Today, artificial intelligence (AI) broadly supports clinical decision-making, with software offering diagnostic assistance, treatment recommendations, and more. The U.S. Food and Drug Administration (FDA) has authorized over 1,900 digital health devices; yet that number may still pale in comparison to the number of apps promoted to the public for management of one’s own health (such as Noom, Calm, or Earkick).

This rapid innovation has the potential to improve individual patients’ lives and population health. At the same time, insufficient oversight allows products to reach users without adequate guardrails or beyond the extent to which they are validated. So as medtech and software companies forge ahead to capture the excitement of AI, regulators struggle to find the right balance. Particularly when it comes to mental health and protecting the more vulnerable, what constitutes appropriate regulation of AI – enough, not too much, and the right kind? The emerging regulatory landscape – with federal agencies, the White House, and state enforcers each addressing different dimensions of the issue – leaves uncertainty for product developers, patients, and providers.

Amid significant unmet need for quality mental health services in the U.S.—a report notes that approximately 42% of high school students experienced persistent feelings of sadness or hopelessness (CDC, Youth Risk Behavior Survey, 2023)—chatbots and other AI-based applications proclaiming to guide and assist users facing anxiety, depression, stress, and more are now available to anyone with a computer or smartphone. This enables more people to attain mental health support; yet there is a wide range in the quality and reliability of these products. Some are clinically validated and FDA-authorized, but many are not, while still claiming to provide clinical benefit. This morass is compounded by the impossibility of truly assuring the safety or reliability of an AI-based product, since there may be infinite potential permutations or outputs. Furthermore, mental health issues vary widely in severity; some individuals may be well-supported by a digitized coach, whereas others require more extensive evaluation and support that should still be conducted—or at least overseen—by a licensed clinical practitioner.

The FDA, which reviews and authorizes software that meets the definition of a medical device, aims to support innovation, but its foremost responsibility is to protect public health and patient safety. FDA has thus taken a cautious approach to AI-based medical products, wary of software that may lead individuals to forgo care from a trained practitioner. During its November 2025 Digital Health Advisory Committee meeting on generative AI mental health products, FDA stressed that GenAI’s benefits are accompanied by significant risks, particularly in mental health, flagging hazards such as failure to detect suicidal ideation, misidentification of medical symptoms, and the formation of unhealthy parasocial relationships with AI systems. The Committee actively deliberated over what clinical evidence should be required to demonstrate safety, effectiveness, and generalizability; which products need authorization at all; and how manufacturers should monitor software that adapts after deployment.

Into the mix, enter the March 2026 White House National Policy Framework for Artificial Intelligence. The Framework expresses the current Administration’s strong support for AI-based innovation and preference for “lighter touch” regulation. It calls on Congress to ensure that technology and AI providers do not face content censorship and to establish regulatory sandboxes supporting American creativity and leadership in AI development. It also urges broad preemption of states’ AI regulation to preclude a “fragmented patchwork” of rules, while simultaneously discouraging over-reaching federal requirements. Thus, while the Framework does not directly address the use of AI in mental health, it appears in spirit to favor business interests over FDA’s protective posture.

However, around child safety, the Framework prominently supports active regulation of AI. Specifically, it suggests that AI platforms and services likely to be accessed by minors be required to implement features reducing the risks of self-harm and sexual exploitation, and that states remain able to enforce their own generally applicable laws protecting children. In this respect, the Framework sketches a regulatory floor without a clear ceiling, and implementation of its various proposals may ultimately prove more burdensome for AI mental health developers than FDA oversight. There is already prominent litigation against AI manufacturers, such as the Character.AI wrongful death suit and various state Attorney General actions, alleging that these tools encourage self-harm and act as unlicensed medical professionals; the Framework would not appear to lessen such scrutiny moving forward.

It is generally agreed that AI can serve a broad range of mental health-related purposes, and that a tool’s indications for use and claims inherently impact its risk/benefit profile and the extent of regulation it warrants. Similarly, few object to the notion that AI tools should be monitored across their life cycle for safety and continued applicability. Yet there is still debate over exactly which AI-based products should trigger FDA regulation. The FDA’s guidance documents continue to evolve, while lacking definitive commentary on GenAI in particular.

Given the difficulty of fitting rapidly iterating software into a regulatory framework designed for hardware devices, many AI manufacturers try to avoid FDA regulation, innovating around clinical decision-making to position their products as “general wellness” or non-device clinical decision support (CDS). However, the burden of FDA regulation carries a countervailing benefit: recognition of having undergone a reasonably rigorous review. For example, the added requirements imposed by the FDA on devices marketed for use by children provide some reassurance that authorized tools used by kids for mental health support are safe and reliable. 

Meanwhile, AI-based mental health tools that stay outside of the FDA’s purview lack defined state or federal requirements, making it unclear how – or if – they will engender comparable trust. The absence also leaves developers less able to mitigate business risk and patients and providers without a standard by which to evaluate these tools. Without coordinated federal or industry-driven action, the regulatory vacuum is likely to be filled by state-by-state litigation and enforcement—precisely the “fragmented patchwork” that neither the White House nor industry welcomes.