Fang Zhang

Background: Generative artificial intelligence (GAI) chatbots are frequently used for mental health–related questions, including suicide-related queries. User interfaces (UIs) may include additional safety controls not present in direct application programming interface (API) access (eg, moderation rules that flag or block harmful content, crisis refusal templates, or system-level instructions). Objective: This observational cross-sectional study aimed to examine how frequently AI chatbots provide direct responses to suicide-related prompts across access modes and whether response patterns differ between consumer-facing UI and API access across models and prompt risk levels. Methods: We evaluated how five widely used consumer AI models (ChatGPT, Claude, Gemini, Grok, and Llama) responded to 30 previously vetted suicide-related prompts spanning five clinician-assigned risk levels. Each prompt was submitted 100 times through both public-facing UI and direct API access, yielding 30,000 total responses. The primary outcome was whether a response directly answered the prompt (ie, provided specific information or guidance related to the question asked). AI responses were categorized using a blinded large language model–based classifier (GPT-4o mini), validated against independent human coding (289/300, 96.3% agreement; Cohen κ=0.89). We estimated mixed-effects logistic regression models that predicted a direct response. The primary predictors were the AI model, access mode (UI vs API), and prompt risk category. Results: Of the 30,000 responses, 69.8% (n=20,934) were direct. Direct responses were more common through APIs than UIs (n=11,609, 77.4%, vs n=9325, 62.2%); however, Gemini had slightly higher direct response rates via UIs than APIs (n=2301, 76.7%, vs n=2131, 71.0%). Differences in the likelihood of a direct response were most pronounced for higher-risk prompts: For very-high-risk prompts, 24.8% (n=743) of API responses were direct compared with 4.6% (n=138) of UI responses, whereas for high-risk prompts, 80.1% (n=2002) of API responses were direct compared with 48.4% (n=1209) of UI responses. Claude and Gemini had the highest direct response rates (n=4688, 78.1%, and n=4432, 73.9%, respectively) across UI and API access modes, whereas ChatGPT, Grok, and Llama had lower rates (n=3878, 64.6%; n=3870, 64.5%; and n=4066, 67.8%, respectively). In mixed-effects models, UI access was associated with lower odds of a direct response than API access (odds ratio 0.09, 95% CI 0.08-0.10). Higher prompt risk was associated with lower direct response probability, and access mode differences varied by risk level and model. Conclusions: AI safety depends on how users access the model. Safety observed in a UI should not be assumed to generalize to API access or apps built on APIs. Evaluations and policies should consider access modes to ensure safety protections for individuals interacting with AI chatbots.

Introduction: Investment in mental health treatment for children has grown substantially, but whether this reflects more treatment recipients, higher spending per user, or both has not been examined. To estimate trends in mental health expenditures, probability of use and expenditures per user for US children aged 2-17 from 2004 to 2023. Methods: Two-part models were estimated to predict inflation adjusted health expenditure data from the 2004-2023 Medical Expenditure Panel Survey (unweighted N = 153 622) to assess overall trends and variation by age-group and insurance coverage. Results: Total annual mental health expenditures increased from $10.9B to $37.9B, rising from 4.4% to 8.1% of total health expenditure per child (mean individual-level ratio among children with any health care expenditure). The probability of mental health service use more than doubled (6.3% to 13.0%, +0.35% points/year), while expenditure per user grew from $2816 to $5178 (+3.3%/year). Discussion: Relative to the 2004 baseline those ages 2-5 had the greatest growth in both service use and expenditure per user. Expenditure per user grew substantially faster for privately-insured than for Medicaid-covered children. Conclusions: Child mental health treatment expenditures have grown dramatically over the past 2 decades, with particularly large increases for those ages 2 to 5.

PURPOSE: An increasing number of oral targeted anticancer medications (OTAMs) are approved with uncertain evidence of clinical benefits, marketed at increasing prices, and subject to insurance coverage mandates. We determined how use of and spending on OTAMs varied by evidence and magnitude of clinical benefit. METHODS: Using claims data between 2016 and 2021 from a large, national insurer, we assessed volumes of use and spending on OTAMs approved by the (US) Food and Drug Administration (FDA) between 2010 and 2021. We categorized drug use by evidence of overall survival (OS) and progression-free survival benefits for each FDA-approved indication using FDA-approved drug labels. We assessed how use in terms of days supplied, spending, and patients exposed was related to the level of evidence available for specific drug indications. RESULTS: Use of OTAMs increased from 50.7 patient-days supplied per 1,000 enrolled patients in 2016 to 120.4 patient-days supplied in 2020 (last full year of data). Approximately half of the dispensings (46.8% in 2016 and 55.6% in 2021) were for patients with indications for which the drugs had label-documented evidence of OS benefit at the time of dispensing. Spending on OTAMs increased from $394 million US dollars (USD) in 2016 to nearly $1 billion USD in 2020. Spending for indications with evidence of OS benefit fluctuated between 41.7% and 47.1% of total spending across all OTAM dispensings. Spending on off-label use was consistently 11%-12% of all OTAM spending annually. CONCLUSION: Between 2016 and 2021, the largest shares of OTAM dispensings and spending were for indications with documented evidence of OS benefits. Nevertheless, significant spending and use were observed for drugs without confirmed OS benefit.

BACKGROUND Generative AI chatbots are frequently used for mental health-related questions, including suicide-related queries. Prior evaluations have focused on user interfaces (UI), which may include additional safety controls not present in direct application programming interface (API) access. OBJECTIVE To examine how frequently AI chatbot models provide direct responses to suicide-related prompts across access pathways and to determine whether safer behavior is intrinsic to the model or depends on UI-level safeguards. METHODS Observational cross-sectional study conducted in February and March 2026. We evaluated how five widely used consumer AI models (ChatGPT, Claude, Gemini, Grok, and Llama) responded to 30 previously vetted suicide-related prompts spanning five clinician-assigned risk levels. Each prompt was submitted 100 times through both public-facing UI and direct API access, yielding 30,000 total responses. The primary outcome was whether a response directly answered the suicide-related prompt. A direct response was defined as one that provided specific information or guidance related to the question asked, rather than refusing to answer, redirecting the user to a crisis resource, or responding only with general safety language. AI responses were categorized using a blinded large language model–based classifier. We estimated mixed-effects logistic regression models that predicted a direct response. The primary predictors were the AI model, access mode (UI vs. API), and prompt risk category. RESULTS 69.8% of responses to the suicide prompts were direct. Direct responses were more common through APIs than UIs (77.4% vs. 62.2%). Differences in the likelihood of a direct response were most pronounced for higher-risk prompts: among very high-risk prompts, 24.8% of API responses were direct compared with 4.6% of UI responses; among high-risk prompts, 80.1% of API responses were direct compared with 48.4% of UI responses. Claude and Gemini had the highest direct response rates (78.1% and 73.9%), whereas ChatGPT, Grok, and Llama had lower rates (64.6%, 64.5%, and 67.8%). In mixed-effects models, UI access was associated with lower odds of a direct response than API access (odds ratio, 0.09; 95% CI, 0.08-0.10). Higher prompt risk was associated with lower direct response probability, and access-mode differences varied by risk level and model. CONCLUSIONS AI safety behavior depends on how the user accesses the model. The safety observed in an AI chatbot’s UI should not be assumed to generalize direct model access through APIs or to downstream applications built on those APIs. Evaluations and policies must consider access channels to ensure comprehensive safety protections for all individuals interacting with AI chatbots.

PURPOSE: State legislations mandating prescriber use of Prescription Drug Monitoring Programs (PDMPs) may have the unintended consequence of restricting opioid analgesics to patients dying of cancer. This study aims to assess associations of comprehensive PDMP mandates with opioid-related outcomes for patients dying of cancer, overall and by decedent race and ethnicity. METHODS: Study population were Medicare decedents who were age 66 years or older, diagnosed with breast, colorectal, lung, or prostate cancer, and died of cancer in 2011-2019. This cross-sectional study used SEER-Medicare data and a difference-in-differences design. Study sample included decedents from 10 states with an operating PDMP on January 1, 2011. Outcomes included the dichotomous event of having one or more opioid days, total and daily morphine milligram equivalents (MMEs) if having opioids, near the end of life. Generalized linear models were estimated for dichotomous (logit link function) and continuous (log) outcomes. RESULTS: This study included 115,256 decedents. Comprehensive PDMP mandates were associated with modest reductions in the rate of one or more opioid days (from 45.1% to 43.9%, difference = 0.011 [95% CI, −0.019 to −0.003]), total dose (from 1,600.6 to 1,521.0 MMEs, difference = 79.6 [95% CI, −131.5 to −27.6]), and daily dose from all opioids (from 75.7 to 72.9 MMEs, difference = 2.7 [95% CI, −5.1 to −0.4]). Compared with non-Hispanic White decedents, Black decedents experienced a four-fold reduction, and Asian/Pacific Islander decedents experienced a two-fold reduction, in the rate of one or more opioid days. CONCLUSION: Comprehensive PDMP mandates were associated with modest reductions in opioid analgesics dispensed to Medicare patients dying of cancer. Non-Hispanic Black and Asian/Pacific Islander decedents experienced larger reductions.