Hims & Hers Health Inc.

Please see the attached comments of Hims & Hers Health, Inc. on Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices — Discussion Paper and Request for Feedback.

Hims & Hers Health, Inc.​ ​ ​ (415) 636 9654
2269 Chestnut Street, #523​ ​ ​ hims.com | forhers.com
San Francisco, CA 94123 ​

Public Comments of Hims & Hers Health, Inc.
Submitted to the
U.S. Food and Drug Administration
Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices — Discussion Paper
and Request for Feedback
Docket No. FDA-2026-N-7874
Submitted: October 19, 2026

I.​ Introduction

Hims & Hers Health, Inc. ("Hims & Hers" or the “Company”) respectfully submits these comments
in response to CDRH's discussion paper on the regulation of Generative AI ("GenAI")-enabled
medical devices. These comments are general in nature and do not make any representations with
regard to a particular Hims & Hers technology, device, product or service. We support FDA's efforts
to develop a risk-proportionate, least-burdensome approach that reflects both the promise of these
technologies and the responsibility their use requires.

Hims & Hers operates the leading health and wellness platform, connecting patients across all 50
states with licensed healthcare providers who evaluate and provide ongoing care, as appropriate. We
have built our platform on the conviction that expanded access to care and rigorous evidence are not
competing values, but complementary imperatives.

Getting FDA’s framework right is essential to fostering innovation, equipping patients and providers
with the best tools, and ensuring that GenAI-enabled medical devices perform as intended. Our
comments are intended to convey the following:

●​ Shift the evidentiary burden from premarket testing to postmarket monitoring,
recognizing that static premarket snapshots have limited utility in capturing full
GenAI behavior, whereas continuous postmarket review utilizes real-world clinical
data to evaluate how a device actually performs so sponsors can respond accordingly.
●​ Formal recognition of high-yield, scalable validation methods, including automated
evaluation using validated LLM adjudication, sampled human review, shadow
deployment, and simulated patient encounters and virtual patient avatars as sufficient
clinical evidence.
●​ Authorization of Predetermined Change Control Plans (PCCPs) based on pre-cleared
benchmarks, allowing sponsors to validate and deploy both internal, sponsor-initiated
 device modifications and third-party foundation model updates without requiring
repeated premarket submissions post-launch of the device.
●​ Refinement of the proposed risk framework to explicitly lower a function’s placement
on the consequences axis when the system incorporates output traceability,
downstream safeguards, formally verified components, and architectures that allow
for human review before a problem can occur.
II.​ Where the Evidence Should Come From
For GenAI-enabled medical devices, shifting the primary evidentiary burden to the postmarket
setting is the most practical and scientifically sound approach. We agree with the paper's core
instinct: premarket testing cannot cover every open-ended input a GenAI-enabled function will see,
and a single premarket snapshot will inevitably miss the failure modes that matter most. Evidence
gathered once, before launch, cannot do the whole job. To truly understand how a device performs,
evidence should be gathered continuously in the real-world population the device actually serves.
Therefore, we support moving a meaningful share of the evidence burden to the postmarket setting as
the default framework. While a prespecified premarket safety benchmark remains the necessary
initial baseline, the balance of clinical confirmation evidence should shift to the product’s total
lifecycle.

This shift is appropriate provided a sponsor’s postmarket monitoring program demonstrates the
capability to promptly detect and contain a problem, proportional to the device’s functional risk. A
monitoring program that only tells a sponsor something went wrong long after a problem arises does
little to limit that problem’s effects. We recommend the Agency make this the actual test for whether
shifting the evidentiary burden postmarket is appropriate: focusing on whether the sponsor can
demonstrate timely, reliable detection and a real way to contain the problem.

Because GenAI device functions and their associated risk will continuously evolve, the mechanisms
for achieving this detection and containment should remain flexible. Depending on the risk profile,
sponsors might utilize options such as periodic re-benchmarking against the original premarket
baseline, the ability to roll back, contain, or stage deployment once a departure from expected
performance is detected, and adaptable corrective actions. Periodic human review of the sampled real
outputs and validated LLM adjudication, both described below, are practical, scalable ways to detect
systematic problems.

Because detection depends on monitoring performance, we also recommend that the Agency allow
sponsors flexibility in demonstrating how they will obtain visibility into real-world device behavior.
As the paper recognizes in Section VI.B, clinicians, healthcare institutions, and others each play a
role in monitoring. Depending on the deployment model, this visibility may come through a
sponsor’s own care delivery, data-sharing arrangements, or by equipping the deploying healthcare
institution with the necessary tools to monitor the device locally. A sponsor seeking a reduced
premarket evidentiary burden should also accept corresponding obligations in proportion to the
device’s risk profile and the sponsor’s role in care delivery, such as prespecified monitoring
thresholds and cadence, as the paper outlines in Section VI.A.

III.​ Formal Recognition of Scalable, High-Yield Validation Methods
Prospective clinical trials sit at the top of this framework's evidentiary ladder, and they are also
among the least suited tools for this category. A trial runs in a controlled and static environment, on a
limited population, at high cost, and none of that reflects how a GenAI-enabled function will actually
behave once it meets the real population it serves. We welcome the paper’s recognition that clinical
confirmation does not need a prospective study in every case. We ask the Agency to carry that logic
further, and stop defaulting to prospective trials simply because they are the most familiar tool, and
instead recognize a set of tools that are highly scalable and more informative for generative models.

Automated evaluation systems utilizing a validated LLM adjudicator grade far more cases than a
human panel, at a fraction of cost, when checked against a smaller human-reviewed sample and a
prespecified acceptance threshold. Periodic, sample-based review of real postmarket outputs by a
human reviewer is a complementary check worth recognizing in its own right, distinct from an LLM
adjudicator, since the two are likely to catch different kinds of problems. Shadow deployment,
running a device silently alongside real care, without acting on its outputs, should not require a
blinded comparison arm to count as meaningful evidence. Blinded designs are not standard
commercial practice, and requiring one here adds cost without adding much insight relative to what a
well-designed shadow deployment already shows. And, simulated patient encounters and virtual
patient avatars, which the paper itself already recognizes as legitimate tools, deserve a clearer, more
affirmative role in the primary evidentiary toolkit.

We ask the Agency to formally recognize this real-world-representative, high-volume, highly
scalable evidence as primary evidence, not a lesser substitute for a trial. Accordingly, the lowest level
of evidentiary rigor sufficient for a function's risk profile should be the presumptively acceptable
evidentiary path, unless there is a specific risk-based reason to require more resource-intensive
methods, provided the sponsor's postmarket monitoring plan can detect and contain a problem. For
lower-risk functions, retrospective evaluation on real patient inputs and shadow deployment should
be recognized as presumptively sufficient clinical confirmation. Finally, any method of clinical
confirmation should use a pragmatic baseline, where a device is used alongside a clinician,
performance should be assessed for the clinician and device working together, and more generally, it
should be assessed against the care likely to occur without use of the device, including the delays and
limitations on access to care.
We propose the Agency establish a dedicated regulatory sandbox and pilot program. This
collaborative environment would allow the FDA to gain familiarity with and recognize novel
methodologies. This sandbox would help both the FDA and industry learn, gain practical experience,
and develop these tools. We invite the Agency to actively work with companies within this pilot
program to co-develop these validation frameworks, ensuring they are robust, pragmatic, and capable
of keeping pace with the rapid evolution of generative AI.

IV.​ Authorization of PCCPs Based on Pre-Cleared Benchmarks

The Predetermined Change Control Plan (PCCP) concept was built for a world where a sponsor can
specify, in advance, exactly what will change. GenAI-enabled devices increasingly do not work that
way; a sponsor cannot predict every internal orchestration refinement, nor can they predict what a
foundation model vendor will change next. But, a sponsor can build a robust performance benchmark
that any acceptable version of that model or overall device has to pass.

We recommend CDRH let a PCCP be built around a pre-cleared benchmark rather than a
pre-specified list of changes. Once the Agency agrees a given benchmark is the right test, any update
that passes it should be treated as already within the sponsor's approved change control plan, whether
the modification is an internal, sponsor-initiated device update or a third-party foundation model
change. Provided the change does not introduce a new material risk or increase the severity of an
existing one, passing the pre-cleared benchmark should allow the sponsor to document and deploy
the modification under its change control plan without requiring a new submission.

We would ask that these benchmarks be specific to the function and clinical action being tested,
rather than a single centralized one applied across every product, since the competencies that matter
for one clinical task are not the competencies that matter for another. As the paper itself recognizes in
Section V. B, publicly available benchmarks are prone to data contamination, saturation, and limited
real-world representativeness. Sponsor-developed benchmarks tailored to a specific device and its
functions therefore have an important role, despite concerns about independence, provided their
development and validation methodology is transparent and prespecified. Benchmarks used this way
should be refreshed on a regular cadence, with a meaningful share of test scenarios rotated each year,
and retired after a defined period, so a model cannot memorize them.

V.​ Evaluate the Function, Not the Product

A single product can combine functions with very different risk profiles: one function that only
surfaces information, and another that recommends or initiates a specific action. We recommend the
evidentiary burden, at every stage, be set at the level of the individual function rather than at the
product as a whole, so a lower-risk function is not held to the standard of the highest-risk function in
the same platform. We would also ask the Agency to recognize that a function whose output can be
traced back to its specific basis, and hybrid architectures where GenAI device function is bounded by
a deterministic or formally verified component whose behavior can be mathematically proven rather
than only sampled and tested, should be able to carry a correspondingly lighter evidentiary burden
for that specific element of the device.

On the risk framework itself, we ask the Agency to build traceability and downstream safeguards
explicitly into the risk framework, because both fundamentally change how much harm an incorrect
output can do. On traceability, the paper already recognizes that a function moves higher on the
consequence axis when a user cannot independently evaluate the basis of an output, as with
measurement and signal-processing functions, in Section IV. The reverse should hold, when an
output cites the specific criterion and record it relied on, an error can be caught before anyone relies
on it, and the function should sit lower on the consequences axis for a given level of accuracy.

On downstream safeguards, we ask the Agency to recognize checks and controls that operate after a
function produces an output and before it affects a patient. While textual disclaimers provide helpful
context and set important user expectations, deterministic software guardrails, such as hard-coded
clinical rules that intercept and block an LLM’s output before it can dictate patient care, provide a
structural layer of safety that actively prevents severe outcomes. The same output, with its potential
harm contained by such safeguards, presents a lower effective consequence than it would without
them and the framework should explicitly reflect this. For example, an eligibility determination that
cites the specific criterion and record excerpt it relied on, and must clear a deterministic check before
it takes effect, carries a measurably lower consequence than the same determination lacking those
controls.

The paper notes that the risk of an action taking function is moderated by the severity of its
consequences. We agree with that, and ask the Agency to confirm that even fully autonomous
action-taking functions with limited consequences, such as continuing an established therapy within
defined and hard-coded clinical parameters, are appropriately placed lower on the grid than those
initiating high-risk treatment.

VI.​ Conclusion

We appreciate CDRH's decision to seek stakeholder input on a framework still under discussion. We
support its goal of a risk-based, least burdensome approach across the total product life cycle, and its
willingness to consider greater reliance on postmarket monitoring. As CDRH continues to develop
its thinking, we recommend above all that any shift of the evidentiary burden to the postmarket
setting be tied to a flexible, tailored postmarket program that detects and contains problems. By
formally recognizing automated validation tools, allowing benchmark-driven change control for both
internal and external updates, and crediting the structural safety of deterministic guardrails, the
Agency can establish a framework that protects patients while enabling the next generation of digital
health innovation. Hims & Hers welcomes the opportunity to continue working with CDRH as this
discussion develops.