Mental health chatbots

In the preprint, with four of six authors from Slingshot AI, the Ash model without its safety classifier produced harmful responses to 0.4% of self-harm and suicide prompts, 8.4% of eating disorder prompts and 9.9% of substance use prompts, against 12% to 58% for three GPT-5 models. In 20,000 sampled conversations from September and December 2025, clinicians reviewed sessions flagged by a language model; crisis resources were missing in three of 80 confirmed self-harm or suicide sessions the classifier had missed, all involving self-injury.

In a 2026 evaluation of crisis features in six apps, by two authors affiliated with Elomia Health, Ash scored 25.0 of 65 points, in the poor range. The registered trial plans to enroll about 500 new users with at least moderate depressive symptoms.

In nine services with 64,862 patients, assessments of patients referred through Limbic Access averaged 41.6 minutes against 54.4 minutes, waits for assessment were 2.2 days shorter, and reliable recovery was 58% against 27.4%. In a study of 129,400 patients, referrals rose 15% at services using the chatbot against 6% at control services, including 179% among nonbinary and 29% among ethnic minority patients. An independent early value assessment for NICE found the evidence non-comparative in the main and identified no information on quality and accuracy.

In a survey of 131 practitioners, clinicians reported higher wellbeing and lower cognitive load for assessments with Limbic reports. In a randomized trial of 540 U.S. adults recruited online, the Limbic Care app increased engagement 2.4 times over digital workbooks, with comparable anxiety and depression outcomes.

In the first randomized trial, 70 students recruited online used Woebot or read an ebook for two weeks; depression scores fell only with Woebot. In the postpartum trial, women were enrolled before discharge at a Stanford hospital; PHQ-9 scores fell 1.32 points with the chatbot against 0.13 with usual care, but Edinburgh Postnatal Depression Scale scores did not differ, and baseline scores were below screening thresholds.

Among 141 adolescents seeking treatment at a children's hospital, the app was noninferior to telehealth group therapy on depression at four weeks, a secondary aim. For substance use, the app cut use occasions more than a waiting list in 180 adults, but not more than emailed psychoeducation in 258 adults. Woebot Health staff were authors on these trials. In an independent trial of 105 adults with tinnitus recruited online, Woebot and a tinnitus chatbot both improved anxiety and depression more than a waiting list.

In the NHS trial, 2,161 patients were screened and 76 randomized 2:1 to Wysa or a control arm; 30 were lost to follow-up, and the authors concluded there was no evidence that Wysa treats depression. Two of its nine authors listed Wysa affiliations. In the online trial, with 34 participants per arm over four weeks, stress did not change in either group.

At an academic orthopedic clinic, 51 patients with elevated depression or anxiety scores given Wysa improved more on depression, pain interference and physical function scores at two months than 51 patients in usual care, in a retrospective, nonrandomized comparison. An earlier company analysis of app users compared frequent with infrequent users. An independent review for NICE found the evidence for Wysa's NHS referral assistant less robust than that for Limbic Access.