington, D.C.; mean age was 32.6 years, and half were already caring for infants <6 months at enrollment. From October 2023 through December 2024, participants were randomized 1:1 to receive access to a maternal-infant health generative AI chatbot—or to an active control: one children’s book mailed monthly. Follow-up lasted 12 months, concluding in December 2025.
Results were entirely negative: at 12 months, adjusted mean differences on the anxiety scale (GAD-7) and depression scale (PHQ-9) were 0.10 (95% CI −0.77 to 0.96) and −0.23 (−1.05 to 0.59), respectively—both straddling zero. Secondary endpoints—including emergency department visits, postpartum checkups, contraceptive use, and sleep duration—also showed no between-group differences. The authors’ conclusion is direct: compared with the active control, receiving chatbot access did not improve maternal-infant anxiety or depression at 12 months. Two exploratory signals merit noting—but not overinterpreting: among the subgroup enrolled while pregnant, self-reported pregnancy-related complications were higher at 12 months in secondary endpoint analyses (not observed in longitudinal analyses); within-group exploratory analysis found that higher chatbot usage correlated with longer sleep duration—but also higher anxiety scores. Conflict-of-interest disclosure notes only one author’s consulting fees and research funding outside the trial, with 'no other disclosures'.
## My Four Judgments
Laid side by side, the three thresholds yield the following synthesis—not any single paper’s conclusion.
First, conversational competence and therapeutic effectiveness are two distinct examinations; currently, no evidence links performance on the first to outcomes on the second. Expert blind evaluation says AI’s CBT competence exceeds that of human therapists; yet among the three layers of evidence covered here, the only RCT using clinical symptoms as primary endpoints and extending follow-up to 12 months yielded a negative result. Competence scoring measures how closely the process resembles good therapy; symptom endpoints measure whether people actually get better—between these two lies a missing causal bridge.
Second, choice of endpoint and comparator nearly predetermines 'success' or 'failure'. The same technical approach yields positive results when well-being is the endpoint (Flourish), but negative results when symptoms are the endpoint (the maternal-infant trial—and even Flourish’s own depression/anxiety/stress sub-analyses). Using 'one children’s book per month' as an active comparator revealed that merely 'receiving access' does not automatically equate to 'receiving help'. Thus, when reading any claim that 'AI mental health is effective', the first question should be: What is the endpoint? The second: Who is the comparator? Winning against 'giving nothing' does not count as winning.
Third, randomization assigned 'access', not 'therapeutic relationship'. The maternal-infant trial distributed chatbot access rights; how deeply participants used it, what they discussed, and whether crisis signals were detected—all remained unmeasured. The exploratory signal—that higher usage correlated with both longer sleep *and* higher anxiety—is intriguing but directionally ambiguous—perhaps anxious individuals simply talk more; because ambiguity persists, this measurement gap is critical: we do not yet know which users should be advised to limit use of such products.
Fourth, density of commercial interest and strength of reported effect show a striking inverse relationship across these three studies: the most commercially entangled study (all nine authors held Limbic equity—eight as current or former employees, one as a paid consultant) reported the most positive results; the developer-led Flourish trial was half-positive/half-negative; and the most transparently disclosed maternal-infant trial reported uniformly negative findings. Three data points are insufficient to prove systematic bias—but they are sufficient to establish a reading discipline: the closer a claim comes to 'treating disease', the more urgently one must first ask: Who funded the evidence—and what, exactly, was measured?
## Takeaways by Role
If you’re a general user: Well-being–focused AI applications may offer modest benefits for resilience and loneliness—but they are not treatments for depression or anxiety. Clinical care remains the answer for symptomatic individuals. 'Talking extensively with AI' does not equal 'getting better'—that exploratory signal warrants self-awareness from every heavy user.
If you’re building a product: Flourish’s pre-registration and honest reporting of negative sub-analyses sets a replicable standard; the maternal-infant trial’s lesson is that access rights constitute the weakest possible intervention form—if engagement depth, content quality, and safety monitoring remain unmeasured, efficacy cannot be proven, nor harm detected.
If you’re in clinical practice or regulation: Expert-rated competence cannot substitute for endpoint-driven trials; vendor-aligned competence assessments require independent replication. The knowledge gaps are clear: mechanisms (which components work), dose–response relationships (how much usage is needed), safety monitoring (who might be harmed), and follow-up extending to symptom remission and functional recovery.
Conversations are becoming increasingly therapy-like; the evidence now needs to catch up—to therapy itself. If an AI tells you, 'I can deliver cognitive behavioral therapy to you,' would you first ask who trained it—or first ask where its 12-month trial is? Join the discussion in the comments.
Source: Generative Artificial Intelligence and Maternal Health Outcomes in Pregnant and Postpartum Individuals: A Randomized Clinical Trial. JAMA Network Open, 2026. https://doi.org/10.1001/jamanetworkopen.2026.41062