[House Hearing, 119 Congress]
[From the U.S. Government Publishing Office]
ADVANCING VA CARE THROUGH
ARTIFICIAL INTELLIGENCE
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON TECHNOLOGY
MODERNIZATION
OF THE
COMMITTEE ON VETERANS' AFFAIRS
U.S. HOUSE OF REPRESENTATIVES
ONE HUNDRED NINETEENTH CONGRESS
FIRST SESSION
__________
MONDAY, SEPTEMBER 15, 2025
__________
Serial No. 119-34
__________
Printed for the use of the Committee on Veterans' Affairs
[GRAPHIC NOT AVAILABLE IN TIFF FORMAT]
Available via http://govinfo.gov
__________
U.S. GOVERNMENT PUBLISHING OFFICE
61-916 WASHINGTON : 2025
-----------------------------------------------------------------------------------
COMMITTEE ON VETERANS' AFFAIRS
MIKE BOST, Illinois, Chairman
AUMUA AMATA COLEMAN RADEWAGEN, MARK TAKANO, California, Ranking
American Samoa, Vice-Chairwoman Member
JACK BERGMAN, Michigan JULIA BROWNLEY, California
NANCY MACE, South Carolina CHRIS PAPPAS, New Hampshire
MARIANNETTE MILLER-MEEKS, Iowa SHEILA CHERFILUS-MCCORMICK,
GREGORY F. MURPHY, North Carolina Florida
DERRICK VAN ORDEN, Wisconsin MORGAN MCGARVEY, Kentucky
MORGAN LUTTRELL, Texas DELIA RAMIREZ, Illinois
JUAN CISCOMANI, Arizona NIKKI BUDZINSKI, Illinois
KEITH SELF, Texas TIMOTHY M. KENNEDY, New York
JEN KIGGANS, Virginia MAXINE DEXTER, Oregon
ABE HAMADEH, Arizona HERB CONAWAY, New Jersey
KIMBERLYN KING-HINDS, Northern KELLY MORRISON, Minnesota
Mariana Islands
TOM BARRETT, Michigan
Jon Clark, Staff Director
Matt Reel, Democratic Staff Director
SUBCOMMITTEE ON TECHNOLOGY MODERNIZATION
TOM BARRETT, Michigan, Chairman
NANCY MACE, South Carolina NIKKI BUDZINSKI, Illinois, Ranking
MORGAN LUTTRELL, Texas Member
SHEILA CHERFILUS-MCCORMICK,
Florida
Pursuant to clause 2(e)(4) of Rule XI of the Rules of the House, public
hearing records of the Committee on Veterans' Affairs are also
published in electronic form. The printed hearing record remains the
official version. Because electronic submissions are used to prepare
both printed and electronic versions of the hearing record, the process
of converting between various electronic formats may introduce
unintentional errors or omissions. Such occurrences are inherent in the
current publication process and should diminish as the process is
further refined.
C O N T E N T S
----------
MONDAY, SEPTEMBER 15, 2025
Page
OPENING STATEMENTS
The Honorable Tom Barrett, Chairman.............................. 1
The Honorable Nikki Budzinski, Ranking Member.................... 3
WITNESSES
Panel I
Mr. Charles Worthington, Chief Technology Officer & Chief
Artificial Intelligence Officer, Office of Information &
Technology, U.S. Department of Veterans Affairs................ 5
Accompanied by:
Dr. Evan Carey, Ph.D., Acting Director, National Artificial
Intelligence Institute, Digital Health Office, Veterans
Health Administration, U.S. Department of Veterans
Affairs
Mr. Sid Ghatak, Chief Technical Advisor, National Artificial
Intelligence Association....................................... 7
Dr. Mohammad Ghassemi, Ph.D., Assistant Professor, Department of
Computer Science and Engineering, College of Engineering,
Michigan State University...................................... 9
Ms. Carol Harris, Director, Information Technology and
Cybersecurity Issues, U.S. Government Accountability Office.... 10
APPENDIX
Prepared Statements Of Witnesses
Mr. Charles Worthington Prepared Statement....................... 31
Mr. Sid Ghatak Prepared Statement................................ 32
Dr. Mohammad Ghassemi, Ph.D. Prepared Statement.................. 35
Ms. Carol Harris Prepared Statement.............................. 40
ADVANCING VA CARE THROUGH
ARTIFICIAL INTELLIGENCE
----------
MONDAY, SEPTEMBER 15, 2025
Subcommittee on Technology Modernization,
Committee on Veterans' Affairs,
U.S. House of Representatives,
Washington, DC.
The subcommittee met, pursuant to notice, at 2:58 p.m., in
room 360, Cannon House Office Building, Hon. Tom Barrett
(chairman of the subcommittee) presiding.
Present: Representatives Barrett, Luttrell, Budzinski, and
Cherfilus-McCormick.
OPENING STATEMENT OF TOM BARRETT, CHAIRMAN
Mr. Barrett. The subcommittee will come to order.
Without objection, the chair may declare a recess at any
time.
Like many who have worn the uniform and received U.S.
Department of Veterans Affairs (VA) health care, I know the
frustration when the system is slow, the paperwork stacks up or
the technology fails, or does not lead us in the direction we
are trying to go. That is why this subcommittee's work is so
critical and why it is important that we have the folks here
joining us today.
It is our duty to ensure VA's technology is efficient and
reliable, helping veterans rather than standing in the way of
their care. That brings us to the focus of today's hearing,
artificial intelligence, or AI, as it is of course commonly
referred to right now. For some, AI sounds like a science
fiction movie--we have all seen many of them--something only
computer scientists worry about or even something scary because
it is unknown and not well understood. It feels like today
everything is about drones or artificial intelligence. The
world has shifted quite a bit.
Within VA, AI is already being used in ways they are making
a real difference for our veterans. In fact, U.S. Government
Accountability Office (GAO) recently released a report
highlighting how VA is among the most active adopters of
artificial intelligence, from analyzing medical images and
workflows to creating summary diagnostic reports. The report
identified more than 200 reported use cases across the system.
As we were preparing for this hearing, my staff had told me
about some, even very early prototype AI systems that the VA
had integrated decades ago.
In clinical care, AI can help doctors detect cancer earlier
and identify warning signs of heart disease before a crisis
occurs.
A recent study led by VA researchers at the VA Long Beach
Health Care System showed how AI can enable providers to detect
the risk of calcium buildup in heart arteries. This new
technology could give providers the chance to prevent heart
attacks rather than respond to them.
AI is also being used to enhance mental healthcare. One of
the greatest challenges we face as a Nation remains around
veteran suicide. In 2017, VA launched the react vet program--or
the Recovery Engagement and Coordination for Health Veterans
Enhanced Treatment (REACH VET) program. This program uses an AI
model to help identify a very small group of veterans who are
at the greatest risk of suicide. The results were promising.
The program helped VA step in early, guiding veterans to care
before crisis strikes.
It is not just about medical breakthroughs. AI is also
helping doctors and nurses relieve the day-to-day burden of
paperwork and things that are not spending time with the
patients. The most common complaint from providers is that they
spend too much time filling forms and not enough time taking
care of their patients.
The VA is exploring AI tools, like Ambient scribes, which
can listen to a provider's conversation with the patient and
automatically create a clean and accurate medical note. On
average, this saves providers 2 or 3 hours per week. Now
multiply that across thousands of staff. It means more time
spent caring for veterans and less time staring at a computer
screen and doing paperwork. The promise of AI is real. I want
to be clear: Our job here is not applaud the promise. It is to
make sure that AI is being used responsibly, safely, and
transparently. Every great innovation comes with risk, and AI
is of course no exception. If the data is biased, the results
can be unfair. If safeguards are weak, privacy is compromised.
If the systems are not carefully monitored, mistakes could harm
the very people they are trying to protect. That is why the
governance of AI matters.
The VA has been one of the first agencies in the government
to step up a framework for how AI should be reviewed and
approved. By law, VA already has some of the strongest privacy
protections in government, and those protections extend
directly to AI technology in veterans health and benefits, that
it cannot be used by vendors for other purposes, period. We are
going to hear more about that today.
Veterans deserve to know when AI is being used in their
care. They deserve to know the technology has been tested and
that it is working for them, not against them. Congress
deserves to see evidence that taxpayer dollars are being well-
spent.
I see today's hearing as an opportunity to highlight what
is working, to dig into what still needs improvement, and to
set clear expectations for the road ahead.
AI does not and will not replace doctors or nurses, nor the
human touch that every veteran deserves when receiving their
care. Done right, AI can give clinicians another tool in their
tool box, helping them focus more on patients and ultimately on
saving lives.
We need to find the balance between moving efficiently
enough to give veterans the benefit of innovation and
cautiously enough to make sure that no veteran is put at risk.
Our veterans should never be guinea pigs for untested
technology, but they should also not be denied the benefits of
safe and proven innovations.
This subcommittee will hold VA to that standard, and I
intend to make sure that we get it right. It is not about
technology; it is about trust. Veterans give this Nation their
trust through their service. When they in turn go to the VA,
they deserve to know that trust will be given back and honored.
It is our duty to make sure that that trust is never broken. I
know the ranking member will join me in this pursuit, and I
appreciate your joining me in this committee today.
Then, last, before I yield to her, I sat next to another
physician on my flight here last night. We were talking about
AI and actually about this hearing coming up. She was saying,
``We are never going to replace the element of patient care
that is done by doctors, nurses, and other medical
professionals.'' If we can expand the reach that they have,
look at it as, in the Army, we used to say a force multiplier
to deliver benefits in a more deliberative fashion that
benefits everybody. That is what we ought to be pursuing here.
With that, I will yield to Ranking Member Budzinski for
your opening statement.
OPENING STATEMENT OF NIKKI BUDZINSKI, RANKING MEMBER
Ms. Budzinski. Thank you very much, Chairman Barrett. I
appreciate our subcommittee coming together to have a frank
conversation about the underlying Information Technology (IT)
challenges at the Department of Veterans Affairs and how we can
support the VA in closing those gaps through technology.
However, today's review of artificial intelligence use
cases at the Veterans Health Administration (VHA) feels like a
distraction. VA is struggling with the basics. We are here
discussing the newest technologies while the VA is still
working with a crumbling IT infrastructure and still grapples
to modernize systems and workflows.
As the ranking member on the Technology Modernization
Subcommittee, I am certainly excited by the potential of both
AI and innovation. AI could improve some of VA's challenges
through large language models and higher processing speeds. We
have seen promising studies of providers using AI to identify
cancers more easily, improve patient outcomes, and ease
clinician burnout by taking on more administrative tasks.
The VA has certainly been the leader in the research and
development and widespread usage of a number of significant and
groundbreaking technologies. It stands to do so again with AI.
However, success in these efforts requires adequate resources
and investments in its budgets, its processes, and its people.
Veterans choose VA for the community it provides, the people it
employs, and for the fact that it is not driven by profit.
What VA does best is make veterans feel seen and
understood. As we have seen, AI can be a tool to provide
decision support, ease provider burdens, and help with
notetaking so doctors can be more present with the patient. We
should also acknowledge that it is not the answer to every
challenge the VA faces. Also, we as a committee and as Congress
need to have a real conversation about AI policy and how to
implement it safely. I am excited about the opportunities that
AI presents. I am not convinced that VA is prepared to deploy
this technology just yet.
I have a number of concerns that I hope to address today,
like the lack of regulation and governance structures and the
need for better transparency around what data is involved in
training such models.
Further, like all technology modernization efforts,
implementing AI successfully requires a highly skilled,
adequately staffed workforce. Almost 2 weeks ago, the acting
head of the Department on Government Efficiency stressed the
need to ``hire and empower great tech talent in government.'' I
could not agree more with that. However, I think we should all
note the irony of that statement considering Office of
Information and Technology (OIT) is proposing a massive
reorganization and intends to cut at least 20 percent of its
workforce.
Success is also reliant on strong IT leadership. If OIT is
in fact undergoing significant changes to its organizational
structure, priorities list, and workforce makeup, we need a
confirmed Chief Information Officer (CIO) at VA. This position
is particularly critical as we see the acceleration and
progression of modernization efforts at the Department. It
seems the VA still lacks a coherent enterprise IT strategy,
leaving projects AI integration to happen in silos. Without
stable and competent leadership, veterans and VA employees will
continue to be stuck with cobbled-together systems and
workflows that do not meet their needs rather than a solid
strategy for technology usage to guide its decision-making.
I hope that we can get some clarity into the
administration's plan to propose a nominee for the CIO position
and that one can be confirmed before many of these substantial
changes occur.
Last, I understand this subcommittee held a similar hearing
in January 2024, though neither I nor the chairman were on this
subcommittee at that point. In that hearing, data privacy was
an intrinsic part of the discussion. I hope that it still is
the case today.
As we become more interconnected through technology
advancements like artificial intelligence, we must become
increasingly aware of the concerns about the privacy of users'
data, especially in healthcare. Since this last hearing, the
Department has been entangled in multiple cybersecurity
incidents, which have potentially placed veterans' data at
risk. Though many of these breaches have been targeted at VA
contractors, veterans' data has still been implicated, and VA
maintains some responsibility for its safety. Though I do feel
that this hearing is perhaps too early, considering VA has yet
to develop and release some of its policies and plans to align
its efforts with the administration's, I hope to hear from our
VA witnesses today about how data privacy and security, as well
as the views of both VA employees and patients, will be
integrated into such plans.
Thank you and I yield back, Mr. Chairman.
Mr. Barrett. Thank you, Ranking Member Budzinski.
I join you in making sure that we have adequate ethics
guardrails around this, and certainly privacy is paramount in
that as well.
I now want to introduce our witnesses. Again, thank you for
joining us today from the Department of Veterans Affairs, we
have Mr. Charles Worthington, the Chief Technology Officer and
Chief Artificial Intelligence Officer. Thank you for being
here. Accompanying Mr. Worthington is Dr. Evan Carey, Acting
Director over the National Artificial Intelligence Institute at
the VA. We also have Mr. Sid Ghatak.
Did I say that correctly?
Mr. Ghatak. Yes, sir.
Mr. Barrett. Thank you. The chief technical advisor from
the National Artificial Intelligence Association. Dr. Mohammad
Ghassemi, assistant professor at Michigan State University. Go
green.
Dr. Ghassemi. Go white.
Mr. Barrett. Thank you for being here today as well.
Finally, from the Government Accountability Office we have
Ms. Carol Harris, a familiar face to all of us on this
committee. Thank you again for being here and joining us. She
is also Director of IT and Cybersecurity at the GAO. Again,
thank you all for being here.
At this time, I ask the witnesses to please stand and raise
your right-hand.
[Witnesses sworn.]
Mr. Barrett. Thank you. Let the record reflect that all
witnesses have answered in the affirmative.
Mr. Charles Worthington, you are now recognized for 5
minutes to deliver your opening statement on behalf of VA.
STATEMENT OF CHARLES WORTHINGTON
Mr. Worthington. Chairman Barrett, Ranking Member
Budzinski, and distinguished members of the subcommittee, thank
you for the opportunity to discuss the Department of Veterans
Affairs' use of artificial intelligence to enhance healthcare
and services for veterans.
Your steadfast support of the veterans and their families
is invaluable. I am joined today by Dr. Evan Carey, Acting
Director of the National AI Institute in the Digital Health
Office of the Veterans Health Administration.
While AI is not new to VA, recent advancements in AI
systems presents a tremendous opportunity to improve VA's
services. When used effectively, AI can improve the efficiency
and accuracy of many time-consuming and error-prone tasks that
create burdens for VA staff and veterans alike. That is why VA
is rapidly working to capitalize on this technology.
Our strategic vision is to make VA a leader in AI,
providing faster services, higher quality care, and more cost-
effective operations. We will aggressively deploy this new
technology while remaining committed to strong controls that
ensure security, privacy, and effectiveness of our technology
systems.
We have distilled this vision into five key priorities.
First, we are aggressively expanding AI across our workforce.
Second, we are reimagining high-impact workflows through AI and
automation.
Third we are prioritizing investment in data and
infrastructure that supports those high potential use cases.
Fourth, we are cultivating an AI-ready workforce. Finally,
we are executing transparent and effective governance, an
essential requirement to maintain veterans' trust. We are
already bringing the strategy to life, making significant
investments in AI-driven tools.
In 2024, our AI inventory had 227 use cases in it, which
was nearly 100 more than the previous year. We expect this
growth to continue in 2025 as we prepare for our December
update to that inventory. These investments are delivering
tangible results. I am pleased to report that all VA employees
now have access to secure generative AI tool to assist them
with their work. In surveys, users of this tool are reporting
that it is saving them over 2 hours per week.
Additionally, over 2,000 VA staff and contract and software
developers are using an AI software development copilot tool,
enabling faster delivery of features that help veterans. AI is
also revolutionizing clinical care. In fact, 82 percent of VA's
AI use cases come from the Veterans Health Administration.
VA's stratification tool for opioid risk mitigation uses
machine learning to identify veterans at high risk of overdose
and suicide, enabling healthcare teams to review and intervene
effectively. Since 2017, the REACH VET program, as you
mentioned, has used AI answer algorithms to identify over
130,000 veterans at elevated risk, improving outpatient care
and reducing suicide attempts.
AI-assisted colonoscopy devices have increased adenoma
detection rates by 21 percent, reducing late stage cancer
incidents and mortality. Thanks to groundbreaking research by
folks like Dr. Raffi Hagopian and Dr. Evan Carey, the VA is
exploring how AI could help providers detect heart disease
earlier by reviewing the millions of Computed Tomography (CT)
scans that are not currently evaluated for cardiovascular
disease risk at all.
As we advance our AI deployments, protecting veterans' data
remains paramount. All AI systems approved for use at VA must
meet VA rigorous security and privacy standards before
receiving an authority to operate. Additionally, consistent
with Office of Management and Budget's (OMB) policy, we conduct
a thorough agency-level review of each AI use case to ensure
that it meets the government the standards.
We will publish the results of this review in our annual AI
inventory, positioning us as one of the most transparent
healthcare systems in the country with regards to our use of
artificial intelligence.
Despite our progress, adopting AI tools does present
challenges. As you mentioned, integrating new AI solutions with
a complex system architecture and balancing innovation with
stringent security compliance is crucial. Recruiting and
retaining AI talent remains difficult. Scaling commercial AI
tools incurs additional costs. This underscores the importance
of full congressional funding for VA to continue this critical
work.
In conclusion, the Department of Veterans Affairs is
committed to harnessing AI to improve the lives of veterans.
Through strategic investments in AI tools and workforce
capabilities, we strive to it deliver faster, higher quality,
and more cost-effective services. Your continued support is
vital for VA to lead in AI innovation and set a benchmark for
responsible AI use in government.
Thank you for the opportunity to discuss our strategy, and
we look forward to your questions.
[The Prepared Statement Of Charles Worthington Appears In
The Appendix]
Mr. Barrett. Thank you, Mr. Worthington.
The written statement of Mr. Worthington will be entered
into the hearing record.
Mr. Ghatak, you are now recognized for 5 minutes to deliver
your opening statement.
STATEMENT OF SID GHATAK
Mr. Ghatak. My name is Sid Ghatak, and for almost three
decades, I have designed and deployed artificial intelligence
forecasting systems across finance, healthcare,
pharmaceuticals, media, and government.
I currently serve as the chief technology adviser for the
National Artificial Intelligence Association, the premier
organization representing 1,500 businesses in the advancement
of AI. I am also the founder and chief executive officer of
Increase Alpha, where we use artificial intelligence to predict
stock prices, and we license these predictions to hedge funds.
In the Federal Government, I served in the General Services
Administration for 4 years where I was a Director of the Data
and Analytics Center of Excellence. In that role, I coauthored
the Federal AI maturity model 3 years before AI took the world
by storm. I also contributed previous executive orders on the
critical issues of data privacy and data security. At Increase
Alpha, I increased a predict--architected a predictive AI model
that generates off of once thought impossible, a deep learning
system that is exceptionally accurate at predicting equity
prices. Increase Alpha far exceeds multiple industry
benchmarks, including accuracy, sharp ratio, and alpha
generation. The solution itself is not based on large language
models at all, but it is purpose built, designed for this
specific need.
I want to emphasize that this company and our solution is
completely unrelated to the Department of Veterans Affairs, and
it has no bearing on today's testimony. I mention it only as an
example of how AI, when carefully designed with a clear
purpose, can achieve exceptional effectiveness.
Taken together, this diverse background, spanning academia
and government and industry, has given me the rare opportunity
to actually build AI systems that work well in the real world.
I have spent my career outside the orthodox roles of academia,
venture capital, and Big Tech, I am also not beholden to herd
mentality. Instead, I bring an expert independent perspective,
which is especially valuable now when much of the world is
caught up in the art of the possible with AI when what is most
urgently needed is a sober understanding of what is safe,
practical, and ready to serve the public.
Large Language Models (LLM) like ChatGPT, Claude, and
Gemini are a powerful subset of AI, but they come with their
own set of problems, specifically in healthcare where
hallucinations and sycophancy on the part of ChatBots can lead
susceptible users down psychological rabbit holes, which is why
it is important to clarify that AI is bigger than just ChatGPT
and its competitors. To use an analogy, the steam engine
transformed society, fueling the Industrial Revolution. While
steam power exists today, it gave way to other forms of power
over time. Until steam engines were used to create the first
railroads, no human had ever traveled faster than a horse. This
new form of transportation opened the world's eyes to what is
possible, just as ChatGPT has shown the world the art of a
possible with artificial intelligence. Early train travel was
dangerously unreliable. Accidents were frequent, derailments
common, and thousands of lives were lost before rail systems
matured into safe networks that we know today.
The lesson is clear: Revolutionary technologies will evolve
and improve over time when the private sector and the
government work in collaboration. The same applies to
artificial intelligence. As the committee gathers information
on how to modernize technology at the VA, I would like to offer
a few pieces of advice from my many decades on the front lines
of building and implementing advanced analytical solutions.
As I mentioned, the last several years, the world has been
consumed with LLMs to the point where AI has become synonymous
with it. However, that is not the case. Many other types of AI
may have similarities to these models but function very
differently, technologies that specialize in interpreting and
understanding images, video, and audio, for example, or
technologies that are better suited to working with numbers and
symbols instead of words, a new technology that is yet to be
invented.
There is an old adage about, when you are a hammer,
everything likes like a nail. The world has become so enamored
with LLMs, and rightfully so, interacting with them can feel
magical, giving you the sense that they are real people, but
they are not. This may be why little to no investment is being
made in these other areas. At Increase Alpha, we demonstrate
clearly what can be done with other forms of artificial
intelligence. I began building our models at the same time as
the research underlying ChatGPT was published. I had also
encountered the same compute cost energy and reliance on a
video that we still see today. I took a different approach to
conserve resources and focus on simplification, using
predictive intelligence which led to leading AI models that use
a minuscule amount of data compared to LLMs and which are small
enough to run on a cell phone.
What does all this mean for the VA and the well-being and
care of veterans? I do not claim to know. No one really does. I
want to leave you with a prediction: I believe that we truly
are on the verge of a scale--of a revolution on the scale of
the Industrial Revolution. If I could leave you with one idea
today, it would be this: AI is much bigger than today's LLMs.
It is these technologies, many of which have yet to be
invented, that will enable the VA to execute on its mission.
Thank you.
[The Prepared Statement Of Sid Ghatak Appears In The
Appendix]
Mr. Barrett. Thank you, Mr. Ghatak.
The written statement of Mr. Ghatak will be entered into
the hearing record. I appreciate your remarks. I think, if we
all use ChatGPT for cat memes, it will not be meeting its full
potential and leaving a lot of things behind. Thank you.
Dr. Ghassemi, you are now recognized for 5 minutes for your
opening statement.
STATEMENT OF MOHAMMAD GHASSEMI
Dr. Ghassemi. Chairman, Ranking Member, and members of the
subcommittee. Thank you for the opportunity to speak today. I
am a scientist and an entrepreneur focused on artificial
intelligence but especially its applications to healthcare. The
views I am going to share today are my own, but they are
informed by roles I played as a professor at Michigan State
University, where I direct a research laboratory on AI and its
applications to health sciences.
I am also going to bring a perspective as the founder of an
AI consultancy Gamut Corporation, which has helped large
pharmaceutical companies, insurance companies, as well as
health systems, plan and execute their AI strategy.
I want to be clear: I am not a veteran health specialist.
My perspective is on how artificial intelligence can broadly
advance care in ways directly relevant to the needs of
patients, and this very critically includes our veterans.
This subcommittee has identified in their invitation letter
three priorities for AI health. These were transforming
healthcare delivery, streamlining services, and improving
outcomes. I am going to frame my remarks around three roles
that AI can play to help with these three priorities. The three
roles are automation, which is reducing low-value work through
the use of machines; augmentation, which is having a machine
assist a human in a task, so to strengthen clinical decision-
making, as an example; and insights, which is allowing us to
extract complex patterns from data, patterns far too complex
for us to discern just with our human intuitions alone. Let us
talk about three.
First, AI can transform what happens during care itself.
Clinicians today spend hours on paperwork, but AI scribes can
generate notes automatically so they can focus more fully on
patients. We have heard that from more than one person in the
conversation today.
In emergency rooms, decision tools powered by AI can help
identify the sickest patient sooner and get them treated
faster. Continuous monitoring assistance can pick up on the
early signs of decline, like sepsis, long before they would be
obvious to our human eyes. These tools make the encounter
safer, timelier, and more patient-centered.
Second, AI cannot only streamline what happens during care;
it can streamline the plumbing of healthcare itself. Missed
appointments waste scarce clinician time. Automated reminder
systems, which do not have to use a large language model or a
sophisticated tool like ChatGPT, can reduce these no-shows and
save that time. Patients also too often fall between the cracks
between primary care and specialist visits. AI can flag the
missing referral information, track follow up, and prevent all
these gaps. When imaging or labs reveal unexpected findings,
like, God forbid, a lung nodule discovered by chance, AI
tracking systems can ensure these findings are followed up on
so that the treatable conditions do not get overlooked. This is
how we reduce wasted effort and ensure smoother, more reliable
care.
In conclusion, artificial intelligence is not a silver
bullet. I say this as a person who has been working on
developing the methods for several years, but it can already
help with the subcommittee's three priorities. It works best
when it reduces low-value work, strengthens rather than
replaces clinical judgment, and turns complex data into
actionable insights.
To succeed, we need disciplined pilots, clear metrics, and
safeguards for safety, equity, and privacy. If deployed with
care, AI can return time from paperwork to patients, ensure
that critical findings are not missed, and support clinicians
in their hardest decisions.
I look forward to our conversation. I am grateful for the
invitation to be here with you today.
[The Prepared Statement Of Mohammad Ghassemi Appears In The
Appendix]
Mr. Barrett. Thank you, doc.
The written statement of Dr. Ghassemi will be entered into
the hearing record.
Ms. Harris, you are now recognized for 5 minutes to deliver
your opening statement on behalf of GAO.
STATEMENT OF CAROL HARRIS
Mr. Harris. Chairman Barrett, Ranking Member Budzinski, and
members of the subcommittee, thank you for inviting us to
testify today on the use of artificial intelligence at VA.
Develops in generative AI, which is a subset of AI, which can
create text, images, video, and other content when prompted by
a user, have revolutionized how the technology can be used in
many industries, including healthcare and at VA and other
Federal agencies.
AI holds substantial promise for improving the operations
of government agencies. However, it can increase risk for
agencies and poses unique oversight challenges because the
source of information used by AI is not always clear or
accurate. Given the fast pace at which AI is evolving, the
government must be proactive in understanding its complexities,
risks, and societal consequences.
It should also be noted that VA has experienced
longstanding challenges in managing its IT projects and
programs, raising questions about the efficiency and
effectiveness of its operations and its ability to deliver
intended capabilities.
As requested, I will briefly summarize our prior work on
the Department's AI use and challenges, as well as principles
and key practices for Federal agencies, including VA, that are
considering and implementing AI systems.
In July 2025, we reported that VA's AI use cases increased
from 40 in 2023 to 229 in 2024. For example, VA is a developing
a generative AI use to automate various medical imaging
processes. This use may enhance VA's ability to analyze medical
images, integrate existing and new data workflows, and create
summary diagnostic reports.
In the health and medical sector, agencies have adopted
generative AI to advance medical research and improve public
outcomes, including at VA. It is also worth noting that, of the
229 use cases, 64 percent were considered to be high-impact AI,
meaning that their capabilities impact the rights and/or safety
of individuals or entities. Looking at just VHA, that
percentage increases to 72 percent.
The Department also reported to us a number of challenges
they face in using and managing generative AI. The full list is
noted in my written statement. I will only highlight a few
here.
Challenge one, complying with existing Federal policies and
guidance. VA officials shared that the existing Federal AI
policy can present obstacles to the adoption of generative AI,
including in the areas of cybersecurity, data privacy, and IT
acquisitions.
Challenge number two, having sufficient technical resources
and budget. Gen AI can require infrastructure with significant
computational and technical resources. VA noted challenges in
obtaining or accessing the needed technical resources and also
in having the funding necessary to establish those resources
and support desired AI initiatives.
The last challenge, hiring and developing an AI workforce.
Among other things, the VA reported difficulties in
establishing and providing ongoing education and technical
skills development for their current workforce.
VA officials told us they are working toward implementing
the new AI requirements in OMB's April 2025 memorandum. Doing
so will provide opportunities to develop and publicly release
AI strategies for identifying and removing barriers and
addressing the challenges I noted.
Additionally, the GAO has identified a framework of key
practices to help ensure accountability and responsible AI use
in the design development, deployment, and continuous
monitoring of AI systems.
Our framework is organized around four complimentary
principles that address governance, data, performance, and
monitoring. Consideration of the key practices in this
framework can help VA as it considers, collects, and implements
AI systems.
Last, I will mention that we have 26 open recommendations
to VA concerning the management of its IT resources. If the
Department implements these recommendations effectively, it
will be better positioned to overcome its longstanding
challenges in managing its IT resources and will improve its
ability to address the rapidly changing AI landscape.
That concludes my statement. I look forward to addressing
your questions.
[The Prepared Statement Of Carol Harris Appears In The
Appendix]
Mr. Barrett. Thank you, Ms. Harris.
The written statement of Ms. Harris will be entered into
the hearing record.
Again, thank you to all of our witnesses.
We will now proceed to questioning. I will recognize myself
for 5 minutes to begin questioning.
I am going to start with Mr. Worthington. The VA--I know we
have got a lot of concerns obviously about data security, data
privacy, what can be used, what can be modeled off of veteran
information. The VA requires vendors to sign contracts directly
stipulating that it will prevent secondary use of veteran data.
Number one, can you kind of walk us through how that works? How
are you making sure that companies actually follow that rule?
Mr. Worthington. Thank you for the question, Chairman
Barrett. I think it is extremely important that everyone
understands that there is not a second set of rules for AI
systems. In the VA, we have a very clear and stringent set of
rules around both security and privacy for any technology
system. Before we bring a system into production, we have to
review that system for its compliance with those requirements
and ensure that the partners that are working with us on those
systems attest to and agree with those requirements. AI systems
receive an authority to operate just like any other system
would before we would put veteran data into the system.
Mr. Barrett. Okay. I appreciate that. For example, though,
I know the large language model, kind of most stereotypical use
of AI, we are going to be looking at, you know, the millions of
records that the VA has and then modeling patient outcomes from
that and then looking kind of retrospectively to see where
people are on that spectrum today, and say, ``Well, we know, if
this condition led to 10 years later a worse condition over
here, how can we stem that off earlier?'' If we allow an AI
vendor to have access to that to cultivate that knowledge, is
that something that could be then used as an outgrowth in
another way for, like, another I guess research tool for other
things? For example, if a person has a predisposition to kidney
disease or diabetes or something like that, we can look
retrospectively at their health record to show that they had
certain indicators ahead of time, would not we want that to be
to the benefit of all medicine and not just within the VA?
Mr. Worthington. Yes. I think that, as you are mentioning,
in the training phase of models, which VA does occasionally do,
that, if we work with a vendor, we make sure that the
agreements say that any protected health information can only
be used for that specific purpose that we have contracted with.
Often, that is taking place in environments that VA already
runs and controls.
Now when we are talking about using a large language model,
those are provided typically via one of the big cloud service
providers, and those environments are set aside in a VA
boundary that basically the vendor has to attest that they
already meet VA security requirements. When we are sending
information to a large language model to get feedback back from
that model, we are using a version of that model that has been
made secure to meet government standards.
Mr. Barrett. Okay. I will fully confess that I am not an
expert on this. Would a large language model allow a
practitioner to say, ``I have a veteran presenting with these
conditions; what are the risk factors that I ought to look for
to, maybe run tests that would not ordinarily be otherwise top
of mind?''
Mr. Worthington. There could be a variety of AI approaches
for a use case like that. Dr. Carey may just quickly provide a
couple of examples of those sorts of decisions support type use
cases.
Dr. Carey. Thank you. It is a fantastic question. I think
there are two versions of that. As you note, there are tools
where providers can get general advice, and they might
specifically articulate the needs of the veteran and for the
conditions that they are looking for, to point out to sort of
follow the different procedures that are recommended and
identify the guidelines. Those tools are available within the
VA.
Mr. Barrett. Okay. After the passage of the The Sergeant
First Class Heath Robinson Honoring our Promise to Address
Comprehensive Toxics (PACT) Act, you know, we have this burn
pit registry and everything, and they are supposed to track
veterans and conditions that arose from that. Obviously, the
specific information about a particular veteran we want to have
protected and not revealed. If there are outcomes of that that
could be useful to, you know, human medicine in total, is there
a way for that to be revealed?
Mr. Worthington. Yes, thank you for the question. VA does
have, as you know, a very large amount of health data. We have
a robust----
Mr. Barrett. More than anybody in the world, I think.
Mr. Worthington. That is right. We have a robust tradition
of research to advance not just VA healthcare but healthcare
overall. We are seeing an increasing interest in using that
data for AI-driven research papers, like the one that Dr. Carey
recently wrote.
Mr. Barrett. Okay, and that is the--like the benefit but
also the concern is we obviously have a large repository of
medical data. If that is being used or to the benefit of a
curator of artificial intelligence, should the VA be, you know,
should that be brought into account for the cost of services
and other things like that? What I do not want is for a
provider to come in and leach that information out solely for
their benefit while not providing a benefit to the VA and to
the veterans as well.
Mr. Worthington. We agree.
Mr. Barrett. Thank you.
Ranking Member Budzinski.
Ms. Budzinski. Thank you, Mr. Chairman.
Dr. Carey and Mr. Worthington, thank you so much for both
being here.
I understand that several of VA's AI use cases, like the
ambient dictation pilot, intend to use an opt-in practice for
consent. For systems that are perhaps less directly veteran-
facing, like the use of AI in benefits determination or medical
assessments, how is the Department educating veterans on these
use cases to ensure for their awareness?
Mr. Worthington. At a very high level--and thank you for
the question. We are using our AI use case inventory as the way
to catalogue all of the uses of AI and make sure that that is
publicly available. When there is not, as you mentioned, like a
one-to-one interaction that provides the opportunity to explain
directly what is happening, as there is in many healthcare
settings, what we are relying on is our publishing of the
overall AI strategy and use case to explain how the Department
is using AI in various products and services.
Ms. Budzinski. Okay. Other than that general awareness--for
veterans, is there any way to kind of draw their attention to
this so that they know that, you know, what their situation
might be using to inform an AI model?
Mr. Worthington. We are always listening for veterans'
feedback through a variety of mechanisms and reacting to that.
That is true of AI situations and non-AI situation. We
certainly want to monitor this for AI in particular, because I
think maintaining veterans' trust in VA as we introduce these
new technologies is going to be critical.
Ms. Budzinski. Okay. Then, Mr. Worthington, I am glad that
you and your teams are committed to the transparency in AI use
cases at the Department. That is commendable. However, there
have been reports that certain employees had access to certain
data sets and systems within VA's enclave which may have been
used for AI related operations. I have some specific employees
I want to mention by name, and then I have some questions for
you. I am going to ask about these employees: Justin Fulcher,
Sahil Lavingia, Christopher Roussos, Payton Rehling, Cary
Volpert, or Jon Koval. I am just looking for, like, a yes or no
to these questions. Did you ever work with any of those
individuals?
Mr. Worthington. Yes, I have come across several of them.
Ms. Budzinski. Okay. Are or were these individuals
affiliated with the Department of Government Efficiency (DOGE)?
Mr. Worthington. I am not exactly clear on the
relationship. I believe they are VA employees. At points, they
were introduced as also being part of the DOGE movement.
Ms. Budzinski. Okay. Did any of these employees access data
sets that included VA patient medical records or other
personally identifiable information?
Mr. Worthington. I am not aware.
Ms. Budzinski. Okay. Were you or anyone you know ever asked
to duplicate data sets by these employees?
Mr. Worthington. No, I was not.
Ms. Budzinski. Okay. Can you commit to me that no veteran's
data was removed from the Department of Veterans Affairs?
Mr. Worthington. As far as I understand, all the VA
employees follow all the VA IT security processes and
procedures and that was a key priority for all of us and always
is a key priority.
Ms. Budzinski. Okay, Okay. Mr. Worthington, almost 2 weeks
ago, the Acting Director of the U.S. Digital Service noted that
the Federal Government needs more tech employees to--and to
hire and empower great talent. Do you believe that VA shares
that sentiment?
Mr. Worthington. Yes, I do. I think having technologists in
government is critically important, as is having great
researchers and doctors.
Ms. Budzinski. Okay. Secretary Collins has often noted the
importance of VA employees in direct care roles, disregarding
the importance of what he might call support employees in the
provision of this work. Do you believe that this type of
rhetoric has helped the Department to recruit and retain tech
talent?
Mr. Worthington. I think the good thing about working at
the VA is our mission is so clear. The mission of serving
veterans is the most important one that I have worked on in my
tech career. I think there are many technologists across the
country that are willing to sign up for that mission. I love
trying to recruit those people on my team.
Ms. Budzinski. Ms. Harris, real quick on a follow up, GAO's
Artificial Intelligence Accountability Framework notes the
workforce is a key component to ensuring effective AI
application. How does a highly skilled technical workforce
ensure adequate scalability of AI applications and protection
of veteran data?
Mr. Harris. Well, while there is great excitement around AI
because of the potential to improve operations, there is also
significant concerns, the ones that I articulated earlier about
cybersecurity, intellectual property, as well built-in bias in
the AI system, as well as environmental and other concerns. We
want to make sure that we have a workforce that understands
both the potential of these systems but also understands the
risks in the AI well. Having those two are vital.
Ms. Budzinski. Okay. Thank you.
I yield back.
Mr. Barrett. Thank you.
Mr. Luttrell.
Mr. Luttrell. Mr. Chairman.
Mr. Ghassemi, you laid out a well-articulated plan of
attack on how the VA could tackle this healthcare, artificial
intelligence kind of combining of forces. The problem is you
have--it sounds like you never worked with the U.S. Government
because that is what kills this effort is the U.S. Government.
Ms. Harris, your opening statement was very well-
articulated, and you hit every single point precisely. The
problem is we have such an issue with the VA because it is a
big machine, and we are trying to compound--we are trying to
bring in artificial intelligence to streamline the process. You
have 172 different VA facilities, plus satellite campuses, and
that is 172 different silos. They do not work together. They do
not communicate very well with each other. We have spent almost
$16 billion trying to push electronic healthcare records across
multiple facilities. Now we are going to try to tackle
artificial intelligence as well. In 2024, we had 229 AI
actions. Correct, Mr. Worthington?
Mr. Worthington. Yes, approximately.
Mr. Luttrell. What site did that come from, because I would
dare say that that did not come from all or every single VA
installation. That sounds like to me that that is collected
from, like, a few. Is that correct?
Mr. Worthington. We did attempt to have a pretty
comprehensive review process to gather all of the uses of AI
across the country. We----
Mr. Luttrell. I did not get anything out of that. That was
almost a yes-or-no question, but go ahead again.
Mr. Worthington. Yes, I believe that AI is being used at
facilities across the country. This inventory covers those
uses.
Mr. Luttrell. The conversations I have with multiple sites
is they do not have artificial intelligence capabilities
because their sites are not ready or they do not have the
infrastructure in place to do that, because we keep compounding
software on top of software. Some sites cannot function at all
with the new software they are trying to implement. That is a
pretty fair statement, correct?
Mr. Worthington. I would agree that having standardized
systems is a challenge at the VA. There is a bit of a
difference in different facilities. Although I do think many of
them are starting to use AI-assisted medical devices, for
example, and a number of those are covered in this inventory.
Mr. Luttrell. How do we fix this problem? Again, I am going
to ask you, sir, because I usually ask everyone who sits in
front of me from the VA: How would we fix this problem? Mr.
Ghatak and Mr. Ghassemi have probably thought about this quite
a bit before they showed up in front of us, but again they have
not--actually, I do not know this for certain--I may be
throwing this at you, and course correct me if you would like--
but I do not think they have had to deal with the U.S.
Government and also the VA. Now how long have you been in this
position, sir?
Mr. Worthington. I have been at the VA nearly 10 years and
this position for about 2 years as chief AI officer.
Mr. Luttrell. Okay. What comes first, the communication
between the sites and the ability to ask that information
questions, which we do not do that or we do not have the
ability to do that--do we run the implementation of artificial
intelligence in parallel with that, or do we have to do one
before the other?
Mr. Worthington. In my personal opinion, we cannot wait,
because AI is here, whether we are ready or not. Increasingly,
every solution we buy from our partners in the private sector
is going to have it embedded inside of it. I think our
challenge is we need to come up with very good standard
templates that every site can use and allow those standard
tools to be deployed, things like the VA GPT school that I
mentioned, which is now available to every VA employee in a
standard way.
Mr. Luttrell. Since the Department of Veterans Affairs
houses the most important data set on the planet arguably, and
everyone wants to touch it, including Dr. Ghassemi at Michigan
State--I would have to guess, especially when you were at
Massachusetts Institute of Technology (MIT) in Cambridge I am
sure. Pretty impressive resume, sir. Everybody is trying to
touch it. Everybody wants to be a part of it, and you have to
deal with every single subject-matter expert that walks through
your door that says, ``I am the best.'' I can assure you every
one of those corporations and companies walks into our office
as well. Question is, who is it? Who do you vet, and who is
going to touch it, because it cannot be everybody? We do not
have it--in my personal opinion, that I am not aware of, we do
not have an enclave that can house all of that information
where everybody can get in there and not steal it.
Implementation of artificial intelligence, which we do not have
the ability to regulate, so the question is who will do that,
or do you have the AI system itself regulate itself?
Mr. Worthington. I think it is a great observation and
concern; it is one we share. The reason why we are putting
every AI use case through that review process is to ensure
that, if it is being used with real veteran data, that it meets
VA's stringent security requirements.
Mr. Luttrell. Thank you, Mr. Chairman. Thank you, sir.
I yield back.
Mr. Barrett. I thank you.
Ms. Cherfilus-McCormick.
Ms. Cherfilus-McCormick. Thank you so much, thank you so
much.
I wanted to kind of piggyback off of some of Representative
Luttrell's questions. You mentioned standardization, and we
know now, from doing this for years, that standardization in
the VA has not been our strong suit. Are there any things that
you have learned from our lack of standardization for all of
our electronic medical records? We have been consistently
having an issue there with standardization. I have two
questions for you first. Are you confident that you can
actually have a standardization mechanism that will be able to
have a smooth transition implementation?
Mr. Worthington. Thank you for the question, and it is a
critical topic for us. I do think that the investments this
committee has helped make over the past years has helped with
that. We do have, for example, in the space of decision
support, we have an investment that allows AI-assisted decision
support tools to be purchased or built and then deployed to
every Veterans Health Information Systems and Technology
Architecture (VistA) site and also to every----
Ms. Cherfilus-McCormick. I guess my question really is,
like I said, we have been trying to be successful here, and it
has not been. How confident are you now? What are the missing
links for standardization when it comes to AI, because AI has
some complexities that I think we can all acknowledge,
especially when it comes to biases? If we are going to
implement AI into our system, we want to make sure that we have
precise implementation, and we are also taking into
consideration responsible implementation of AI, which actually
addresses the biases immediately, that deals with security
immediately. I was going to go into those questions first, but
I said, ``I cannot even go there if we do not deal with
standardization.'' What have we learned? How confident are you,
or should we really be taking some time to step back and look
at standardization again but through a magnifying glass to make
sure we get it right?
Mr. Worthington. I do feel confident that we are
approaching this in an enterprise approach. That is why
partnerships with the VHA and our colleagues, like Dr. Carey,
is so critical. AI is both a new area--it is one we need to be
able to experiment in before we commit to that enterprise
solution. Then, once we commit, we do not want to have, you
know, every medical center buying its own version of the same
product. We have got a pretty careful balance of that
innovation. We are doing structured pilots to help us decide
what to purchase and what to deploy to the enterprise.
Ms. Cherfilus-McCormick. I wanted to talk more about the
implementation development because we know that most of the
biases will be during the development phase and also the
implementation phase. What are you doing specifically to make
sure that these biases are not being inherently put into the
system, to make sure that all of our veterans actually have
access to equitable care?
Mr. Worthington. That is a great question, and it is a
concern that is of critical importance for us as we adopt AI.
The Office of Management and Budget in their policy has
determined, defined high-impact use cases. Those would be
things involved in healthcare benefits. They have provided us a
set of requirements that any AI needs to meet before they are
used. Some of the highlights of those are pre-deployment
testing to make sure the model performs well across different
demographic groups, but not just pre-deployment testing but
also ongoing monitoring so that we can make sure that the
models perform over time.
Ms. Cherfilus-McCormick. Could you tell me how you are
doing that? We have been reading--I have been loving this AI
conversation I have been looking at through all spectrums. One
of the articles that I am going to actually ask to put into the
record, it talks about the clinical decision-making the
implementations. I also want to hear from Ms. Harris about, are
we matching the need right now to identify bias?
Mr. Worthington. I do believe that we, through the AI use
case control process and the governance we put in place with
our partners in VHA, that we do have a commitment from all the
use case owners to meet those standards in the OMB
requirements.
Ms. Cherfilus-McCormick. Ms. Harris, what would you like to
see when it comes to actually being vigilant on making sure
that we are not utilizing a system that has inherent biases in
it?
Mr. Harris. For sure. One thing to note--even Mr.
Worthington talked about these high-impact systems--VHA has 72
percent of their AI use cases as being high impact, so meaning
that they affect people and entities and their rights. That is
quite a number, a high number of systems that have that
implication. Yes, you have to go through additional hoops, as
he had mentioned, with pre-deployment and during monitoring to
make sure that, you know, rights are not compromised. The VA
has told us that there is a need for more privacy officers to
handle increased data security demands. We would like to see
more of those positions being filled to ensure that privacy is
really taken care of as it relates to these high-impact uses
case.
Ms. Cherfilus-McCormick. I have a few seconds left, but I
did want to ask Dr. Ghassemi, are there any cases that you have
seen in public usage or private usage where they have done an
excellent job in actually removing the biases, identifying them
immediately?
Dr. Ghassemi. There is a really active domain of
researchers who are trying to solve exactly that problem. A lot
of the studies are happening with, for example, the Medical
Information Mart for Intensive Care (MIMIC) data base, which is
based out of the Boston area, something that I actually
contributed to.
To summarize, I think the broader domain of that research
activity, in a few words, is it is possible to do it, but it
requires a thoughtful approach, and each data set is different.
What you have in the VA and the bias in that will be different
than if you are doing it in the context of a data set in Boston
with somewhere else.
Ms. Cherfilus-McCormick. Thank you. I yield back. Thank you
for your time.
Mr. Barrett. Thank you.
I will recognize myself for 5 minutes again.
Dr. Ghassemi, I wanted to come back to you, and you have
listened to some of the back and forth testimony and some of
the responses, both from the VA and from members here. You are
outside of the VA. You have the benefit of being removed from
some of this internal stuff. I am curious, you know, kind of
what your thoughts are to me, and to Mr. Luttrell's point is we
are trying to upgrade this legacy health record system on a I
guess parallel track, to use the term you used. We are trying
to modernize some of the easy lift items that can be done
through assisted technology or augmented, I think somebody said
in their testimony as well. Do you think that is achievable,
number one? You know, how do you think that the VA can do this
responsibly to make sure that it is done in the appropriate
way?
Dr. Ghassemi. The short answer is I think it is achievable.
How can it be done responsibly? It has to start first and
foremost with unification of the data. I heard earlier
conversations that----
Mr. Barrett. In unification of data, are you talking about
having a singular system, or are you talking about the data
itself not being fragmented across all these different VA
facilities?
Dr. Ghassemi. What I mean is that you need a singular way
to represent the data so that an AI system that operates in one
system can move and operate in another. Now, actually the good
news is that artificial intelligence can be used to help with
that unification process itself. I will speak about some of my
external experiences here and say why I think there is room to
be helpful. It is a common problem in industry for corporations
to deal with. They have a large data base of customers, or
health systems have a large data base of patients, and they
want to enrich that with some data from outside of their
ecosystem. That is a common problem. There is reconciliation of
two complex data sets where column names in these data sets do
not match, representations of values inside these data sets do
not match. There is so many things that are misaligned here.
The same, instead of thinking of AI's role as coming in after
you gave done a very heavy duty and costly and inglorious task
of aligning that data, you can use the AI tools to perform
alignment of that data, right, to ask how you do the
combination of the information, the debiasing considerations
that were brought up earlier, and so on.
Mr. Barrett. Thank you, I appreciate that. How do you think
balancing, you know, the access to this and the benefit that
comes from it with keeping the paramount interest of, you know,
veterans' consent and privacy and all of those things that we
cannot miss the mark on as well? I would be interested in your
thoughts on that.
Dr. Ghassemi. Yes, I think disclosure is really important
transparency. You know, when we go to a supermarket and we turn
around an item that is on the shelf? On the back is disclosed
to us through nutrition label what are the contents inside of
the food that we purchase. In a similar way, if you think of
care that we receive as an item, then you need a similar way to
inspect what components, which parts of the ingredients in that
care came from which sources. Did they come from a model that
Oracle trained on their Cerner ecosystem? Did they come from an
academic paper? Did they come from a clinician's judgment? The
traceability of that decision and making it transparent back to
the end consumer of the care, which is the veteran, that is
really important because they have a right to know how care is
being derived prior to consenting to receive it. I think that
transparency sits at the beating heart of doing this correctly.
The reason there is trepidation, as far as I understand it,
behind the use of AI--not just in healthcare by the way, but in
a large number of industries, is because the transparency is an
issue, right? It could tell you--hallucinations--I think maybe
some of you have heard of this concept. If you have not, I will
quickly define it--is when a model basically confidently tells
you the wrong answer. There are ways to overcome this. They
require some expertise, but it is solvable.
Mr. Barrett. Thank you. I appreciate it. I am out of time.
Ranking Member Budzinski, I will recognize for you 5
minutes.
Ms. Budzinski. Thank you very much. September is Suicide
Prevention Month, and our full committee has not had a hearing
for many years on suicide prevention, which I think is
something that is a very big missed opportunity and something I
am hoping we can be getting to.
I can use this opportunity at this subcommittee hearing to
ask the VA some questions around suicide prevention and then
the connection with AI and how AI might be a useful tool
suicide in prevention, like the REACH VET algorithm model, in
particular. My question is for actually Dr. Carey. Can you
speak to how VA is planning to use its AI inventory to build on
this success?
Dr. Carey. Absolutely. Thank you so much for the question.
As you know, it is incredibly important that we take care of
our veterans, especially in this context of mental health
needs. We have been operating the REACH VET model for a number
of years, as Mr. Worthington noted, since 2017 successfully. We
have updated that model recently to ensure it has ongoing high
performance of identifying identification of veterans at the
highest risk core tiles. Then we implement that model as part
of a multipronged strategy to ensure veterans get the care they
need. Their receipt of the care they need does not depend only
on identification of an AI tool or being flagged as being at
high risk. It is just one of many strategies we use to ensure
that veterans are regularly screened, and, as you noted, in the
opening statement, if anybody falls through the cracks, that
they have an opportunity to still receive the care they need.
Ms. Budzinski. One of my concerns is just we do not want to
prevent human involvement from being a part of suicide
prevention. We can use AI as a tool. How does the VA look at-
you know, working to ensure that human involvement is not
eliminated as a part of the critical nature of the care that we
want to be able to provide to a veteran with suicide prevention
efforts?
Dr. Carey. Thank you. That is a fantastic question, and we
completely agree. I want to make it absolutely clear that VA
clinicians deliver care to veterans. VA clinicians are in
control of the care that veterans receive. While we do use AI
tools to surface risks and ensure that all veterans are flagged
to get the care they need, what happens next is that a human at
the VA reaches out to that veteran, where it first reviews the
information and decides if outreach is necessary.
Ms. Budzinski. Okay. Could you commit for me that the VA
will never use AI, including chatbots, as a substitute for
frontline staff responders for mental health crisis
intervention?
Dr. Carey. We do not currently have any plans that I am
aware of to use AI as a treatment device instead of providers.
I personally have been a part of many conversations where we
ensure that continues to be the case.
Ms. Budzinski. Okay. Thank you.
Ms. Harris, could I ask, what risks are posed by using AI
tools for use cases other than their intended purpose, like the
use of chatbots that were developed for programs like Veterans
Readiness and Employment (VR&E) or home loans and crisis
intervention support?
Ms. Harris. Well, I think that there would be significant
risks in a tool that is not being performed as intended.
For example, if you are using an AI chatbot for one
program, but, you know, obviously if you use that same bot for
another program, it is going to produce poor results. That is
because the data that was used to teach that tool would not be
relevant to the expected role for that other program. We would
certainly think that there is significant risks in dealing with
what you have asked.
Ms. Budzinski. Okay. Then I guess the VA's Office of
Inspector General reported in April that Veterans Benefits
Administrations (VBA) automated decision support tool was
ineffective in helping claims processes assign the correct
effective date for PACT Act claims. This resulted in at least
$7 billion in improper payments. I worry that VHA's rushed to
expand automation will lead to similar errors that could put
patients at risk.
Shifting gears, Mr. Worthington, how do you plan to measure
accuracy of implemented and piloted AI tools?
Mr. Worthington. That is a great question. I think by
having all of the use cases documented, along with the owner of
each AI use case, we will have the consistency plans available
to us so that then our colleagues and VHA can be regularly
following up to see what they found. We agree that continuous
monitoring of AI in production is very important.
I do think our healthcare system is particularly well
designed to monitor for those sorts of things because that is
part of what they do in a non-AI context as well.
Ms. Budzinski. Okay. Just a quick follow up. At the hearing
on this topic, Mr. Worthington, last year, you mentioned that
the key to understanding how any particular AI may introduce
biases is to understand the data that it was trained on and the
outputs it provides.
Considering the efforts of this administration to limit
what kind of data may be available in research data sets or in
a veteran's medical file, do you believe that this will impact
the efficacy of VA's AI tools?
Mr. Worthington. I would have to get into the specifics of
any given case. I think, at a high level, it is very important
to understand what data went into the training and do pre-
deployment testing before we use something in production.
Ms. Budzinski. Okay.
I yield back.
Mr. Barrett. Thank you. I will now recognize Mr. Luttrell.
Mr. Luttrell. Thank you, Mr. Chairman.
Dr. Ghassemi, I am fascinated with your previous statement.
Clean data, dirty data, retrospective, prospective data, the
transfer of information is very challenging. I am not going to
say impossible. I will never say that.
Currently, the VA does not house all of veterans' data. It
sits in the different silos of the different hospitals. I think
the death records lives in one spot, but everyone else is
assimilated, right? Correct?
Mr. Worthington. There is definitely siloed systems,
although our health data is pretty consolidated.
Mr. Luttrell. Consolidated. It make senses to me--and I do
not know the price tag on this, if this is even possible, that
if all the data lived in one enclave, the entire veterans space
lived under just say the VA data center--which I do not even
know what that would look like--but then the VA could control
access to anybody, including all the sites, plus every single
university and research student, whoever it wants to touch it,
and they could prevent the ability for data theft. Is that a
fair statement? Anybody?
Mr. Worthington. I do think that consolidating data into
secure platforms can be a good enabler of this sort of
technology for sure.
Mr. Luttrell. Are we even having that discussion inside the
VA? You can say no.
Mr. Worthington. Yes, we were actively working and, in
fact, have done a number of data consolidations to make that
possible.
Mr. Luttrell. I have been here for about 3 years now, and
the word ``activity working,'' it does not really resonate in
this place.
Are we really wanting to do this, or is this just something
that is just something you are throwing at me?
Mr. Worthington. No, I think like an example, like the
REACH Vet model that we just tried is a model that was created
based on that consolidated data set that draws on data from all
the different medical centers as well as other data into one
central data warehouse.
Mr. Luttrell. Everybody can touch it. If somebody in
Conroe, Texas, a VA facility that I have says, ``Hey, look I
have a veteran here that has this,'' they can reach out to that
data center, populate from tens of trillions of data points,
and send back, ``Hey, most likely this is what we are looking
at''?
Mr. Worthington. Well, when you are using it--it gets
complicated quickly, as you know.
Mr. Luttrell. I know.
Mr. Worthington. Different use cases have different degrees
of connectedness. In terms of building places where we can
create those models that we just went through like REACH VET,
we do already have investments that help with that.
Mr. Luttrell. Okay. If we do have the willingness to do
this, somebody is going to have to have the software in place
to do it. Mr. Ghatak, I am not going to let you out of here
without saying something. Okay.
Who can handle something like this? Company-wise, industry,
whoever? Do not say Michigan State because you are sitting in
the room with me.
Mr. Ghatak. No, sir. I would say University of Michigan
where I went to school, they could probably take----
Mr. Luttrell. They are pretty good, too? Okay. Yes.
Mr. Ghatak. Sir, I spent 4 years in the Federal Government,
I have worked in the General Services Administration under
Technology Transformation Services (TTS), and I had the
opportunity to work with a lot of different agencies in that
capacity. What I saw there was what I had seen throughout my
commercial career, which is, as I put in my written statement,
organizations have way more data than they realize. That data
exists in more locations than they are aware of. That data
means different things in different places at the fundamental
root level in terms of where the data exist. The number one
reason that projects fail--if it is an AI project or if it is
any other technology project, it is because of the data. If the
data is not there, then no matter what position, what solution
you have, it will never really work. It is sort of like what we
call lipstick on a pig, in other words. You have to solve that
problem.
Now, who solves that problem? That is an enterprise wide
problem. That is an enterprise wide acknowledgment that the
problem exists, and then an enterprise wide effort to make the
investment in solving that problem from a----
Mr. Luttrell. Multiple agencies are going to have to come
in on top of this.
Mr. Ghatak. I would say multiple departments within an
agency would report up through a business leader, a chief
officer, reporting up at the highest level to make that
investment and to solve that problem at the fundamental level.
Because if it is not solved fundamentally, then the underlying
structure of any solution will not work.
Mr. Luttrell. I am going to make the assumption, which I
probably should not. This is what is going to have to happen.
Yes?
Mr. Worthington. I think we need to find ways to get the
exact right piece of data from everything that VA and U.S.
Department of Defense (DOD) have access to the person that
needs it at the right time. I actually think that search-and-
summarization capability is actually one of the things that we
are excited about AI may be being able to help with.
Mr. Luttrell. This is what AI will do for us.
Mr. Worthington. I think it could help with those sorts of
things to sift through all----
Mr. Luttrell. I do not think the human brain could process
that many data sets.
Mr. Worthington. That is right. This is one of the areas we
are actively investing in.
Mr. Luttrell. I should not say that. The human brain could
absolutely do anything; the human being cannot.
Mr. Worthington. I think it gives an opportunity to empower
people to act on more information than they would be able to do
manually.
Mr. Luttrell. That is something that--the kind of
downstream I would like to--you know, I would like to see the--
how we are laying this out. At the end of the day, as
appropriators, in Congress, we are going to have to put a
dollar sign on that. Since Electronic Health Record (EHR) is
really giving us a great time, you kind of see where I am going
with this?
Thank you, Mr. Chairman. I yield back.
Mr. Barrett. Yes. Thank you, Mr. Luttrell.
I will recognize myself for 5 minutes.
Mr. Ghatak, you mentioned in your testimony, kind of
compared AI to the early days of the railroad, right? You know,
this was a great advancement, but it was fraught with all these
problems and challenges, and, you know, over time was
perfected--and I guess never truly perfected, but certainly
perfected to the degree that we can reasonably get to.
I think when it comes to artificial intelligence, there is
a greater risk than the occupants of a train rolling down a
railroad track. This could have catastrophic outcome if left,
you know, unguarded or breach of information or, you know, who
knows what. It could be truly problematic.
What are the guardrails that you think are appropriate and
necessary right now to make sure that that does not happen with
AI? Like how are we going to look over the horizon of what
could happen and prevent it from happening on the front end?
Mr. Ghatak. Thank you. It is a great question. I think it
is a very--it is sort of a fundamental question in terms of AI
and what it is and what it is not. As I said in my statement,
when you interact with AI tools today, it feels like you are
talking to a human being, but it is not a human being. It has
no moral conscience. It does not really understand the words
that is actually being given to it or the words that it is
producing.
There are a number of ways to really address this issue.
One of those is really understanding the difference between
correlation and causation without getting into great
statistical detail.
There is nearly a perfect correlation--as I put in my
testimony--in terms of the number of Google searches for the
word ``Nintendo'' and the number of librarians in the State of
Michigan. Most statistical models will rely on this relation--
since I am using correlation--to identify patterns and then
reproduce those patterns in its output. What is really needed
is an emphasis on causation, understanding the inputs that a
model uses, how those inputs relate to each other, and how
those relate to the outputs.
There is very little effort being placed on that type of
technology and that type of investment because the dollars are
already chasing correlation. Correlation is a lot easier to do
than causation. That is where a lot of the investment goes.
I would say one of fundamental areas is--and I do not know
if it can be mandated, but I would think--I would hope that the
scientific and research community would realize that is the
power of AI, is to unlock the true potential of it, is to
really mimic how a human mind works, which is it sees something
and reacts to it, and then produces something else. To mimic
that with other technology would be great.
The other thing that I did want to say is going back to the
data itself. A model is trained. I think the question was
around bias, right? The data that the model is given, if it is
not inherently debiased, if a lot of thought is not given to
the data itself that the model receives, then the output will
be inherently biased. It could be biased because of the way it
is engineered. It could be biased because of the data that it
is given. Because these models are so complex and so little
work has been done to understand how they work, we will never
know if it is the model that is biased or the data that was
biased.
Again, a principle, a development principle, a research
principle, a standardization that is adopted by industry to
address all of those would be very helpful.
Mr. Barrett. Yes, thank you. That correlation-causation
thing is really important. I would bet or guess that a lot of
information at the beginning is correlation information. Over
time, maybe it can be perfected or improved into the causative
and non-causative, you know, parts of that. At the beginning,
it is ``if this, then that'' correlation. We may not know why
or how, but these things, especially when you are dealing with
medical information over a long period of time, and, you know,
if enough people come in with a correlating condition, enough
times we begin to believe it is causative for a risk factor for
something else.
I guess how do we--like how do we make good decisions based
on that? You know, because we may not even understand the
causative nature of it, but if it is enough correlation data
there, maybe it does tell us something.
Mr. Ghatak. Absolutely. I think correlation has a purpose
in terms of identifying patterns or identifying things that are
outside of the norm, absolutely. It is a wonderful tool, and it
is a critical tool. My position would be that it just cannot be
used in a vacuum. That coupled with understanding causation and
investing more in those types of tools to help understand the
true relations between these things and why one is causing the
other.
As I mentioned, there is no obvious relationship between
Google searches and the number of librarians. The problem is
correlation models do not know that.
Mr. Barrett. Right.
Mr. Ghatak. They just Google that number and run with it.
There are a lot of crazy examples that I can give, but that is
a good one and relevant. That emphasis on causation I think is
really one that has not been invested in as much as it should
be. It is something that we found that is really helpful and
powerful that helps us understand our models and why they came
that way. Also, when they fail, we understand why they fail. It
is likely because something broke in that relationship or did
not work in that relationship.
Mr. Barrett. Thank you. I am out of time.
I am going to yield to Ranking Member Budzinski for 5
minutes.
Ms. Budzinski. Thank you. Thank you, Mr. Chairman.
I wanted to ask Ms. Harris some follow-up questions. Just
as ranking member, I have spent now a lot of time asking the VA
how it plans to juggle all of these different numerous
modernization efforts the Department is pursuing, like
Electronic Health Record Modernization (EHRM), of course,
supply chain, HR modernization, and now AI.
I was wondering if you could speak to the types of
resources that the VA will need to consider having at its
disposal as it deploys these systems.
Ms. Harris. Yes, thank you for the question. I mean, first
and foremost, I think it is hugely problematic that VA does not
have a permanent CIO in place. I know you mentioned it in your
opening statement. That is because, under his or her
leadership, that is where these, you know, various IT
modernizations get prioritized, you know. Plus our work has
shown that, you know, when you have that steady leadership
over, you know, 3-to 4-year time period, that is essential for
any successful major IT initiative, including all the AI
initiatives, Zero Trust, EHRM, all those things.
The second point, OIT is obviously going through a major
restructuring right now. They have requested almost $300
million less in Fiscal Year 2026 than the previous year. They
have also reduced staff by 931 staff.
Now more than ever, VA needs to fully understand--have the
comprehensive grasp on the skills and inventories that they
have in their IT workforce, and at this time, they do not know
that. They are not in a position to effectively assess what
they need if they do not know what they have. That is an open
recommendation that we have. That is first and foremost
something that they need to do in order to answer your
question.
Ms. Budzinski. Okay. Thank you.
Mr. Worthington, in GAO's review from July, VA noted that
it faced challenges with implementing generative AI use cases
due to a lack of sufficient technical resources and budget.
Your testimony highlights this issue of cost as well.
As it is currently funded and staffed, do you believe the
VA is capable of implementing additional AI-use cases on top of
these other modernization efforts that I have mentioned.
Mr. Worthington. Thank you for the question, Ranking
Member. I do think that we have the resources to implement
high-impact AI, but it is a tough environment. Everything is
competing for resources with each other. It is a matter of
prioritizing those things that are going to have the most
amount of veteran impact with the resources that we have.
Ms. Budzinski. Okay. I guess I just go back to what Ms.
Harris' recommendation, getting a CIO I think is really
critical to helping to prioritize all of these different really
important initiatives.
Mr. Worthington, do you believe the VA's challenges with
retaining AI experts and other technical employees may impact
VA's ability to scale AI tools and other modernization efforts?
Mr. Worthington. I definitely think having AI experts on
the VA side will help make us a better purchaser of these
solutions, and it is an important thing for us to do. We have
invested a lot in trying to build this team, especially through
partnerships with things like the United States Digital Corps
and the Presidential Innovation Fellows Program. You want to
lean into those sorts of partnerships to help us bring AI
experts in, in addition to those that we can recruit ourselves.
Ms. Budzinski. Okay. Great. Mr. Worthington, we are hearing
reports of VA's ambient listening pilot will be rolled out
across ten facilities by the end of this year. What is the
Department determining a success for this pilot?
Mr. Worthington. Thank you for the question. I will let Dr.
Carey give you some details on that.
Dr. Carey. Thank you for the question. We have established
a series of criteria and evaluation as we roll this out that is
focused on user acceptance testing, veterans' perceptions of
the tool as its used in their ongoing trust, and the care they
receive, and just overall performance of the tool. We will
continue to monitor that during the pilot.
Ms. Budzinski. Okay. Are you measuring clinician burden,
and what are your targets?
Dr. Carey. We are--I can take that for the record to get
back to you with the specifics. In general, we are measuring
clinician burden and getting clinician feedback both
synchronously and through survey mechanisms to understand the
impacts.
Mr. Worthington. One thing I would love to add is the users
of our generative AI tool that is deployed to the workforce as
a whole, in a survey, 73 percent of the users of that tool
reported that they were able to spend more time fully using
their professional skills, and 68 percent reported increased
job satisfaction. I do think that these tools are going to be
value adds to our workforce to help them do more to serve
veterans.
Ms. Budzinski. Well, it seems to me that we are placing a
massive burden on providers. That is a concern. From being an
ambassador to the tool for veterans, ensuring the tools'
accuracy, and then reporting and mediating issues as they
arise, how is the Department working to be proactive about
receiving feedback from providers on issues with this tool?
Dr. Carey. Thank you. That is a great question. Just
briefly, I want to recognize, it is so important to balance
that survey response burden and burden on the clinicians that
are also providing care. We have been partnering with
clinicians on day one, designing this as they are the end
users. We just have ongoing conversations with them about the
best way to balance those competing things.
Ms. Budzinski. Okay. Thank you. I yield back.
Mr. Barrett. Thank you. I will--we are going to close here
momentarily. I just have one quick question.
On that listening and automation transcribing, is that file
of that recording, is that deleted after it is transcribed? Is
there some protection there to make sure that it is not
archived or held someplace?
Mr. Worthington. We do have procedures on that and would be
happy to get that back to you for the record. I do not have the
details in front of me, but, yes, we have got that accounted
for.
Mr. Barrett. Thank you. I will now yield to Ranking Member
Budzinski for her closing statement.
Ms. Budzinski. Okay. Thank you. I just want to thank the
panelists for being here today to have this conversation. I do
very much appreciate it.
I do want to go back, though, Mr. Worthington, to a
conversation we had earlier about the six VA employees that had
been working with DOGE and a letter that Ranking Member Takano
had written to the VA back in June. We have not gotten a
response. We just want similar transparency around access to
the data that those six employees had. That is veterans' data.
I just want transparency and some additional information on
that.
Anything you can do to help us get a response back for
Ranking Member Takano would be very appreciated. Thank you.
Mr. Barrett. Thank you, Ranking Member Budzinski. I
appreciate it.
I want to thank our panelists and the members today for
joining us for this important hearing. This hearing has made
clear that VA has both made a tremendous--we have had both a
tremendous opportunity as well as a serious responsibility when
it comes do using artificial intelligence within the VA.
VA has access to some of the best data and research assets
in the world. I know Mr. Luttrell pointed that out in some of
the questioning too.
If used the right way, AI could help doctors detect cancer
earlier, prevent heart disease, cut down on paperwork, and,
most importantly, save veterans' lives and hopefully prevent
veteran suicides in the process.
Programs like REACH Vet show us it is possible when
technology is focused on the mission, and we can improve
outcomes. Let us be clear, AI is a tool, not a replacement for
doctors, nurses, and care teams. I appreciate the VA
stipulating that we are not trying to replace practitioners
with AI tools.
It can help identify risks earlier and provide clinical
pathways, but it cannot and must not replace treatment or human
judgment. That is the reason we send doctors to college, right,
because we want them to be experts on what they are doing.
Veterans deserve both cutting-edge technology and a strong
medical team working together on their behalf. That means
vigilance and self-responsibility--and a sense of
responsibility are still required. If VA fails to safeguard
veterans' data or to maintain transparency, trust will be lost,
and progress is going to stall.
This subcommittee will continue to hold the VA accountable
to ensure that AI enhances care, reduces red tape, and
strengthens--not substitutes--the human touch needed in
medicine.
I ask unanimous consent that all members have 5 legislative
days to revise and extend their remarks and include extraneous
material.
Without objection, that is so ordered, and this hearing is
adjourned.
[Whereupon, at 4:23 p.m., the subcommittee was adjourned.]
=======================================================================
A P P E N D I X
=======================================================================
Prepared Statements of Witnesses
----------
Prepared Statement of Charles Worthington
Chairman Barrett, Ranking Member Budzinski, and distinguished
Members of the Subcommittee, thank you for the opportunity to testify
regarding VA's opportunity to use Artificial Intelligence (AI) to
improve health care and services to Veterans. Your longstanding support
of Veterans and their families is greatly appreciated. I am accompanied
today by Dr. Evan Carey, Acting Director of the National Artificial
Intelligence Institute, Digital Health Office, Veterans Health
Administration.
While the use of AI at VA is not new, recent advances in the
capabilities of AI systems represent a significant opportunity for VA.
Many of the most time-consuming tasks VA employees and Veterans must
now complete manually could, in the future, be made dramatically faster
and more accurate when assisted by effective AI-enabled software. VA,
in partnership with industry, academia, and other Federal agencies, is
working rapidly to seize this opportunity.
VA's strategic vision is to make the Department an industry leader
in AI that improves Veterans' lives by delivering faster, higher
quality, and more cost-efficient services, with strong governance and
trust.
We have distilled this strategy into five execution priorities: (1)
expanding AI access across the VA workforce; (2) reimagining high-
impact workflows with AI and automation; (3) ensuring the most
promising AI projects receive prioritized investment; (4) building an
AI-ready workforce; and (5) running transparent and effective AI
governance.
To realize this strategy, VA is increasingly investing in AI-driven
tools that enhance productivity, reduce manual burden, and improve
service delivery to Veterans. In VA's 2024 inventory, we reported 227
AI use cases, representing nearly 100 more use cases than in the 2023
report. We expect this increase to continue in our 2025 report.
These investments are yielding tangible results. In one highly
anticipated use case, VA now offers an on-network generative AI tool
known as VA GPT. Over 85,000 users are engaged with the tool which
assists with basic administrative tasks such as drafting emails and
summarizing documents and meetings notes. A survey of VA GPT users
found that the tool saves its users an average of 2.5 hours per week,
with more than 80 percent agreeing that it has made them more
efficient.
Furthermore, we have successfully piloted and scaled an AI-assisted
software development tool called GitHub Copilot, now used by over 2,000
developers within OIT and our contract partners. These software
developers indicate this AI-assisted software development tool is
helping them deliver capabilities faster and saving them over 8 hours a
week. This includes faster development of Veteran-facing features on
VA.gov, making it easier to refill prescriptions and apply for
benefits, and the improvement of backend systems that accelerate claims
processing.
AI-augmented tools are also driving improvements in clinical care,
with 82 percent of the over 200 use cases in VA's inventory coming from
the Veterans Health Administration (VHA). VA's Stratification Tool for
Opioid Risk Mitigation (STORM) uses machine learning to identify and
mitigate the risk of overdose and suicide among Veterans prescribed
opioids or with opioid use disorder. By summarizing patient risk
factors, STORM identifies high risk Veterans for review by expert
health teams. Health care teams reviewed the care of over 28,700
Veterans identified by STORM in the past year alone, decreasing
mortality in high-risk patients by 22 percent. Since its launch in
2017, the REACH VET program has used tools like STORM to identify and
bring clinical attention to nearly 135,500 Veterans, improve outpatient
care, reduce suicide attempts, and decrease the number of mental health
emergencies.
Additionally, VHA has deployed 84 AI-assisted devices that have
been authorized by the , including one that uses computer vision to
enhance clinical outcomes such as early tumor detection. One VA study
showed that using AI-assisted colonoscopy devices increased adenoma
detection rates by 21 percent, which is associated with lower late-
stage cancer incidence and reduced mortality.
VA is committed to implementing innovative, AI-powered tools that
advance health care for Veterans, improve the experience of care teams,
and optimize VA's workforce. As part of this commitment, VA will pilot
ambient scribe technology at 10 sites beginning this fall. Ambient
scribe is an AI technology that listens to and documents the
conversations between health care providers and patients. AI processes
a transcript of the encounter to generate secondary outputs like
clinical encounter notes and coding recommendations. It has the
potential to transform health care by reducing clinician burdens,
enhancing efficacy, improving patient care quality and experience, and
engaging with clinical decision support services. Ultimately, it allows
the provider to spend more time face-to-face with Veterans.
As we progress, protecting Veterans' data privacy while responsibly
leveraging AI's potential is a top priority for the Department. Like
all software approved for use at VA, AI systems must meet VA's rigorous
security and privacy standards before they receive an Authority to
Operate. Additionally, consistent with the Office of Management and
Budget memorandum M-25-21, our team is facilitating an agency-level
review of each AI use case to ensure the tool meets the Government's
standards for innovation, governance, and public trust. Each use case
undergoes an AI Impact Assessment to identify and mitigate risks.
Further, VA has established and is committed to maintaining an
annual AI use case inventory. First released in December 2024, we are
on track to provide an update to this inventory in December 2025. This
inventory positions VA among the most transparent health care systems
in the country regarding AI.
Looking ahead, our focus over the next 12 months will be
implementing our strategic execution priorities by expanding employee
access to generative AI to 100 percent of VA staff, reimagining high-
impact workflows, prioritizing investment strategy to high return-on-
investment AI solutions, releasing new AI training opportunities for
employees, and maintaining transparent and effective AI governance by
ensuring 100 percent of VA's high-impact AI use cases meet the
Administration's standards.
Despite our industry-leading progress, VA acknowledges the adoption
of AI tools presents significant challenges. Among them is integrating
new AI solutions within VA's highly complex existing system
architecture, and as a Government entity entrusted with Veterans'
private information, balancing adoption of new and emerging tools and
vendors with the Government's strict security compliance standards is
crucial. Retention of AI experts is a challenge. Finally, scaling
commercial AI tools will incur additional costs, making it an ongoing
effort to align these costs with available technology funding. Cost is
one of many reasons the Department encourages Congress to fully fund VA
next year in lieu of another continuing resolution.
In conclusion, VA remains steadfast in its commitment to harnessing
the power of AI to improve the lives of Veterans. By strategically
investing in AI tools and enhancing our workforce's capabilities, we
aim to deliver faster, higher quality, and more cost-efficient
services. While we acknowledge the complexities and challenges inherent
in this transformation, we are dedicated to maintaining the highest
standards of governance, transparency, and ethical use of AI. With your
continued support, we can ensure that VA leads in AI innovation and
sets a benchmark for responsible AI use in public service. Thank you
for the opportunity to testify before you today. I look forward to your
questions.
Prepared Statement of Sid Ghatak
Chairman Barrett and distinguished Members of the Subcommittee:
Thank you for the opportunity to testify. My name is Sid Ghatak,
and for almost three decades, I have designed and deployed artificial
intelligence and forecasting systems across finance, healthcare,
pharmaceuticals, media, and government.
I currently serve as the Chief Technical Advisor for the National
Artificial Intelligence Association, the premier organization
representing over 1,500 businesses in the advancement of AI, and am
also the founder and Chief Executive Officer of Increase Alpha, LLC,
where we use artificial intelligence to predict stock prices and
license these predictions to hedge funds.
In the Federal Government, I served in the General Services
Administration for almost 4 years, where I was a Director of the Data &
Analytics Center of Excellence. In that role, I co-authored the Federal
AI Maturity Model 3 years before AI took the world by storm and
contributed to previous Executive Orders on AI, specifically on the
critical issues of data security and privacy.
At Increase Alpha, I architected a predictive AI model that
generates alpha once thought impossible--a deep learning system
exceptionally accurate at predicting equity prices. Increase Alpha far
exceeds multiple industry benchmarks, including accuracy, Sharpe ratio,
and alpha generation. This solution is not based on Large Language
Models but is a purpose-built predictive engine designed for a very
specific need.
I want to emphasize that it is entirely unrelated to the Department
of Veterans Affairs and has no bearing on today's testimony. I mention
it only as an example of how AI, when carefully designed with a clear
purpose, can achieve exceptional effectiveness.
Taken together, this diverse background--spanning academia,
government, and industry--has given me the rare opportunity to actually
build AI systems that work well in the real world. Because I have spent
my career outside the orthodox worlds of academia, venture capital, and
big tech, I am also not beholden to herd mentality. Instead, I bring an
expert, independent perspective which is especially valuable now, when
much of the world is caught up in the `art of the possible' with AI,
when what is most urgently needed is a sober understanding of what is
safe, practical, and ready to serve the public.
LLMs like ChatGPT, Claude, and Gemini are a powerful subset of AI,
but they come with their own set of problems, specifically in
healthcare, where hallucinations and sycophancy on the part of chatbots
can lead susceptible users down psychological rabbit holes. Which is
why it's important to clarify that AI is bigger than just ChatGPT and
its competitors.
To use an analogy: the steam engine transformed society, fueling
the Industrial Revolution. While steam power still exists today, it
gave way to other forms of power over time. Until steam engines were
used to create the first railroads, no human had ever traveled faster
than a horse. This new form of transportation opened the world's eyes
to what was possible, just as ChatGPT has shown the world the art of
the possible with AI. But early train travel was dangerously
unreliable. Accidents were frequent, derailments common, and thousands
of lives were lost before rail systems matured into the safe networks
we know today.
The lesson is clear: revolutionary technologies will evolve and
improve over time when the private sector and government work in
collaboration. The same applies to AI.
As the Committee gathers information on how to modernize technology
at the VA, I would like to offer three pieces of advice from my decades
at the front lines of building and implementing advanced analytical
solutions:
1. Expand the playing field: For the last several years, the
world has been consumed with Large Language Models to the point
where AI has become synonymous with it; however, that is not
the case. Many other types of AI may have similarities to these
models, but function very differently. Technologies that
specialize in interpreting and understanding images, video, and
audio, for example. Or technologies that are better suited to
working with numbers and symbols instead of words. And new tech
that has yet to be invented.
There is an old adage that when you are a hammer, everything
looks like a nail. The world has become so enamored with LLMs,
and rightfully so. Interacting with them can feel magical,
giving you the sense that they are real people, though they are
not. This may be why little to no investment is being made into
these other areas.
At Increase Alpha, we have demonstrated clearly what can be
done with other forms of Artificial Intelligence. I began
building our models at the same time as the research underlying
ChatGPT was published. I had also encountered the same compute,
cost, energy, and reliance on Nvidia GPUs issues we still see
today. I also took a different approach to conserve resources
and focus on simplification using Predictive Intelligence,
which led to lean AI models that use a minuscule amount of data
compared to LLMs, and which are small enough to run on a cell
phone.
Over 4 years, the success of my models directly contradicts the
notion that massive amounts of data--along with their
associated infrastructural and operational costs--are needed to
build AI solutions that are extremely accurate, innovative, and
reliable. Not to mention that they also consume ever-increasing
amounts of energy and utilize models that produce outputs that
are often incomprehensible and unexplainable. I have proven, in
one of the most competitive and challenging tech arenas, that
modern AI does not require all this if it is designed correctly
from the outset. The Administration, in its recent AI Action
Plan, does not limit AI to the narrow definition of LLM and
provides support for numerous types of technologies to be
developed.
2. Correlation is not Causation: The difference between
correlation and causation is best understood through an
example. There is a near-perfect correlation between the number
of Google searches for the word `Nintendo' and the number of
librarians in Michigan. It doesn't take a rocket scientist to
understand that there is no actual relationship between the two
trends.
Why is this so important? Because AI solutions today, such as
ChatGPT, are based on correlations, even if those correlations
are nonsensical. It is why they hallucinate (make up answers
based on nothing), and why they have an inherent bias. While
they give the impression of understanding and reasoning through
their rapid generation of coherent text, they have no idea what
the words themselves actually mean. They excel at predicting
the next best word based on a vast network of correlations and
are even better at providing the user with the answer they want
to hear, even if it's not accurate.
To achieve true artificial intelligence, these systems would
also have to know why the next word was predicted, which cannot
currently be explained. They would need to know the truth
behind every output. This is causation. That is how the human
mind works. Until AI systems can understand and explain the
`why' of their inner workings and outputs, and become reliable
sources of truth, they will never be truly intelligent. I
remain hopeful that I will experience this in my lifetime, but
it has not happened yet, nor is it likely to happen soon.
3. Data, Data, Data: AI is an engine that requires data. But
not just any data. Accurate, functional AI systems that produce
explainable and auditable outputs require vetted and cleaned
data, which we feel 100 percent confident using. By some
estimates, the Federal Government has more data than any other
organization in the world.
As a former Federal employee, I had the opportunity to work on
projects that required this type of clean data to achieve their
envisioned solutions. What I saw firsthand was the same thing I
had seen in every other large organization. There was always
more data than anyone realized. No one really knew where all of
it was located or what it meant, and the sheer effort to
gather, clean, and organize that data for proper use would have
been enormous and cost-prohibitive.
This is one of the key reasons many AI and data analytics
projects fail. Unless an organization is willing to make the
investments in organizing and cleaning their data, these
solutions--to put it bluntly--will be like lipstick on a pig.
They will not work over the long run, and even when they do,
they will not be reliable because they are not explainable.
We can see this already in current versions of Large Language
Models, which are aptly named because they are built on
unfathomably large amounts of language data. Some of it is
factually correct. Some of it is factually wrong. Some of it
has good intentions. Some of it is prejudiced, with built-in
hate, discrimination, and the bias of their very human authors.
As the old saying goes, garbage in, garbage out.
What does this all mean for the VA and the well-being and care of
our veterans? I can't claim to know. No one does. But I want to leave
you with a prediction of my own. I believe we truly are on the verge of
a revolution on the scale of the Industrial Revolution. So, if I could
leave you with one idea today, it would be this: AI is actually much
bigger than today's LLMs. And it is these technologies, many of which
have yet to be invented, that will enable the VA to execute its mission
``To fulfill President Lincoln's promise to care for those who have
served in our Nation's military and for their families, caregivers, and
survivors.''
In light of this, the Committee's work is vitally important to
ensure that the investments the Federal Government makes into AI
solutions will actually fulfill its mission. Our Veterans have given
their bodies, minds, and very lives so that we all can enjoy ours, and
we owe them more than our thanks and gratitude. We owe them the help
and services they need when and how they need them.
I am privileged to be here today, amongst my esteemed colleagues,
and I look forward to answering your questions. Thank you for this
opportunity.
Prepared Statement of Mohammad Ghassemi
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Prepared Statement of Carol Harris
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
[all]