David Ruttenberg
Themes it raises
The comment as filed
2 October 2026
Dockets Management Staff (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Room 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
To the Food and Drug Administration’s Center for Devices and Radiological Health:
I am Dr. David Ruttenberg, a researcher working at the intersection of neurodiversity, AI ethics, and digital mental health. I appreciate the opportunity to comment on FDA’s discussion paper concerning generative-AI-enabled medical devices.
My central concern is that a device can perform acceptably across an aggregate test population while performing poorly for autistic, ADHD, or otherwise neurodivergent patients. In behavioral health, communication support, sensory monitoring, and care triage, those errors may be consequential. A system might mistake an atypical communication pattern for lack of distress—or interpret a harmless, familiar behavior as a sign of crisis. Neither aggregate accuracy nor a general assurance of “representative” testing is enough to establish safe performance for an intended population that includes neurodivergent people.
I urge FDA to address this concern explicitly as it considers the questions in its discussion paper, particularly those concerning risk assessment, subgroup performance, clinical confirmation, multi-turn conversations, care escalation, and postmarket monitoring.
1.Make subgroup performance specific to intended use. For devices likely to encounter neurodivergent patients, manufacturers should explain whether and how autistic, ADHD, and other relevant users were represented in evaluation data. They should report meaningful subgroup results—including false negatives and false positives where clinically relevant—rather than aggregate performance alone. FDA’s paper already identifies subgroup performance as a possible benchmarking element; I ask the agency to make clear that clinically relevant neurotype is a potential subgroup consideration, not an assumed synonym for the general patient population.
2.Test realistic interactions, not just isolated prompts. A conversational device may move from providing information to effectively directing action over several turns. Evaluations should include clinically realistic conversations with patients who communicate distress atypically, need additional processing time, or express uncertainty in ways the model may misinterpret. Tests should examine whether the device asks an appropriate follow-up question, communicates uncertainty, and escalates to a qualified human when needed.
3.Evaluate both directions of care-escalation error. Under-escalation may delay needed care. Over-escalation can also cause harm, including unnecessary intervention, anxiety, or loss of trust. FDA should ask manufacturers to examine both error types across relevant patient groups and deployment settings, rather than reporting a single overall escalation score.
4.Confirm performance in the intended setting. Benchmark results should be supplemented, proportionate to risk, by clinical or clinically representative evaluation involving the people and workflows for which a device is intended. Relevant patient perspectives and first-person reports can identify failures that a clinician-rated or model-scored benchmark misses.
5.Monitor subgroup harms after deployment. Changes to a foundation model, prompts, guardrails, or user interface may change how a device responds to the same patient. Postmarket plans should specify how manufacturers will detect emerging subgroup performance gaps, investigate reported misclassifications, communicate material risks, and correct or pause affected functions.
These requests need not presume that every AI-enabled wellness product falls under FDA device jurisdiction. They concern products and software functions within the scope of FDA’s medical-device oversight. Nor do they require FDA to adopt a single uniform test for every device: the appropriate evidence should follow the device’s intended use and the consequences of getting its output wrong.
The discussion paper recognizes the importance of subgroup performance, real-world confirmation, and monitoring across the product life cycle. I ask FDA to carry those principles through explicitly for neurodivergent patients, so that a device is not judged safe for everyone on the strength of results that obscure whom it fails.
Respectfully yours,
Dr David Ruttenberg, PhD, FRSA, FIoHE, AFHEA, HSRF
UCL Senior Honorary Research Fellow · Fulbright Specialist Awardee
United States Department of State | Bureau of Educational and Cultural Affairs
ORCID: https://orcid.org/0000-0002-8459-121X
t.: +1.561.206.2160 | w.: www.davidruttenberg.com | e.: d.ruttenberg@ucl.ac.uk
P.O. Box 970692 | Boca Raton, FL 33497 | United States of America
Attachment
Dr David Ruttenberg, PhD
COGNITIVE COMPASS
CLARITY. DIRECTION. IMPACT.
2 October 2026
Dockets Management Staff (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Room 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
To the Food and Drug Administration’s Center for Devices and Radiological Health:
I am Dr. David Ruttenberg, a researcher working at the intersection of neurodiversity, AI ethics, and digital
mental health. I appreciate the opportunity to comment on FDA’s discussion paper concerning generative-AIenabled medical devices.
My central concern is that a device can perform acceptably across an aggregate test population while
performing poorly for autistic, ADHD, or otherwise neurodivergent patients. In behavioral health,
communication support, sensory monitoring, and care triage, those errors may be consequential. A system
might mistake an atypical communication pattern for lack of distress—or interpret a harmless, familiar
behavior as a sign of crisis. Neither aggregate accuracy nor a general assurance of “representative” testing is
enough to establish safe performance for an intended population that includes neurodivergent people.
I urge FDA to address this concern explicitly as it considers the questions in its discussion paper, particularly
those concerning risk assessment, subgroup performance, clinical confirmation, multi-turn conversations,
care escalation, and postmarket monitoring.
1. Make subgroup performance specific to intended use. For devices likely to encounter neurodivergent
patients, manufacturers should explain whether and how autistic, ADHD, and other relevant users were
represented in evaluation data. They should report meaningful subgroup results—including false
negatives and false positives where clinically relevant—rather than aggregate performance alone. FDA’s
paper already identifies subgroup performance as a possible benchmarking element; I ask the agency to
make clear that clinically relevant neurotype is a potential subgroup consideration, not an assumed
synonym for the general patient population.
2. Test realistic interactions, not just isolated prompts. A conversational device may move from
providing information to effectively directing action over several turns. Evaluations should include
clinically realistic conversations with patients who communicate distress atypically, need additional
processing time, or express uncertainty in ways the model may misinterpret. Tests should examine
www.davidruttenberg.com | david@davidruttenberg.com | +1.561.206.2160
Dr David Ruttenberg, PhD
whether the device asks an appropriate follow-up question, communicates uncertainty, and escalates
to a qualified human when needed.
3. Evaluate both directions of care-escalation error. Under-escalation may delay needed care. Overescalation can also cause harm, including unnecessary intervention, anxiety, or loss of trust. FDA
should ask manufacturers to examine both error types across relevant patient groups and deployment
settings, rather than reporting a single overall escalation score.
4. Confirm performance in the intended setting. Benchmark results should be supplemented,
proportionate to risk, by clinical or clinically representative evaluation involving the people and
workflows for which a device is intended. Relevant patient perspectives and first-person reports can
identify failures that a clinician-rated or model-scored benchmark misses.
5. Monitor subgroup harms after deployment. Changes to a foundation model, prompts, guardrails, or
user interface may change how a device responds to the same patient. Postmarket plans should specify
how manufacturers will detect emerging subgroup performance gaps, investigate reported
misclassifications, communicate material risks, and correct or pause affected functions.
These requests need not presume that every AI-enabled wellness product falls under FDA device jurisdiction.
They concern products and software functions within the scope of FDA’s medical-device oversight. Nor do they
require FDA to adopt a single uniform test for every device: the appropriate evidence should follow the device’s
intended use and the consequences of getting its output wrong.
The discussion paper recognizes the importance of subgroup performance, real-world confirmation, and
monitoring across the product life cycle. I ask FDA to carry those principles through explicitly for neurodivergent
patients, so that a device is not judged safe for everyone on the strength of results that obscure whom it fails.
Respectfully yours,
Dr David Ruttenberg, PhD, FRSA, FIoHE, AFHEA, HSRF
UCL Senior Honorary Research Fellow · Fulbright Specialist Awardee
United States Department of State | Bureau of Educational and Cultural Affairs
ORCID: https://orcid.org/0000-0002-8459-121X
t.: +1.561.206.2160 | w.: www.davidruttenberg.com | e.: d.ruttenberg@ucl.ac.uk
P.O. Box 970692 | Boca Raton, FL 33497 | United States of America
www.davidruttenberg.com | david@davidruttenberg.com | +1.561.206.2160