Nicholas C. Jacobson and Michael V. Heinz

Comment from Nicholas C. Jacobson, PhD, and Michael V. Heinz, MD (Geisel School of Medicine at Dartmouth). We are the developers of Therabot and ran the first randomized controlled trial of a fully generative AI therapy chatbot. We support most of the discussion paper and urge CDRH to answer Question 18 with a yes: for patient-facing conversational devices whose errors can be caught and corrected under monitoring, accept a risk-proportionate premarket package in exchange for a prespecified, enforceable postmarket program. The attached comment also addresses layered postmarket monitoring and supervisory agents (Q19 to Q21), process-based change control and PCCPs (Q22 to Q24), comparators and mental-health-specific benchmark elements (Q7 to Q16), the risk framework (Q3, Q5, Q6), and foundation model master files (Q25). Full comment attached.

Nicholas C. Jacobson, Ph.D. Center for Technology and Behavioral Health
Associate Professor of Biomedical Data Science, 46 Centerra Parkway, Suite 300
Psychiatry, and Computer Science Lebanon, NH 03766
Geisel School of Medicine at Dartmouth Nicholas.C.Jacobson@dartmouth.edu
(603) 646-7037
Michael V. Heinz, M.D.
Assistant Professor of Psychiatry
Geisel School of Medicine at Dartmouth
Dartmouth Health

October 19, 2026

Dockets Management Staff (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback

Introduction and summary
We write as the developers of Therabot, a generative AI psychotherapy system, and as the investigators
who ran the first randomized controlled trial of a fully generative AI therapy chatbot for clinical-level
mental health symptoms (Heinz et al., 2025). Dr. Jacobson is Associate Professor of Biomedical Data
Science, Psychiatry, and Computer Science at Dartmouth's Geisel School of Medicine and directs the
AIM HIGH Laboratory at the Center for Technology and Behavioral Health. Dr. Heinz is an assistant
professor of psychiatry at Dartmouth and Dartmouth Health, co-leads the Therabot clinical trials, and
treats adults with mood and anxiety disorders.
Dr. Jacobson was an invited speaker at the November 6, 2025 meeting of the Digital Health Advisory
Committee on generative AI-enabled digital mental health devices, and together we submitted a written
comment to that docket (FDA-2025-N-2338-0108). This comment builds on both.
We think the discussion paper is a real step forward, and most of it is right. Evaluating the final userfacing device rather than the foundation model, scaling evidence to risk, allowing clinical confirmation to
take forms other than a large prospective trial, and using the premarket benchmark as the baseline for
later changes are all sound, and several of them answer concerns raised at the 2025 meeting.
Our main point is about where the weight should sit. For conversational devices like ours, premarket
testing can only sample a small part of what the device will say once real people use it. The failures that
mattered in our own trial showed up in real conversations, and we caught them because people
reviewed transcripts as they came in. We therefore urge CDRH to answer Question 18 with a clear yes:
for most patient-facing GenAI devices whose errors can be caught and corrected before harm becomes
fixed, a test we define below that most conversational mental health functions pass under continuous
monitoring and that autonomous functions such as prescribing do not, FDA should accept a lighter, riskproportionate premarket package in exchange for a demanding, prespecified postmarket program with
real consequences attached.
Our recommendations in brief:

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 1
 1. Put most of the evidentiary weight in postmarket monitoring for devices outside the highest-risk
corner of the two-axis framework, and make a monitoring program with defined triggers, data,
and remedies the condition for doing so (Q18).
2. Build postmarket monitoring in layers: automated screening of every interaction by a separately
validated screener, human review of flagged and randomly sampled conversations, periodic rebenchmarking, and outcome tracking, with each layer checking the one before it (Q19 to Q21).
3. Define change control by process rather than by a list of anticipated modifications, so that a
PCCP prespecifies the test battery, thresholds, and rollback rules any change must pass (Q22 to
Q24).
4. Judge performance against the care people would actually receive without the device, which
today is often a general-purpose chatbot or no care at all (Q14, Q15).
5. Add benchmark elements specific to mental health, including fidelity to evidence-based
treatment and whether the device reinforces pathology, sycophancy, or dependence (Q9).
6. Keep the framework from penalizing purpose-built tools relative to the general-purpose models
that already serve as a default mental health resource for many people (Q25).

What the discussion paper gets right
Several choices in the paper match what we have learned building and testing Therabot since 2019.
• Evaluating the deployed device. Our early models, trained on peer-support forum posts and
then on psychotherapy training transcripts, produced responses that were unsafe or unhelpful.
What made the system safe and effective was the curated training data, the guardrails, and the
oversight built around the model. The final configured device (Section V.A) is the right unit of
evaluation.
• Risk-proportionate evidence. The two-axis framework is a reasonable organizing heuristic. As
Question 1 suggests, we would add reversibility and the availability of human follow-up as
modifiers. By reversibility we mean whether an incorrect output can be detected and corrected
before the harm it could cause becomes fixed, which turns on whether the output is observable
after the fact and whether a person can still intervene. A supportive conversation that goes
poorly can usually be corrected the next day; an autonomous prescription often cannot.
Reversibility is not harmlessness, though, and a missed imminent risk is the clearest exception;
we return to it under Question 6.
• Flexible clinical confirmation. Not every device needs a large prospective trial before any
patient uses it (Section V.C).
• Trajectory-level risk and two-sided escalation errors. Mental health treatment unfolds over
many turns and many sessions, so scoring single responses misses most of what matters. The
paper also recognizes that over-escalation has real costs, which is especially true in mental
health.
• The premarket benchmark as the baseline for change. This is the piece that makes a
postmarket-centered approach workable, and we build on it below.
• Manufacturer accountability. We agree that the developer and deployer should carry
responsibility when a tool directly causes harm, as Dr. Jacobson told the Committee in 2025.

Question 18: Most of the evidentiary weight belongs in postmarket monitoring
We support accepting greater premarket uncertainty in exchange for greater reliance on postmarket
monitoring. For conversational mental health devices, this should be the default approach outside the
highest-risk cells of the framework, for four reasons.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 2
First, premarket testing samples a narrow slice of real use. The paper says as much when it notes that it
may be unreasonable to evaluate every conceivable input. Our trial bears this out. Trained clinicians and
researchers reviewed every Therabot response after it was sent. Staff stepped in 15 times for participant
safety concerns and 13 times to correct inappropriate responses, commonly when participants asked for
medical advice outside the system's mental health scope (Heinz et al., 2025). We had framed the model
tightly around mental health and had not anticipated how often people would bring it general medical
questions. None of our premarket test scenarios surfaced that. Reading real transcripts did, quickly.
Second, the technology changes in months, not years. A premarket package built around one model
version goes stale fast. If every meaningful update needs substantial new premarket evidence,
developers will keep shipping older models that are often less capable and less safe. Postmarket
evidence can keep pace with the product in a way premarket evidence cannot.
Third, delay has a cost, and patients bear it. Most people with a mental health disorder in the United
States do not receive minimally adequate treatment, and many already use general-purpose chatbots
for support. In one recent survey, 48.7% of U.S. adults with a mental health condition who had used a
large language model in the past year had used it for psychological support (Rousmaniere et al., 2025).
Each month a purpose-built, monitored device spends in premarket review is a month those users spend
with tools that have no clinical design and no monitoring.
Fourth, the paper's own analogy points this way. The competency-based approach is modeled on how
clinicians are credentialed: examinations first, then supervised practice with progressively greater
independence, then ongoing review. Much of what assures a clinician's safety happens after initial
licensure. A framework built on that analogy should give comparable weight to supervised real-world
use.
Conditions. A reduced premarket package has to come with conditions that make the postmarket
program enforceable. At the 2025 meeting, one public speaker warned that products are too often
"green-lighted ahead of time" with a plan to study them later. We share that concern; a postmarket
commitment without enforcement protects no one. We suggest:
1. A premarket baseline. Prespecified benchmarks passed for the safety elements (S.1, safetycritical recognition and escalation; S.2, scope maintenance and boundary adherence; S.3,
calibration, uncertainty communication, and clinical deferral) at minimum, plus clinical
confirmation proportionate to risk. For a patient-facing therapeutic device, a controlled or
closely monitored prospective study with systematic transcript review is a reasonable floor.
However, a large multi-site RCT should not be required before limited deployment.
2. A prespecified monitoring plan submitted with the application, stating what is measured, how
often, the thresholds that trigger action, and the action each trigger requires.
3. Staged deployment. Availability starts limited in scale or setting and expands as prespecified
safety and performance gates are met, much like the phased go/no-go approach Dr. Patricia
Areán (University of Washington; formerly director of the Division of Services and Intervention
Research at NIMH) described to the Committee in 2025, in which each stage proceeds only after
the prior one has met prespecified criteria. Model updates during staged deployment would
follow the same process-based change control we describe under Questions 22 to 24.
4. Technical capacity to act. Version control, the ability to roll back to the last authorized
configuration within a defined time, and the ability to suspend a function.
5. Reporting. Periodic reports to FDA on safety events, escalations, re-benchmark results, and
changes, plus a public summary of aggregate safety metrics.
6. Consequences. Authorization that FDA can narrow or suspend if the sponsor fails to carry out its
monitoring plan or exceeds prespecified safety thresholds without a corrective response.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 3
Where this approach does not fit. We would not apply it to functions that fail the reversibility test
described above: action-taking functions whose errors are severe and cannot be undone once acted on,
such as autonomous prescribing or medication dose changes, or functions whose errors leave no
observable trace after the fact, since monitoring cannot correct what it cannot see. Within mental
health, a device that independently assigns a diagnosis that drives treatment decisions warrants more
premarket evidence than one that delivers structured psychotherapy with crisis escalation and human
follow-up.

Questions 19 to 21: What a credible postmarket program looks like
We support all three approaches the paper names. We suggest treating them as layers of one program
rather than as alternatives, and adding automated screening of every interaction and tracking of clinical
outcomes.
1. Automated screening of every interaction. In our trial, a separate risk classifier screened
messages for crisis content and triggered an in-app emergency module pointing to 911, the
national suicide hotline, and the Crisis Text Line. Staff also reviewed every message in near real
time. Human review of every message works in a trial but does not scale. Automated screening
of every message, with human review of what it flags, does. The screener then becomes a
component that needs its own validation, which we address under Question 20. This does not
simply move the point of failure. The two models fail in different ways, so one's errors are not
the other's blind spots, and the random-sample review in the next layer catches and counts
what the screener misses, so its sensitivity in deployment is measured rather than assumed.
2. Sample-based independent clinician review. A random sample of conversations, stratified by
risk level, user characteristics, and conversation length, reviewed against prespecified rubrics by
clinicians independent of the sponsor. Random sampling matters because reviewing only flagged
conversations cannot find the failures the screener misses. Review results could also inform
decisions about expanding availability during staged deployment.
3. Periodic re-benchmarking against the premarket battery (adapted as needed), on a fixed
schedule and after any trigger.
4. Drift monitoring of inputs, outputs, and safety-flag rates, including drift in subgroup
performance. A Committee member asked at the 2025 meeting whether bias testing should be
part of drift detection; we think it should.
5. Outcome tracking. For therapeutic devices, routinely collected symptom measures such as the
PHQ-9 show whether users are improving, deteriorating, or not responding. Deterioration is
among the most important safety signals, and transcript review alone does not detect it.
Without a concurrent control group, aggregate outcome data cannot establish effectiveness,
and we do not propose using them that way. Their value postmarket is as a safety signal: an
individual whose scores worsen across sessions should prompt human follow-up whatever the
cause, and a population-level deterioration rate can be compared against the rate observed in
the premarket study and against published rates for untreated or waitlisted samples, with a
prespecified excess triggering review. Engagement patterns matter too, since unusually heavy
use can signal dependence rather than benefit.
Cadence and triggers. Cadence should scale with risk and with how often the product changes. We
suggest three kinds of trigger:
• Event triggers: any change to model weights, foundation model version, system prompts,
retrieval sources, or guardrails.
• Signal triggers: a prespecified rise in flagged-event, deterioration, or complaint rates, or
detected input drift.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 4
 • Calendar triggers: for example, quarterly re-benchmarking and sample review for a patientfacing therapeutic device in its first year, set in the monitoring plan and relaxing as a stable
record builds up.
Machine-based supervisory agents (Q20). Yes, with limits. A machine supervisor is the only practical
way to screen every conversation, and full coverage is worth more than any feasible amount of human
sampling. Our own risk classifier was a non-generative model trained to detect risk across several
psychiatric categories. We suggest four conditions:
• The supervisor is validated on its own, with its sensitivity for safety-critical events measured
against clinician review and reported.
• It is versioned and change-controlled separately, so a device update cannot silently change the
monitor.
• It is built differently from the device, using a different model family, a different developer, or a
non-generative classifier, so the two do not share blind spots. This is the same worry the paper
raises about synthetic data in Question 13.
• It screens and triages; it does not replace human judgment where escalation beyond the
device's built-in crisis resources is needed.
Other stakeholders (Q21). Clinicians and health systems are well placed to report adverse events and to
contribute de-identified cases to shared test sets. Professional societies and standards bodies could
maintain reference scenarios and scoring rubrics for mental health, which would make sponsor
benchmarks comparable. Those reference sets would need to be held in sequestered form and
refreshed regularly rather than published, so that they do not leak into training data (the contamination
problem the paper raises in Question 10). Secured and rotating item banks are how board examinations
handle the same risk. None of this should dilute manufacturer accountability. Dr. Vaile Wright, senior
director for health care innovation at the American Psychological Association, suggested to the
Committee a public repository of FDA-cleared products; adding summary safety metrics to such a
registry would help clinicians and patients tell monitored tools from unmonitored ones.

Questions 22 to 24: Define change control by process
At the 2025 meeting and in our docket comment, we recommended that FDA review GenAI mental
health developers at the organizational level rather than freeze specific model versions. Using the
premarket benchmark as the baseline for re-evaluation gives that recommendation a concrete form.
PCCPs when modifications cannot be fully specified (Q23). Few sponsors can enumerate in advance
which prompts, guardrails, retrieval sources, or model versions will change over the life of a device, so a
PCCP written as a list will always lag the product. Rather than listing future modifications, a GenAI PCCP
should prespecify the procedure every modification must pass: the relevant portions of the benchmark
battery; the acceptance criteria, for example non-inferiority to the authorized version on the safety
elements and no meaningful loss in clinical proficiency or subgroup performance; the sample-based
clinician review that follows larger changes after deployment; and the rollback rule. Because the plan
specifies the test rather than the change, it covers modifications no one anticipated when it was written,
which for this technology is many of them. A change that passes and stays within the authorized
intended use, user population, and activity level could then be deployed under the quality system with
documentation. FDA's role becomes reviewing the procedure up front and auditing that it is followed.
FDA's Software Precertification (Pre-Cert) pilot ran from 2017 to 2022 and tested whether review could
rest in part on a developer's demonstrated organizational excellence, with streamlined review of
individual products. It ran into problems, and FDA concluded in 2022 that it could not be fully
implemented without new statutory authority. Still, its core idea of appraising the developer's process
alongside product-level evidence fits this technology better than version-by-version review.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 5
Scaling evidence to the change (Q22). We suggest three tiers:

Tier Oversight Examples What is required

Tier 1 (minor) Quality system Interface wording, bug fixes, infrastructure Record in the quality
documentation or latency changes that do not affect model management system
outputs, added crisis resources

Tier 2 Re-benchmark New fine-tuning data or model weights, a Pass the relevant
(intermediate) under the PCCP new foundation model version, changes to prespecified tests
system prompts, retrieval sources, or before deployment;
guardrails, whether or not the specific sample review follows
change was anticipated when the PCCP was changes to model
written, provided intended use, user weights or foundation
population, and activity level are model version
unchanged

Tier 3 (major) New FDA review Moving along either axis of the risk Premarket submission
framework: adding diagnosis or medication
recommendations, extending to minors or
a new care setting, removing a humanoversight step

Third-party foundation models (Q24). Silent upstream changes are the largest change-control risk in
this space. Manufacturers should pin model versions and should not deploy on endpoints a vendor can
update without notice. Where a manufacturer depends on a commercial API, contracts should require
advance notice of changes and continued access to the prior version for a defined period, and any
upstream change should trigger re-benchmarking. Fine-tuning and hosting open-weight models directly
puts version control in the manufacturer's hands. Therabot was built this way: we fine-tuned open
models and serve them on infrastructure we control, so the deployed model changes only when we
change it (Heinz et al., 2025). FDA should not mandate an architecture, but it should give credit for
designs that leave the manufacturer in full control of versions.

Questions 7 to 16: Premarket evaluation
Benchmarking followed by clinical confirmation (Q7) is a useful framework. Our comments focus on the
comparator, gaps in the benchmark elements for mental health, the scope of indications, and keeping
the approach accessible.
Comparators (Q14, Q15). We strongly support assessing performance relative to what would happen
without the device. For most people who would use a patient-facing mental health device, the realistic
alternative is a waitlist, no care, or a general-purpose chatbot with no clinical design, not an expert
therapist, and even expert care carries risk of its own. In our 2025 comment we asked FDA to compare
digital therapeutics against foundation models, since those are the default people already turn to, and
that remains our view. A sensible standard for a patient-facing therapeutic device would be superiority
to no treatment or waitlist on clinical outcomes, together with safety performance on the safety
elements (S.1 to S.3) at least as good as widely used general-purpose models on the same scenarios.
Where a clinician panel is the comparator, it should reflect the median clinician in practice rather than
an idealized consensus. Human therapists also miss warning signs and give unhelpful responses. A
standard no clinician meets would work as an access barrier, not a safety standard.
Benchmark elements for mental health (Q9). The proposed elements are a good start. For
psychotherapy devices we would add, or name explicitly:
Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 6
 • Therapeutic fidelity: whether responses follow the evidence-based treatment the device claims
to deliver, as rated by trained clinicians.
• Non-reinforcement of pathology: whether the device validates or amplifies hopelessness,
maladaptive beliefs, disordered eating cognitions, or avoidance. A response can contain nothing
a content filter would flag and still deepen the very pattern treatment is meant to change.
• Sycophancy: agreeing with the user when clinical judgment calls for gentle challenge. The
Committee flagged this directly in 2025.
• Dependence and appropriate disengagement: whether the device encourages real-world
activity and human connection, and can suggest a break or close a session, rather than
maximizing time in the app.
• Role boundaries: not drifting into romantic or companion roles, and being consistently clear
that it is not a human.
Several of these could sit under E.4 or S.2, but naming them would give sponsors and reviewers a shared
vocabulary.
Indications and comorbidity. The paper says benchmarking should cover adjacent clinical domains when
a device's outputs affect comorbid conditions. The same logic should shape indications. Most people
who seek help for their mental health have more than one condition, and a good therapeutic
conversation addresses them together rather than one diagnosis at a time. Diagnostic categories also
overlap heavily. Poor sleep, low energy, trouble concentrating, and restlessness, for example, are
criteria for both depression and generalized anxiety, and users describe them as one experience rather
than two diagnoses. A device indicated for depression alone would have to redirect every time the
conversation turned to worry, leaving the same symptoms in scope one minute and out of scope the
next. Comorbidity was common in our trial, and we analyzed outcomes for depression, anxiety, and
eating disorder risk within the same sample (Heinz et al., 2025). We urge CDRH to allow transdiagnostic
indications tied to the therapeutic approach, for example cognitive behavioral therapy for mood,
anxiety, and related conditions, with benchmarking and confirmation scoped to that indication instead
of separate evidence packages for each diagnosis.
Clinical confirmation (Q11). For patient-facing therapeutic devices, confirmation should include real
users with clinical-level symptoms and systematic transcript review by qualified clinicians, because
benchmarks and standardized patients do not reproduce how people actually describe their distress. A
modest controlled study can supply that evidence, provided the postmarket program described above
follows. Retrospective evaluation and shadow deployment are of limited use for therapeutic chat, since
there is rarely a record of what a good response would have been at that point in that conversation.
Independent third parties (Q16). Sequestered test sets and independent adjudicators would improve
trust in benchmarks, and we support both, with two cautions. Third-party certification should not
become a fee-based gate that only large companies can afford; much of the careful work in this field
comes from academic groups and small companies. And sequestered sets need regular refreshing, since
static test sets leak into training data over time.

Questions 3, 5, and 6: Notes on the risk framework
Patient-facing information (Q3). Patient-facing therapeutic conversation should sit higher on the
consequences axis than clinician-facing information, but the assessment has to include the alternative.
For a person on a months-long waitlist, the realistic choice is often between a monitored therapeutic
device and an unmonitored general-purpose chatbot. Useful safeguards include clear and repeated
disclosure that the user is talking with software (Therabot carries "bot" in its name and shows a robot
avatar in every interaction for this reason), crisis detection with immediate routes to human help, scope

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 7
limits with redirection for medical questions, and in-app symptom monitoring that flags deterioration
for follow-up.
Multi-turn conversations (Q5). We agree that risk should be assessed across realistic conversational
trajectories. For a mental health intervention, the multi-turn exchange is the intervention itself.
Intended use should therefore be defined by the device's therapeutic role and its boundaries, with
testing under the paper's S.2 element (scope maintenance and boundary adherence) that probes drift
across long conversations and across sessions. Premarket simulation covers the trajectories developers
think to write. Postmarket transcript review finds the ones they did not.
Escalation (Q6). Both directions of error matter, and they should be weighed against clinical practice.
Missing imminent risk is the more serious error and deserves the strictest threshold. A device that
answers every mention of sadness or passive thoughts of death with a scripted hotline referral also does
harm: it breaks the therapeutic alliance, teaches users to stop disclosing, and pushes them toward tools
with no safeguards. Clinicians respond to passive ideation with assessment and continued engagement.
Devices should be held to a similar graded standard, with high sensitivity for imminent risk and
proportionate, clinically appropriate responses below that threshold. The paper's own S.1 element
already treats over-escalation as a failure; we would make this graded standard explicit in how S.1 is
scored.

Question 25: Foundation models and the unregulated default
Voluntary Foundation Model Device Master Files (MAFs) would help sponsors and reviewers, and we
support them. As the paper describes them (Section VII.A), these would extend FDA's existing Device
Master File program: a foundation model developer could file model or system cards and related
documentation confidentially with FDA, and a device sponsor building on that model could reference
the file in its own submission with the developer's authorization, without the developer having to
disclose the material to the sponsor. The most useful content for mental health devices would be
version histories and deprecation schedules, update notification commitments, known failure modes in
emotionally charged conversations, and documented behavior on crisis scenarios.
We also ask CDRH to keep the larger picture in view. The general-purpose models that would file these
MAFs already function as one of the largest sources of mental health support in the United States, and
they sit largely outside any device framework because they make no clinical claims. Two consequences
follow for how CDRH approaches this framework. First, a framework that is demanding for purpose-built
tools and silent on general-purpose ones will move users toward the less safe option. Second, FDA
assesses intended use from a product's features as well as its claims, so a general-purpose model that
adds crisis detection risks being treated as a device, which discourages developers from adding it. As we
wrote in our 2025 comment, safety features that carry no therapeutic claim deserve a safe harbor or
lighter oversight. Any benefit-risk judgment about a purpose-built device should account for what
patients would otherwise use, and for mental health, that is often one of these models.

Conclusion
The paper's framework can work well for conversational mental health devices if the weight is placed
correctly. Set a sound premarket baseline, deploy in stages, monitor every conversation in layers, rebenchmark on every meaningful change, and attach real consequences to monitoring commitments.
That approach would get tools built for mental health care to patients sooner, and it would give FDA
better evidence about real-world safety than any premarket package can. We would be glad to share
more detail on our monitoring methods or to take part in further discussion.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 8
Respectfully submitted,

Nicholas C. Jacobson, Ph.D.
Associate Professor, Departments of Biomedical Data Science, Psychiatry, and Computer Science
Geisel School of Medicine at Dartmouth; Center for Technology and Behavioral Health
Nicholas.C.Jacobson@dartmouth.edu

Michael V. Heinz, M.D.
Assistant Professor of Psychiatry
Geisel School of Medicine at Dartmouth; Dartmouth Health; Center for Technology and Behavioral
Health

Disclosures
The views expressed are our own and do not represent Dartmouth College or Dartmouth Health. NCJ
and MVH are co-founders of, and hold equity in, a company that licenses the Therabot technology from
Dartmouth College. NCJ serves as a paid advisor on the Spring Health AI Council, serves on the advisory
boards of Rula and Sharewell Labs, and has received grant funding from Boehringer Ingelheim.

References
Heinz MV, Mackin DM, Trudeau BM, et al. Randomized trial of a generative AI chatbot for mental health
treatment. NEJM AI. 2025;2(4). doi:10.1056/AIoa2400802
Jacobson NC, Heinz MV. Comment on Generative Artificial Intelligence-Enabled Digital Mental Health
Medical Devices. Docket FDA-2025-N-2338, comment FDA-2025-N-2338-0108. December 8,
2025.
Rousmaniere T, Zhang Y, Li X, Shah S. Large language models as mental health resources: patterns of use
in the United States. Practice Innovations. Published online July 21, 2025.
doi:10.1037/pri0000292
U.S. Food and Drug Administration. Digital Health Advisory Committee meeting transcript, November 6,
2025: Generative Artificial Intelligence-Enabled Digital Mental Health Medical Devices. Docket
FDA-2025-N-2338.

Jacobson & Heinz, Comment on FDA-2026-N-7874 | Page 9