JMIR Publications

ABSTRACT

Generative artificial intelligence (GenAI) chatbots reach millions seeking mental health support, often without clinical oversight, validated efficacy, or adverse-event reporting. Existing frameworks, including pharmacovigilance, device regulation, and professional misconduct, address parts of the problem but are not configured for harm involving diffuse accountability, difficult attribution, population-scale exposure, and cost incentives favouring deployment. Synthesising evidence across patient safety, privacy and liability, therapeutic alliance, and algorithmic bias, we extend iatrogenesis to GenAI, proposing four overlapping harm dimensions arising under two conditions that impede governance, detection, and redress. A provisional four-layer theory of change links political economy and deployment conditions to iatrogenic mechanisms and harm outcomes. Mandatory adverse-event reporting, premarket assessment, liability standards, and longitudinal harm surveillance are the most urgent priorities. A provisional Pre-Deployment Risk Screen, seven knock-out criteria and a red-flag component, offers a hard gate for research and funding decisions and a template for regulators, procurers, and accreditors pending validation.