GreyMatter Capital (Andrew Barr and Danish Munir)

Please find our comments attached.

Oct 19, 2026

To whom it may concern,

GreyMatter appreciates the opportunity to comment on the FDA’s discussion paper,
“Considerations for the Regulation of Generative AI-Enabled Medical Devices.” GreyMatter is a
venture capital firm built exclusively for mental and brain health, investing across AI-driven care
delivery, neurotech, and therapeutic discovery, and drawing on a network of more than 3,000
founders, payors, researchers, and policy experts. We support the FDA’s risk-based, total product
life cycle approach and offer five recommendations specific to generative AI tools for behavioral
health.

1. Set distinct expectations for clinician-facing and patient-facing tools

Clinicians reviewing AI output can often recognize an inaccurate or unsafe response; patients
frequently cannot, and should not be expected to verify the clinical accuracy of AI-generated
guidance. Patient-facing tools may therefore warrant stronger evidence and safeguards. Those
requirements should remain proportionate to risk. Many people cannot reach a therapist because
of cost, location, or workforce shortages, and well-designed, well-tested patient-facing tools can
help close that gap. We believe that requirements should not block access to tools that have
demonstrated safety and benefit.

2. Recognize a third category: patient-facing tools used under clinician oversight

Many behavioral health tools are used directly by patients while a licensed clinician monitors
progress, reviews flagged interactions, and intervenes when needed. Neither “clinician-facing” nor
“patient-facing” captures this model. We recommend the FDA treat clinician-supervised use as its
own category, with expectations that reflect the added safeguard of oversight, while requiring
evidence that the oversight is real: who reviews what, how quickly, and with what authority to act.

3. Draw clear lines between purpose-built and general-purpose chatbots

General-purpose chatbots are widely used for mental health support even though they were not
designed or validated for it. If purpose-built behavioral health tools face rigorous requirements
while general-purpose tools performing similar functions face none, the framework will reward
building general tools that avoid oversight, the opposite of its intent. We urge the FDA to define
clearly which functions, claims, and patterns of use bring a tool within scope, and to make the
purpose-built pathway predictable, proportionate, and worth pursuing, so developers who invest
in clinical validation are not disadvantaged.

4. Support a shared, independent crisis-response benchmark

Every mental health chatbot must respond appropriately when a user is in crisis, including suicidal
ideation, self-harm, or risk of harm to or from others. Today, developers largely test this with their
own methods, making results hard to compare and hard to trust. No single company can build a
benchmark that clinicians, the field, and the public will accept as neutral. We encourage the FDA to
support a common, independently governed crisis-response benchmark, developed with clinicians,
researchers, developers, and people with lived experience, that tests multi-turn conversations and
the full escalation workflow, not just detection. FDA recognition of such a benchmark, for example
as a consensus standard, would give developers a clear target and the public a meaningful signal.
5. Establish a shared definition of a reportable event

Existing adverse event reporting was not designed for conversational AI, where harm may emerge
gradually over a conversation rather than from a single malfunction. Developers lack a common
understanding of what counts as a reportable problem, such as a missed crisis signal,
reinforcement of harmful behavior, or guidance outside the tool’s intended role. We recommend
the FDA work with stakeholders to define reportable events for generative AI behavioral health
tools, specify where and how they are reported, and clarify how reports inform corrective action.
Consistent definitions would make postmarket data comparable across products and surface
problems earlier.

Generative AI can expand access to behavioral health support, but only if the regulatory
framework rewards developers who build and test tools for clinical use. We welcome the
opportunity to work with the FDA on these recommendations.

Sincerely,

​
Andrew Barr and Danish Munir,

Founding Partners at GreyMatter Capital​