The team retrospectively analyzed Ash conversation logs from May 2025 to January 2026, a study determined exempt by the BRANY IRB. An automated classifier followed by clinician review of every flagged case confirmed 419 perinatal users, whose 4,087 sessions formed the basis for the emotion analysis. Each session was coded by an LLM judge against a new 26-label, six-domain Perinatal Emotion Expression Taxonomy anchored in established perinatal theory, and a separate model produced continuous PHQ-9 estimates over two-week windows. A date-matched cohort of 46,134 non-perinatal users served as comparison.

Perinatal users differed from the general cohort mainly in what they discussed rather than in overall distress. Parenting (Cohen's d = +1.34), family (+0.72), and life transitions (+0.62) were markedly elevated, while anxiety and depression were if anything slightly lower. Sessions were emotionally multidimensional, averaging 5.38 active emotion labels with 77.8% expressing three or more concurrent states, and relational and attachment content was dominant, appearing in 75.9% of sessions. Positive emotions were rarely expressed in isolation: only 7% of sessions endorsed positive content without any accompanying distress.

Distress did not follow a simple rise or fall. Two troughs stood out, the second trimester and the early-infant phase from six weeks to three months postpartum. Crucially, distress did not resolve after the early postpartum weeks: across the first postpartum year, positive affect declined, with hope and optimism falling from 68.5% to 54.6%, while identity and role strain and emotional exhaustion rose steadily, with emotional exhaustion climbing from 28.4% to 38.4%. This pattern runs counter to the common assumption, and to maternal-role theory, that postpartum distress eases over the first year, and the authors argue it strengthens the case for screening that extends well beyond the standard six-week visit.

On symptoms, the signal concentrated among those who needed it most. Among users entering at moderate severity or above, about half the sample, model-estimated PHQ-9 scores declined by 2.41 points over a median eight-week follow-up (d = -0.58), with 27.1% meeting a reliable-improvement threshold, while users entering below that threshold stayed stable, with no sign of worsening. A multilevel model showed the same downward trend over time. Because the study is naturalistic with no control group, the authors are explicit that improvement cannot be attributed to Ash alone and that regression to the mean and natural recovery cannot be ruled out.

Engagement differed in rhythm rather than volume: perinatal users took longer breaks between sessions and used the platform on fewer days but sustained it over a longer span, and their first-to-second-week retention was higher than the general cohort (80.7% versus 74.3%), consistent with on-demand support fitting around caregiving. Safety-relevant content appeared in 11% of sessions, and a blinded clinician audit found that Ash escalated to crisis resources appropriately in every high-risk and crisis case, with no confirmed safety gaps. The authors note that perinatal safety monitoring needs population-specific calibration, for instance distinguishing unwanted intrusive thoughts of harm from genuine suicidal intent.

The authors are clear about the study's limits: it is observational with no comparison condition, the sample is self-selected and skews higher-distress, emotion coding relied on LLM judgments that still need broader validation against human raters, concurrent use of conventional care was not captured, and Slingshot AI developed the tool under study. They frame the work as a foundation for prospective trials and for perinatal-specific AI design rather than as causal evidence.