Oversight: Advancing VA Care Through Artificial Intelligence

1[House Hearing, 119 Congress]2[From the U.S. Government Publishing Office]34                       ADVANCING VA CARE THROUGH5                        ARTIFICIAL INTELLIGENCE67=======================================================================89                                HEARING1011                               BEFORE THE1213                        SUBCOMMITTEE ON TECHNOLOGY14                               MODERNIZATION1516                                 OF THE1718                     COMMITTEE ON VETERANS' AFFAIRS1920                     U.S. HOUSE OF REPRESENTATIVES2122                    ONE HUNDRED NINETEENTH CONGRESS2324                             FIRST SESSION2526                               __________2728                       MONDAY, SEPTEMBER 15, 20252930                               __________3132                           Serial No. 119-343334                               __________3536       Printed for the use of the Committee on Veterans' Affairs3738[GRAPHIC NOT AVAILABLE IN TIFF FORMAT]3940                    Available via http://govinfo.gov4142                                __________4344                   U.S. GOVERNMENT PUBLISHING OFFICE4561-916                  WASHINGTON : 20254647-----------------------------------------------------------------------------------4849                     COMMITTEE ON VETERANS' AFFAIRS5051                     MIKE BOST, Illinois, Chairman5253AUMUA AMATA COLEMAN RADEWAGEN,       MARK TAKANO, California, Ranking54    American Samoa, Vice-Chairwoman      Member55JACK BERGMAN, Michigan               JULIA BROWNLEY, California56NANCY MACE, South Carolina           CHRIS PAPPAS, New Hampshire57MARIANNETTE MILLER-MEEKS, Iowa       SHEILA CHERFILUS-MCCORMICK,58GREGORY F. MURPHY, North Carolina        Florida59DERRICK VAN ORDEN, Wisconsin         MORGAN MCGARVEY, Kentucky60MORGAN LUTTRELL, Texas               DELIA RAMIREZ, Illinois61JUAN CISCOMANI, Arizona              NIKKI BUDZINSKI, Illinois62KEITH SELF, Texas                    TIMOTHY M. KENNEDY, New York63JEN KIGGANS, Virginia                MAXINE DEXTER, Oregon64ABE HAMADEH, Arizona                 HERB CONAWAY, New Jersey65KIMBERLYN KING-HINDS, Northern       KELLY MORRISON, Minnesota66    Mariana Islands67TOM BARRETT, Michigan6869                       Jon Clark, Staff Director70                  Matt Reel, Democratic Staff Director7172                SUBCOMMITTEE ON TECHNOLOGY MODERNIZATION7374                    TOM BARRETT, Michigan, Chairman7576NANCY MACE, South Carolina           NIKKI BUDZINSKI, Illinois, Ranking77MORGAN LUTTRELL, Texas                   Member78                                     SHEILA CHERFILUS-MCCORMICK,79                                         Florida8081Pursuant to clause 2(e)(4) of Rule XI of the Rules of the House, public82hearing records of the Committee on Veterans' Affairs are also83published in electronic form. The printed hearing record remains the84official version. Because electronic submissions are used to prepare85both printed and electronic versions of the hearing record, the process86of converting between various electronic formats may introduce87unintentional errors or omissions. Such occurrences are inherent in the88current publication process and should diminish as the process is89further refined.9091                         C  O  N  T  E  N  T  S9293                              ----------9495                       MONDAY, SEPTEMBER 15, 20259697                                                                   Page9899                           OPENING STATEMENTS100101The Honorable Tom Barrett, Chairman..............................     1102The Honorable Nikki Budzinski, Ranking Member....................     3103104                               WITNESSES105                                Panel I106107Mr. Charles Worthington, Chief Technology Officer & Chief108  Artificial Intelligence Officer, Office of Information &109  Technology, U.S. Department of Veterans Affairs................     5110111        Accompanied by:112113    Dr. Evan Carey, Ph.D., Acting Director, National Artificial114        Intelligence Institute, Digital Health Office, Veterans115        Health Administration, U.S. Department of Veterans116        Affairs117118Mr. Sid Ghatak, Chief Technical Advisor, National Artificial119  Intelligence Association.......................................     7120121Dr. Mohammad Ghassemi, Ph.D., Assistant Professor, Department of122  Computer Science and Engineering, College of Engineering,123  Michigan State University......................................     9124125Ms. Carol Harris, Director, Information Technology and126  Cybersecurity Issues, U.S. Government Accountability Office....    10127128                                APPENDIX129                    Prepared Statements Of Witnesses130131Mr. Charles Worthington Prepared Statement.......................    31132Mr. Sid Ghatak Prepared Statement................................    32133Dr. Mohammad Ghassemi, Ph.D. Prepared Statement..................    35134Ms. Carol Harris Prepared Statement..............................    40135136                       ADVANCING VA CARE THROUGH137                        ARTIFICIAL INTELLIGENCE138139                              ----------140141                       MONDAY, SEPTEMBER 15, 2025142143  Subcommittee on Technology Modernization,144                    Committee on Veterans' Affairs,145                             U.S. House of Representatives,146                                                    Washington, DC.147    The subcommittee met, pursuant to notice, at 2:58 p.m., in148room 360, Cannon House Office Building, Hon. Tom Barrett149(chairman of the subcommittee) presiding.150    Present: Representatives Barrett, Luttrell, Budzinski, and151Cherfilus-McCormick.152153           OPENING STATEMENT OF TOM BARRETT, CHAIRMAN154155    Mr. Barrett. The subcommittee will come to order.156    Without objection, the chair may declare a recess at any157time.158    Like many who have worn the uniform and received U.S.159Department of Veterans Affairs (VA) health care, I know the160frustration when the system is slow, the paperwork stacks up or161the technology fails, or does not lead us in the direction we162are trying to go. That is why this subcommittee's work is so163critical and why it is important that we have the folks here164joining us today.165    It is our duty to ensure VA's technology is efficient and166reliable, helping veterans rather than standing in the way of167their care. That brings us to the focus of today's hearing,168artificial intelligence, or AI, as it is of course commonly169referred to right now. For some, AI sounds like a science170fiction movie--we have all seen many of them--something only171computer scientists worry about or even something scary because172it is unknown and not well understood. It feels like today173everything is about drones or artificial intelligence. The174world has shifted quite a bit.175    Within VA, AI is already being used in ways they are making176a real difference for our veterans. In fact, U.S. Government177Accountability Office (GAO) recently released a report178highlighting how VA is among the most active adopters of179artificial intelligence, from analyzing medical images and180workflows to creating summary diagnostic reports. The report181identified more than 200 reported use cases across the system.182As we were preparing for this hearing, my staff had told me183about some, even very early prototype AI systems that the VA184had integrated decades ago.185    In clinical care, AI can help doctors detect cancer earlier186and identify warning signs of heart disease before a crisis187occurs.188    A recent study led by VA researchers at the VA Long Beach189Health Care System showed how AI can enable providers to detect190the risk of calcium buildup in heart arteries. This new191technology could give providers the chance to prevent heart192attacks rather than respond to them.193    AI is also being used to enhance mental healthcare. One of194the greatest challenges we face as a Nation remains around195veteran suicide. In 2017, VA launched the react vet program--or196the Recovery Engagement and Coordination for Health Veterans197Enhanced Treatment (REACH VET) program. This program uses an AI198model to help identify a very small group of veterans who are199at the greatest risk of suicide. The results were promising.200The program helped VA step in early, guiding veterans to care201before crisis strikes.202    It is not just about medical breakthroughs. AI is also203helping doctors and nurses relieve the day-to-day burden of204paperwork and things that are not spending time with the205patients. The most common complaint from providers is that they206spend too much time filling forms and not enough time taking207care of their patients.208    The VA is exploring AI tools, like Ambient scribes, which209can listen to a provider's conversation with the patient and210automatically create a clean and accurate medical note. On211average, this saves providers 2 or 3 hours per week. Now212multiply that across thousands of staff. It means more time213spent caring for veterans and less time staring at a computer214screen and doing paperwork. The promise of AI is real. I want215to be clear: Our job here is not applaud the promise. It is to216make sure that AI is being used responsibly, safely, and217transparently. Every great innovation comes with risk, and AI218is of course no exception. If the data is biased, the results219can be unfair. If safeguards are weak, privacy is compromised.220If the systems are not carefully monitored, mistakes could harm221the very people they are trying to protect. That is why the222governance of AI matters.223    The VA has been one of the first agencies in the government224to step up a framework for how AI should be reviewed and225approved. By law, VA already has some of the strongest privacy226protections in government, and those protections extend227directly to AI technology in veterans health and benefits, that228it cannot be used by vendors for other purposes, period. We are229going to hear more about that today.230    Veterans deserve to know when AI is being used in their231care. They deserve to know the technology has been tested and232that it is working for them, not against them. Congress233deserves to see evidence that taxpayer dollars are being well-234spent.235    I see today's hearing as an opportunity to highlight what236is working, to dig into what still needs improvement, and to237set clear expectations for the road ahead.238    AI does not and will not replace doctors or nurses, nor the239human touch that every veteran deserves when receiving their240care. Done right, AI can give clinicians another tool in their241tool box, helping them focus more on patients and ultimately on242saving lives.243    We need to find the balance between moving efficiently244enough to give veterans the benefit of innovation and245cautiously enough to make sure that no veteran is put at risk.246Our veterans should never be guinea pigs for untested247technology, but they should also not be denied the benefits of248safe and proven innovations.249    This subcommittee will hold VA to that standard, and I250intend to make sure that we get it right. It is not about251technology; it is about trust. Veterans give this Nation their252trust through their service. When they in turn go to the VA,253they deserve to know that trust will be given back and honored.254It is our duty to make sure that that trust is never broken. I255know the ranking member will join me in this pursuit, and I256appreciate your joining me in this committee today.257    Then, last, before I yield to her, I sat next to another258physician on my flight here last night. We were talking about259AI and actually about this hearing coming up. She was saying,260``We are never going to replace the element of patient care261that is done by doctors, nurses, and other medical262professionals.'' If we can expand the reach that they have,263look at it as, in the Army, we used to say a force multiplier264to deliver benefits in a more deliberative fashion that265benefits everybody. That is what we ought to be pursuing here.266    With that, I will yield to Ranking Member Budzinski for267your opening statement.268269      OPENING STATEMENT OF NIKKI BUDZINSKI, RANKING MEMBER270271    Ms. Budzinski. Thank you very much, Chairman Barrett. I272appreciate our subcommittee coming together to have a frank273conversation about the underlying Information Technology (IT)274challenges at the Department of Veterans Affairs and how we can275support the VA in closing those gaps through technology.276    However, today's review of artificial intelligence use277cases at the Veterans Health Administration (VHA) feels like a278distraction. VA is struggling with the basics. We are here279discussing the newest technologies while the VA is still280working with a crumbling IT infrastructure and still grapples281to modernize systems and workflows.282    As the ranking member on the Technology Modernization283Subcommittee, I am certainly excited by the potential of both284AI and innovation. AI could improve some of VA's challenges285through large language models and higher processing speeds. We286have seen promising studies of providers using AI to identify287cancers more easily, improve patient outcomes, and ease288clinician burnout by taking on more administrative tasks.289    The VA has certainly been the leader in the research and290development and widespread usage of a number of significant and291groundbreaking technologies. It stands to do so again with AI.292However, success in these efforts requires adequate resources293and investments in its budgets, its processes, and its people.294Veterans choose VA for the community it provides, the people it295employs, and for the fact that it is not driven by profit.296    What VA does best is make veterans feel seen and297understood. As we have seen, AI can be a tool to provide298decision support, ease provider burdens, and help with299notetaking so doctors can be more present with the patient. We300should also acknowledge that it is not the answer to every301challenge the VA faces. Also, we as a committee and as Congress302need to have a real conversation about AI policy and how to303implement it safely. I am excited about the opportunities that304AI presents. I am not convinced that VA is prepared to deploy305this technology just yet.306    I have a number of concerns that I hope to address today,307like the lack of regulation and governance structures and the308need for better transparency around what data is involved in309training such models.310    Further, like all technology modernization efforts,311implementing AI successfully requires a highly skilled,312adequately staffed workforce. Almost 2 weeks ago, the acting313head of the Department on Government Efficiency stressed the314need to ``hire and empower great tech talent in government.'' I315could not agree more with that. However, I think we should all316note the irony of that statement considering Office of317Information and Technology (OIT) is proposing a massive318reorganization and intends to cut at least 20 percent of its319workforce.320    Success is also reliant on strong IT leadership. If OIT is321in fact undergoing significant changes to its organizational322structure, priorities list, and workforce makeup, we need a323confirmed Chief Information Officer (CIO) at VA. This position324is particularly critical as we see the acceleration and325progression of modernization efforts at the Department. It326seems the VA still lacks a coherent enterprise IT strategy,327leaving projects AI integration to happen in silos. Without328stable and competent leadership, veterans and VA employees will329continue to be stuck with cobbled-together systems and330workflows that do not meet their needs rather than a solid331strategy for technology usage to guide its decision-making.332    I hope that we can get some clarity into the333administration's plan to propose a nominee for the CIO position334and that one can be confirmed before many of these substantial335changes occur.336    Last, I understand this subcommittee held a similar hearing337in January 2024, though neither I nor the chairman were on this338subcommittee at that point. In that hearing, data privacy was339an intrinsic part of the discussion. I hope that it still is340the case today.341    As we become more interconnected through technology342advancements like artificial intelligence, we must become343increasingly aware of the concerns about the privacy of users'344data, especially in healthcare. Since this last hearing, the345Department has been entangled in multiple cybersecurity346incidents, which have potentially placed veterans' data at347risk. Though many of these breaches have been targeted at VA348contractors, veterans' data has still been implicated, and VA349maintains some responsibility for its safety. Though I do feel350that this hearing is perhaps too early, considering VA has yet351to develop and release some of its policies and plans to align352its efforts with the administration's, I hope to hear from our353VA witnesses today about how data privacy and security, as well354as the views of both VA employees and patients, will be355integrated into such plans.356    Thank you and I yield back, Mr. Chairman.357    Mr. Barrett. Thank you, Ranking Member Budzinski.358    I join you in making sure that we have adequate ethics359guardrails around this, and certainly privacy is paramount in360that as well.361    I now want to introduce our witnesses. Again, thank you for362joining us today from the Department of Veterans Affairs, we363have Mr. Charles Worthington, the Chief Technology Officer and364Chief Artificial Intelligence Officer. Thank you for being365here. Accompanying Mr. Worthington is Dr. Evan Carey, Acting366Director over the National Artificial Intelligence Institute at367the VA. We also have Mr. Sid Ghatak.368    Did I say that correctly?369    Mr. Ghatak. Yes, sir.370    Mr. Barrett. Thank you. The chief technical advisor from371the National Artificial Intelligence Association. Dr. Mohammad372Ghassemi, assistant professor at Michigan State University. Go373green.374    Dr. Ghassemi. Go white.375    Mr. Barrett. Thank you for being here today as well.376    Finally, from the Government Accountability Office we have377Ms. Carol Harris, a familiar face to all of us on this378committee. Thank you again for being here and joining us. She379is also Director of IT and Cybersecurity at the GAO. Again,380thank you all for being here.381    At this time, I ask the witnesses to please stand and raise382your right-hand.383    [Witnesses sworn.]384    Mr. Barrett. Thank you. Let the record reflect that all385witnesses have answered in the affirmative.386    Mr. Charles Worthington, you are now recognized for 5387minutes to deliver your opening statement on behalf of VA.388389                STATEMENT OF CHARLES WORTHINGTON390391    Mr. Worthington. Chairman Barrett, Ranking Member392Budzinski, and distinguished members of the subcommittee, thank393you for the opportunity to discuss the Department of Veterans394Affairs' use of artificial intelligence to enhance healthcare395and services for veterans.396    Your steadfast support of the veterans and their families397is invaluable. I am joined today by Dr. Evan Carey, Acting398Director of the National AI Institute in the Digital Health399Office of the Veterans Health Administration.400    While AI is not new to VA, recent advancements in AI401systems presents a tremendous opportunity to improve VA's402services. When used effectively, AI can improve the efficiency403and accuracy of many time-consuming and error-prone tasks that404create burdens for VA staff and veterans alike. That is why VA405is rapidly working to capitalize on this technology.406    Our strategic vision is to make VA a leader in AI,407providing faster services, higher quality care, and more cost-408effective operations. We will aggressively deploy this new409technology while remaining committed to strong controls that410ensure security, privacy, and effectiveness of our technology411systems.412    We have distilled this vision into five key priorities.413First, we are aggressively expanding AI across our workforce.414Second, we are reimagining high-impact workflows through AI and415automation.416    Third we are prioritizing investment in data and417infrastructure that supports those high potential use cases.418    Fourth, we are cultivating an AI-ready workforce. Finally,419we are executing transparent and effective governance, an420essential requirement to maintain veterans' trust. We are421already bringing the strategy to life, making significant422investments in AI-driven tools.423    In 2024, our AI inventory had 227 use cases in it, which424was nearly 100 more than the previous year. We expect this425growth to continue in 2025 as we prepare for our December426update to that inventory. These investments are delivering427tangible results. I am pleased to report that all VA employees428now have access to secure generative AI tool to assist them429with their work. In surveys, users of this tool are reporting430that it is saving them over 2 hours per week.431    Additionally, over 2,000 VA staff and contract and software432developers are using an AI software development copilot tool,433enabling faster delivery of features that help veterans. AI is434also revolutionizing clinical care. In fact, 82 percent of VA's435AI use cases come from the Veterans Health Administration.436    VA's stratification tool for opioid risk mitigation uses437machine learning to identify veterans at high risk of overdose438and suicide, enabling healthcare teams to review and intervene439effectively. Since 2017, the REACH VET program, as you440mentioned, has used AI answer algorithms to identify over441130,000 veterans at elevated risk, improving outpatient care442and reducing suicide attempts.443    AI-assisted colonoscopy devices have increased adenoma444detection rates by 21 percent, reducing late stage cancer445incidents and mortality. Thanks to groundbreaking research by446folks like Dr. Raffi Hagopian and Dr. Evan Carey, the VA is447exploring how AI could help providers detect heart disease448earlier by reviewing the millions of Computed Tomography (CT)449scans that are not currently evaluated for cardiovascular450disease risk at all.451    As we advance our AI deployments, protecting veterans' data452remains paramount. All AI systems approved for use at VA must453meet VA rigorous security and privacy standards before454receiving an authority to operate. Additionally, consistent455with Office of Management and Budget's (OMB) policy, we conduct456a thorough agency-level review of each AI use case to ensure457that it meets the government the standards.458    We will publish the results of this review in our annual AI459inventory, positioning us as one of the most transparent460healthcare systems in the country with regards to our use of461artificial intelligence.462    Despite our progress, adopting AI tools does present463challenges. As you mentioned, integrating new AI solutions with464a complex system architecture and balancing innovation with465stringent security compliance is crucial. Recruiting and466retaining AI talent remains difficult. Scaling commercial AI467tools incurs additional costs. This underscores the importance468of full congressional funding for VA to continue this critical469work.470    In conclusion, the Department of Veterans Affairs is471committed to harnessing AI to improve the lives of veterans.472Through strategic investments in AI tools and workforce473capabilities, we strive to it deliver faster, higher quality,474and more cost-effective services. Your continued support is475vital for VA to lead in AI innovation and set a benchmark for476responsible AI use in government.477    Thank you for the opportunity to discuss our strategy, and478we look forward to your questions.479480    [The Prepared Statement Of Charles Worthington Appears In481The Appendix]482483    Mr. Barrett. Thank you, Mr. Worthington.484    The written statement of Mr. Worthington will be entered485into the hearing record.486    Mr. Ghatak, you are now recognized for 5 minutes to deliver487your opening statement.488489                    STATEMENT OF SID GHATAK490491    Mr. Ghatak. My name is Sid Ghatak, and for almost three492decades, I have designed and deployed artificial intelligence493forecasting systems across finance, healthcare,494pharmaceuticals, media, and government.495    I currently serve as the chief technology adviser for the496National Artificial Intelligence Association, the premier497organization representing 1,500 businesses in the advancement498of AI. I am also the founder and chief executive officer of499Increase Alpha, where we use artificial intelligence to predict500stock prices, and we license these predictions to hedge funds.501    In the Federal Government, I served in the General Services502Administration for 4 years where I was a Director of the Data503and Analytics Center of Excellence. In that role, I coauthored504the Federal AI maturity model 3 years before AI took the world505by storm. I also contributed previous executive orders on the506critical issues of data privacy and data security. At Increase507Alpha, I increased a predict--architected a predictive AI model508that generates off of once thought impossible, a deep learning509system that is exceptionally accurate at predicting equity510prices. Increase Alpha far exceeds multiple industry511benchmarks, including accuracy, sharp ratio, and alpha512generation. The solution itself is not based on large language513models at all, but it is purpose built, designed for this514specific need.515    I want to emphasize that this company and our solution is516completely unrelated to the Department of Veterans Affairs, and517it has no bearing on today's testimony. I mention it only as an518example of how AI, when carefully designed with a clear519purpose, can achieve exceptional effectiveness.520    Taken together, this diverse background, spanning academia521and government and industry, has given me the rare opportunity522to actually build AI systems that work well in the real world.523I have spent my career outside the orthodox roles of academia,524venture capital, and Big Tech, I am also not beholden to herd525mentality. Instead, I bring an expert independent perspective,526which is especially valuable now when much of the world is527caught up in the art of the possible with AI when what is most528urgently needed is a sober understanding of what is safe,529practical, and ready to serve the public.530    Large Language Models (LLM) like ChatGPT, Claude, and531Gemini are a powerful subset of AI, but they come with their532own set of problems, specifically in healthcare where533hallucinations and sycophancy on the part of ChatBots can lead534susceptible users down psychological rabbit holes, which is why535it is important to clarify that AI is bigger than just ChatGPT536and its competitors. To use an analogy, the steam engine537transformed society, fueling the Industrial Revolution. While538steam power exists today, it gave way to other forms of power539over time. Until steam engines were used to create the first540railroads, no human had ever traveled faster than a horse. This541new form of transportation opened the world's eyes to what is542possible, just as ChatGPT has shown the world the art of a543possible with artificial intelligence. Early train travel was544dangerously unreliable. Accidents were frequent, derailments545common, and thousands of lives were lost before rail systems546matured into safe networks that we know today.547    The lesson is clear: Revolutionary technologies will evolve548and improve over time when the private sector and the549government work in collaboration. The same applies to550artificial intelligence. As the committee gathers information551on how to modernize technology at the VA, I would like to offer552a few pieces of advice from my many decades on the front lines553of building and implementing advanced analytical solutions.554    As I mentioned, the last several years, the world has been555consumed with LLMs to the point where AI has become synonymous556with it. However, that is not the case. Many other types of AI557may have similarities to these models but function very558differently, technologies that specialize in interpreting and559understanding images, video, and audio, for example, or560technologies that are better suited to working with numbers and561symbols instead of words, a new technology that is yet to be562invented.563    There is an old adage about, when you are a hammer,564everything likes like a nail. The world has become so enamored565with LLMs, and rightfully so, interacting with them can feel566magical, giving you the sense that they are real people, but567they are not. This may be why little to no investment is being568made in these other areas. At Increase Alpha, we demonstrate569clearly what can be done with other forms of artificial570intelligence. I began building our models at the same time as571the research underlying ChatGPT was published. I had also572encountered the same compute cost energy and reliance on a573video that we still see today. I took a different approach to574conserve resources and focus on simplification, using575predictive intelligence which led to leading AI models that use576a minuscule amount of data compared to LLMs and which are small577enough to run on a cell phone.578    What does all this mean for the VA and the well-being and579care of veterans? I do not claim to know. No one really does. I580want to leave you with a prediction: I believe that we truly581are on the verge of a scale--of a revolution on the scale of582the Industrial Revolution. If I could leave you with one idea583today, it would be this: AI is much bigger than today's LLMs.584It is these technologies, many of which have yet to be585invented, that will enable the VA to execute on its mission.586Thank you.587588    [The Prepared Statement Of Sid Ghatak Appears In The589Appendix]590591    Mr. Barrett. Thank you, Mr. Ghatak.592    The written statement of Mr. Ghatak will be entered into593the hearing record. I appreciate your remarks. I think, if we594all use ChatGPT for cat memes, it will not be meeting its full595potential and leaving a lot of things behind. Thank you.596    Dr. Ghassemi, you are now recognized for 5 minutes for your597opening statement.598599                 STATEMENT OF MOHAMMAD GHASSEMI600601    Dr. Ghassemi. Chairman, Ranking Member, and members of the602subcommittee. Thank you for the opportunity to speak today. I603am a scientist and an entrepreneur focused on artificial604intelligence but especially its applications to healthcare. The605views I am going to share today are my own, but they are606informed by roles I played as a professor at Michigan State607University, where I direct a research laboratory on AI and its608applications to health sciences.609    I am also going to bring a perspective as the founder of an610AI consultancy Gamut Corporation, which has helped large611pharmaceutical companies, insurance companies, as well as612health systems, plan and execute their AI strategy.613    I want to be clear: I am not a veteran health specialist.614My perspective is on how artificial intelligence can broadly615advance care in ways directly relevant to the needs of616patients, and this very critically includes our veterans.617    This subcommittee has identified in their invitation letter618three priorities for AI health. These were transforming619healthcare delivery, streamlining services, and improving620outcomes. I am going to frame my remarks around three roles621that AI can play to help with these three priorities. The three622roles are automation, which is reducing low-value work through623the use of machines; augmentation, which is having a machine624assist a human in a task, so to strengthen clinical decision-625making, as an example; and insights, which is allowing us to626extract complex patterns from data, patterns far too complex627for us to discern just with our human intuitions alone. Let us628talk about three.629    First, AI can transform what happens during care itself.630Clinicians today spend hours on paperwork, but AI scribes can631generate notes automatically so they can focus more fully on632patients. We have heard that from more than one person in the633conversation today.634    In emergency rooms, decision tools powered by AI can help635identify the sickest patient sooner and get them treated636faster. Continuous monitoring assistance can pick up on the637early signs of decline, like sepsis, long before they would be638obvious to our human eyes. These tools make the encounter639safer, timelier, and more patient-centered.640    Second, AI cannot only streamline what happens during care;641it can streamline the plumbing of healthcare itself. Missed642appointments waste scarce clinician time. Automated reminder643systems, which do not have to use a large language model or a644sophisticated tool like ChatGPT, can reduce these no-shows and645save that time. Patients also too often fall between the cracks646between primary care and specialist visits. AI can flag the647missing referral information, track follow up, and prevent all648these gaps. When imaging or labs reveal unexpected findings,649like, God forbid, a lung nodule discovered by chance, AI650tracking systems can ensure these findings are followed up on651so that the treatable conditions do not get overlooked. This is652how we reduce wasted effort and ensure smoother, more reliable653care.654    In conclusion, artificial intelligence is not a silver655bullet. I say this as a person who has been working on656developing the methods for several years, but it can already657help with the subcommittee's three priorities. It works best658when it reduces low-value work, strengthens rather than659replaces clinical judgment, and turns complex data into660actionable insights.661    To succeed, we need disciplined pilots, clear metrics, and662safeguards for safety, equity, and privacy. If deployed with663care, AI can return time from paperwork to patients, ensure664that critical findings are not missed, and support clinicians665in their hardest decisions.666    I look forward to our conversation. I am grateful for the667invitation to be here with you today.668669    [The Prepared Statement Of Mohammad Ghassemi Appears In The670Appendix]671672    Mr. Barrett. Thank you, doc.673    The written statement of Dr. Ghassemi will be entered into674the hearing record.675    Ms. Harris, you are now recognized for 5 minutes to deliver676your opening statement on behalf of GAO.677678                   STATEMENT OF CAROL HARRIS679680    Mr. Harris. Chairman Barrett, Ranking Member Budzinski, and681members of the subcommittee, thank you for inviting us to682testify today on the use of artificial intelligence at VA.683Develops in generative AI, which is a subset of AI, which can684create text, images, video, and other content when prompted by685a user, have revolutionized how the technology can be used in686many industries, including healthcare and at VA and other687Federal agencies.688    AI holds substantial promise for improving the operations689of government agencies. However, it can increase risk for690agencies and poses unique oversight challenges because the691source of information used by AI is not always clear or692accurate. Given the fast pace at which AI is evolving, the693government must be proactive in understanding its complexities,694risks, and societal consequences.695    It should also be noted that VA has experienced696longstanding challenges in managing its IT projects and697programs, raising questions about the efficiency and698effectiveness of its operations and its ability to deliver699intended capabilities.700    As requested, I will briefly summarize our prior work on701the Department's AI use and challenges, as well as principles702and key practices for Federal agencies, including VA, that are703considering and implementing AI systems.704    In July 2025, we reported that VA's AI use cases increased705from 40 in 2023 to 229 in 2024. For example, VA is a developing706a generative AI use to automate various medical imaging707processes. This use may enhance VA's ability to analyze medical708images, integrate existing and new data workflows, and create709summary diagnostic reports.710    In the health and medical sector, agencies have adopted711generative AI to advance medical research and improve public712outcomes, including at VA. It is also worth noting that, of the713229 use cases, 64 percent were considered to be high-impact AI,714meaning that their capabilities impact the rights and/or safety715of individuals or entities. Looking at just VHA, that716percentage increases to 72 percent.717    The Department also reported to us a number of challenges718they face in using and managing generative AI. The full list is719noted in my written statement. I will only highlight a few720here.721    Challenge one, complying with existing Federal policies and722guidance. VA officials shared that the existing Federal AI723policy can present obstacles to the adoption of generative AI,724including in the areas of cybersecurity, data privacy, and IT725acquisitions.726    Challenge number two, having sufficient technical resources727and budget. Gen AI can require infrastructure with significant728computational and technical resources. VA noted challenges in729obtaining or accessing the needed technical resources and also730in having the funding necessary to establish those resources731and support desired AI initiatives.732    The last challenge, hiring and developing an AI workforce.733Among other things, the VA reported difficulties in734establishing and providing ongoing education and technical735skills development for their current workforce.736    VA officials told us they are working toward implementing737the new AI requirements in OMB's April 2025 memorandum. Doing738so will provide opportunities to develop and publicly release739AI strategies for identifying and removing barriers and740addressing the challenges I noted.741    Additionally, the GAO has identified a framework of key742practices to help ensure accountability and responsible AI use743in the design development, deployment, and continuous744monitoring of AI systems.745    Our framework is organized around four complimentary746principles that address governance, data, performance, and747monitoring. Consideration of the key practices in this748framework can help VA as it considers, collects, and implements749AI systems.750    Last, I will mention that we have 26 open recommendations751to VA concerning the management of its IT resources. If the752Department implements these recommendations effectively, it753will be better positioned to overcome its longstanding754challenges in managing its IT resources and will improve its755ability to address the rapidly changing AI landscape.756    That concludes my statement. I look forward to addressing757your questions.758759    [The Prepared Statement Of Carol Harris Appears In The760Appendix]761762    Mr. Barrett. Thank you, Ms. Harris.763    The written statement of Ms. Harris will be entered into764the hearing record.765    Again, thank you to all of our witnesses.766    We will now proceed to questioning. I will recognize myself767for 5 minutes to begin questioning.768    I am going to start with Mr. Worthington. The VA--I know we769have got a lot of concerns obviously about data security, data770privacy, what can be used, what can be modeled off of veteran771information. The VA requires vendors to sign contracts directly772stipulating that it will prevent secondary use of veteran data.773Number one, can you kind of walk us through how that works? How774are you making sure that companies actually follow that rule?775    Mr. Worthington. Thank you for the question, Chairman776Barrett. I think it is extremely important that everyone777understands that there is not a second set of rules for AI778systems. In the VA, we have a very clear and stringent set of779rules around both security and privacy for any technology780system. Before we bring a system into production, we have to781review that system for its compliance with those requirements782and ensure that the partners that are working with us on those783systems attest to and agree with those requirements. AI systems784receive an authority to operate just like any other system785would before we would put veteran data into the system.786    Mr. Barrett. Okay. I appreciate that. For example, though,787I know the large language model, kind of most stereotypical use788of AI, we are going to be looking at, you know, the millions of789records that the VA has and then modeling patient outcomes from790that and then looking kind of retrospectively to see where791people are on that spectrum today, and say, ``Well, we know, if792this condition led to 10 years later a worse condition over793here, how can we stem that off earlier?'' If we allow an AI794vendor to have access to that to cultivate that knowledge, is795that something that could be then used as an outgrowth in796another way for, like, another I guess research tool for other797things? For example, if a person has a predisposition to kidney798disease or diabetes or something like that, we can look799retrospectively at their health record to show that they had800certain indicators ahead of time, would not we want that to be801to the benefit of all medicine and not just within the VA?802    Mr. Worthington. Yes. I think that, as you are mentioning,803in the training phase of models, which VA does occasionally do,804that, if we work with a vendor, we make sure that the805agreements say that any protected health information can only806be used for that specific purpose that we have contracted with.807Often, that is taking place in environments that VA already808runs and controls.809    Now when we are talking about using a large language model,810those are provided typically via one of the big cloud service811providers, and those environments are set aside in a VA812boundary that basically the vendor has to attest that they813already meet VA security requirements. When we are sending814information to a large language model to get feedback back from815that model, we are using a version of that model that has been816made secure to meet government standards.817    Mr. Barrett. Okay. I will fully confess that I am not an818expert on this. Would a large language model allow a819practitioner to say, ``I have a veteran presenting with these820conditions; what are the risk factors that I ought to look for821to, maybe run tests that would not ordinarily be otherwise top822of mind?''823    Mr. Worthington. There could be a variety of AI approaches824for a use case like that. Dr. Carey may just quickly provide a825couple of examples of those sorts of decisions support type use826cases.827    Dr. Carey. Thank you. It is a fantastic question. I think828there are two versions of that. As you note, there are tools829where providers can get general advice, and they might830specifically articulate the needs of the veteran and for the831conditions that they are looking for, to point out to sort of832follow the different procedures that are recommended and833identify the guidelines. Those tools are available within the834VA.835    Mr. Barrett. Okay. After the passage of the The Sergeant836First Class Heath Robinson Honoring our Promise to Address837Comprehensive Toxics (PACT) Act, you know, we have this burn838pit registry and everything, and they are supposed to track839veterans and conditions that arose from that. Obviously, the840specific information about a particular veteran we want to have841protected and not revealed. If there are outcomes of that that842could be useful to, you know, human medicine in total, is there843a way for that to be revealed?844    Mr. Worthington. Yes, thank you for the question. VA does845have, as you know, a very large amount of health data. We have846a robust----847    Mr. Barrett. More than anybody in the world, I think.848    Mr. Worthington. That is right. We have a robust tradition849of research to advance not just VA healthcare but healthcare850overall. We are seeing an increasing interest in using that851data for AI-driven research papers, like the one that Dr. Carey852recently wrote.853    Mr. Barrett. Okay, and that is the--like the benefit but854also the concern is we obviously have a large repository of855medical data. If that is being used or to the benefit of a856curator of artificial intelligence, should the VA be, you know,857should that be brought into account for the cost of services858and other things like that? What I do not want is for a859provider to come in and leach that information out solely for860their benefit while not providing a benefit to the VA and to861the veterans as well.862    Mr. Worthington. We agree.863    Mr. Barrett. Thank you.864    Ranking Member Budzinski.865    Ms. Budzinski. Thank you, Mr. Chairman.866    Dr. Carey and Mr. Worthington, thank you so much for both867being here.868    I understand that several of VA's AI use cases, like the869ambient dictation pilot, intend to use an opt-in practice for870consent. For systems that are perhaps less directly veteran-871facing, like the use of AI in benefits determination or medical872assessments, how is the Department educating veterans on these873use cases to ensure for their awareness?874    Mr. Worthington. At a very high level--and thank you for875the question. We are using our AI use case inventory as the way876to catalogue all of the uses of AI and make sure that that is877publicly available. When there is not, as you mentioned, like a878one-to-one interaction that provides the opportunity to explain879directly what is happening, as there is in many healthcare880settings, what we are relying on is our publishing of the881overall AI strategy and use case to explain how the Department882is using AI in various products and services.883    Ms. Budzinski. Okay. Other than that general awareness--for884veterans, is there any way to kind of draw their attention to885this so that they know that, you know, what their situation886might be using to inform an AI model?887    Mr. Worthington. We are always listening for veterans'888feedback through a variety of mechanisms and reacting to that.889That is true of AI situations and non-AI situation. We890certainly want to monitor this for AI in particular, because I891think maintaining veterans' trust in VA as we introduce these892new technologies is going to be critical.893    Ms. Budzinski. Okay. Then, Mr. Worthington, I am glad that894you and your teams are committed to the transparency in AI use895cases at the Department. That is commendable. However, there896have been reports that certain employees had access to certain897data sets and systems within VA's enclave which may have been898used for AI related operations. I have some specific employees899I want to mention by name, and then I have some questions for900you. I am going to ask about these employees: Justin Fulcher,901Sahil Lavingia, Christopher Roussos, Payton Rehling, Cary902Volpert, or Jon Koval. I am just looking for, like, a yes or no903to these questions. Did you ever work with any of those904individuals?905    Mr. Worthington. Yes, I have come across several of them.906    Ms. Budzinski. Okay. Are or were these individuals907affiliated with the Department of Government Efficiency (DOGE)?908    Mr. Worthington. I am not exactly clear on the909relationship. I believe they are VA employees. At points, they910were introduced as also being part of the DOGE movement.911    Ms. Budzinski. Okay. Did any of these employees access data912sets that included VA patient medical records or other913personally identifiable information?914    Mr. Worthington. I am not aware.915    Ms. Budzinski. Okay. Were you or anyone you know ever asked916to duplicate data sets by these employees?917    Mr. Worthington. No, I was not.918    Ms. Budzinski. Okay. Can you commit to me that no veteran's919data was removed from the Department of Veterans Affairs?920    Mr. Worthington. As far as I understand, all the VA921employees follow all the VA IT security processes and922procedures and that was a key priority for all of us and always923is a key priority.924    Ms. Budzinski. Okay, Okay. Mr. Worthington, almost 2 weeks925ago, the Acting Director of the U.S. Digital Service noted that926the Federal Government needs more tech employees to--and to927hire and empower great talent. Do you believe that VA shares928that sentiment?929    Mr. Worthington. Yes, I do. I think having technologists in930government is critically important, as is having great931researchers and doctors.932    Ms. Budzinski. Okay. Secretary Collins has often noted the933importance of VA employees in direct care roles, disregarding934the importance of what he might call support employees in the935provision of this work. Do you believe that this type of936rhetoric has helped the Department to recruit and retain tech937talent?938    Mr. Worthington. I think the good thing about working at939the VA is our mission is so clear. The mission of serving940veterans is the most important one that I have worked on in my941tech career. I think there are many technologists across the942country that are willing to sign up for that mission. I love943trying to recruit those people on my team.944    Ms. Budzinski. Ms. Harris, real quick on a follow up, GAO's945Artificial Intelligence Accountability Framework notes the946workforce is a key component to ensuring effective AI947application. How does a highly skilled technical workforce948ensure adequate scalability of AI applications and protection949of veteran data?950    Mr. Harris. Well, while there is great excitement around AI951because of the potential to improve operations, there is also952significant concerns, the ones that I articulated earlier about953cybersecurity, intellectual property, as well built-in bias in954the AI system, as well as environmental and other concerns. We955want to make sure that we have a workforce that understands956both the potential of these systems but also understands the957risks in the AI well. Having those two are vital.958    Ms. Budzinski. Okay. Thank you.959    I yield back.960    Mr. Barrett. Thank you.961    Mr. Luttrell.962    Mr. Luttrell. Mr. Chairman.963    Mr. Ghassemi, you laid out a well-articulated plan of964attack on how the VA could tackle this healthcare, artificial965intelligence kind of combining of forces. The problem is you966have--it sounds like you never worked with the U.S. Government967because that is what kills this effort is the U.S. Government.968    Ms. Harris, your opening statement was very well-969articulated, and you hit every single point precisely. The970problem is we have such an issue with the VA because it is a971big machine, and we are trying to compound--we are trying to972bring in artificial intelligence to streamline the process. You973have 172 different VA facilities, plus satellite campuses, and974that is 172 different silos. They do not work together. They do975not communicate very well with each other. We have spent almost976$16 billion trying to push electronic healthcare records across977multiple facilities. Now we are going to try to tackle978artificial intelligence as well. In 2024, we had 229 AI979actions. Correct, Mr. Worthington?980    Mr. Worthington. Yes, approximately.981    Mr. Luttrell. What site did that come from, because I would982dare say that that did not come from all or every single VA983installation. That sounds like to me that that is collected984from, like, a few. Is that correct?985    Mr. Worthington. We did attempt to have a pretty986comprehensive review process to gather all of the uses of AI987across the country. We----988    Mr. Luttrell. I did not get anything out of that. That was989almost a yes-or-no question, but go ahead again.990    Mr. Worthington. Yes, I believe that AI is being used at991facilities across the country. This inventory covers those992uses.993    Mr. Luttrell. The conversations I have with multiple sites994is they do not have artificial intelligence capabilities995because their sites are not ready or they do not have the996infrastructure in place to do that, because we keep compounding997software on top of software. Some sites cannot function at all998with the new software they are trying to implement. That is a999pretty fair statement, correct?1000    Mr. Worthington. I would agree that having standardized1001systems is a challenge at the VA. There is a bit of a1002difference in different facilities. Although I do think many of1003them are starting to use AI-assisted medical devices, for1004example, and a number of those are covered in this inventory.1005    Mr. Luttrell. How do we fix this problem? Again, I am going1006to ask you, sir, because I usually ask everyone who sits in1007front of me from the VA: How would we fix this problem? Mr.1008Ghatak and Mr. Ghassemi have probably thought about this quite1009a bit before they showed up in front of us, but again they have1010not--actually, I do not know this for certain--I may be1011throwing this at you, and course correct me if you would like--1012but I do not think they have had to deal with the U.S.1013Government and also the VA. Now how long have you been in this1014position, sir?1015    Mr. Worthington. I have been at the VA nearly 10 years and1016this position for about 2 years as chief AI officer.1017    Mr. Luttrell. Okay. What comes first, the communication1018between the sites and the ability to ask that information1019questions, which we do not do that or we do not have the1020ability to do that--do we run the implementation of artificial1021intelligence in parallel with that, or do we have to do one1022before the other?1023    Mr. Worthington. In my personal opinion, we cannot wait,1024because AI is here, whether we are ready or not. Increasingly,1025every solution we buy from our partners in the private sector1026is going to have it embedded inside of it. I think our1027challenge is we need to come up with very good standard1028templates that every site can use and allow those standard1029tools to be deployed, things like the VA GPT school that I1030mentioned, which is now available to every VA employee in a1031standard way.1032    Mr. Luttrell. Since the Department of Veterans Affairs1033houses the most important data set on the planet arguably, and1034everyone wants to touch it, including Dr. Ghassemi at Michigan1035State--I would have to guess, especially when you were at1036Massachusetts Institute of Technology (MIT) in Cambridge I am1037sure. Pretty impressive resume, sir. Everybody is trying to1038touch it. Everybody wants to be a part of it, and you have to1039deal with every single subject-matter expert that walks through1040your door that says, ``I am the best.'' I can assure you every1041one of those corporations and companies walks into our office1042as well. Question is, who is it? Who do you vet, and who is1043going to touch it, because it cannot be everybody? We do not1044have it--in my personal opinion, that I am not aware of, we do1045not have an enclave that can house all of that information1046where everybody can get in there and not steal it.1047Implementation of artificial intelligence, which we do not have1048the ability to regulate, so the question is who will do that,1049or do you have the AI system itself regulate itself?1050    Mr. Worthington. I think it is a great observation and1051concern; it is one we share. The reason why we are putting1052every AI use case through that review process is to ensure1053that, if it is being used with real veteran data, that it meets1054VA's stringent security requirements.1055    Mr. Luttrell. Thank you, Mr. Chairman. Thank you, sir.1056    I yield back.1057    Mr. Barrett. I thank you.1058    Ms. Cherfilus-McCormick.1059    Ms. Cherfilus-McCormick. Thank you so much, thank you so1060much.1061    I wanted to kind of piggyback off of some of Representative1062Luttrell's questions. You mentioned standardization, and we1063know now, from doing this for years, that standardization in1064the VA has not been our strong suit. Are there any things that1065you have learned from our lack of standardization for all of1066our electronic medical records? We have been consistently1067having an issue there with standardization. I have two1068questions for you first. Are you confident that you can1069actually have a standardization mechanism that will be able to1070have a smooth transition implementation?1071    Mr. Worthington. Thank you for the question, and it is a1072critical topic for us. I do think that the investments this1073committee has helped make over the past years has helped with1074that. We do have, for example, in the space of decision1075support, we have an investment that allows AI-assisted decision1076support tools to be purchased or built and then deployed to1077every Veterans Health Information Systems and Technology1078Architecture (VistA) site and also to every----1079    Ms. Cherfilus-McCormick. I guess my question really is,1080like I said, we have been trying to be successful here, and it1081has not been. How confident are you now? What are the missing1082links for standardization when it comes to AI, because AI has1083some complexities that I think we can all acknowledge,1084especially when it comes to biases? If we are going to1085implement AI into our system, we want to make sure that we have1086precise implementation, and we are also taking into1087consideration responsible implementation of AI, which actually1088addresses the biases immediately, that deals with security1089immediately. I was going to go into those questions first, but1090I said, ``I cannot even go there if we do not deal with1091standardization.'' What have we learned? How confident are you,1092or should we really be taking some time to step back and look1093at standardization again but through a magnifying glass to make1094sure we get it right?1095    Mr. Worthington. I do feel confident that we are1096approaching this in an enterprise approach. That is why1097partnerships with the VHA and our colleagues, like Dr. Carey,1098is so critical. AI is both a new area--it is one we need to be1099able to experiment in before we commit to that enterprise1100solution. Then, once we commit, we do not want to have, you1101know, every medical center buying its own version of the same1102product. We have got a pretty careful balance of that1103innovation. We are doing structured pilots to help us decide1104what to purchase and what to deploy to the enterprise.1105    Ms. Cherfilus-McCormick. I wanted to talk more about the1106implementation development because we know that most of the1107biases will be during the development phase and also the1108implementation phase. What are you doing specifically to make1109sure that these biases are not being inherently put into the1110system, to make sure that all of our veterans actually have1111access to equitable care?1112    Mr. Worthington. That is a great question, and it is a1113concern that is of critical importance for us as we adopt AI.1114The Office of Management and Budget in their policy has1115determined, defined high-impact use cases. Those would be1116things involved in healthcare benefits. They have provided us a1117set of requirements that any AI needs to meet before they are1118used. Some of the highlights of those are pre-deployment1119testing to make sure the model performs well across different1120demographic groups, but not just pre-deployment testing but1121also ongoing monitoring so that we can make sure that the1122models perform over time.1123    Ms. Cherfilus-McCormick. Could you tell me how you are1124doing that? We have been reading--I have been loving this AI1125conversation I have been looking at through all spectrums. One1126of the articles that I am going to actually ask to put into the1127record, it talks about the clinical decision-making the1128implementations. I also want to hear from Ms. Harris about, are1129we matching the need right now to identify bias?1130    Mr. Worthington. I do believe that we, through the AI use1131case control process and the governance we put in place with1132our partners in VHA, that we do have a commitment from all the1133use case owners to meet those standards in the OMB1134requirements.1135    Ms. Cherfilus-McCormick. Ms. Harris, what would you like to1136see when it comes to actually being vigilant on making sure1137that we are not utilizing a system that has inherent biases in1138it?1139    Mr. Harris. For sure. One thing to note--even Mr.1140Worthington talked about these high-impact systems--VHA has 721141percent of their AI use cases as being high impact, so meaning1142that they affect people and entities and their rights. That is1143quite a number, a high number of systems that have that1144implication. Yes, you have to go through additional hoops, as1145he had mentioned, with pre-deployment and during monitoring to1146make sure that, you know, rights are not compromised. The VA1147has told us that there is a need for more privacy officers to1148handle increased data security demands. We would like to see1149more of those positions being filled to ensure that privacy is1150really taken care of as it relates to these high-impact uses1151case.1152    Ms. Cherfilus-McCormick. I have a few seconds left, but I1153did want to ask Dr. Ghassemi, are there any cases that you have1154seen in public usage or private usage where they have done an1155excellent job in actually removing the biases, identifying them1156immediately?1157    Dr. Ghassemi. There is a really active domain of1158researchers who are trying to solve exactly that problem. A lot1159of the studies are happening with, for example, the Medical1160Information Mart for Intensive Care (MIMIC) data base, which is1161based out of the Boston area, something that I actually1162contributed to.1163    To summarize, I think the broader domain of that research1164activity, in a few words, is it is possible to do it, but it1165requires a thoughtful approach, and each data set is different.1166What you have in the VA and the bias in that will be different1167than if you are doing it in the context of a data set in Boston1168with somewhere else.1169    Ms. Cherfilus-McCormick. Thank you. I yield back. Thank you1170for your time.1171    Mr. Barrett. Thank you.1172    I will recognize myself for 5 minutes again.1173    Dr. Ghassemi, I wanted to come back to you, and you have1174listened to some of the back and forth testimony and some of1175the responses, both from the VA and from members here. You are1176outside of the VA. You have the benefit of being removed from1177some of this internal stuff. I am curious, you know, kind of1178what your thoughts are to me, and to Mr. Luttrell's point is we1179are trying to upgrade this legacy health record system on a I1180guess parallel track, to use the term you used. We are trying1181to modernize some of the easy lift items that can be done1182through assisted technology or augmented, I think somebody said1183in their testimony as well. Do you think that is achievable,1184number one? You know, how do you think that the VA can do this1185responsibly to make sure that it is done in the appropriate1186way?1187    Dr. Ghassemi. The short answer is I think it is achievable.1188How can it be done responsibly? It has to start first and1189foremost with unification of the data. I heard earlier1190conversations that----1191    Mr. Barrett. In unification of data, are you talking about1192having a singular system, or are you talking about the data1193itself not being fragmented across all these different VA1194facilities?1195    Dr. Ghassemi. What I mean is that you need a singular way1196to represent the data so that an AI system that operates in one1197system can move and operate in another. Now, actually the good1198news is that artificial intelligence can be used to help with1199that unification process itself. I will speak about some of my1200external experiences here and say why I think there is room to1201be helpful. It is a common problem in industry for corporations1202to deal with. They have a large data base of customers, or1203health systems have a large data base of patients, and they1204want to enrich that with some data from outside of their1205ecosystem. That is a common problem. There is reconciliation of1206two complex data sets where column names in these data sets do1207not match, representations of values inside these data sets do1208not match. There is so many things that are misaligned here.1209The same, instead of thinking of AI's role as coming in after1210you gave done a very heavy duty and costly and inglorious task1211of aligning that data, you can use the AI tools to perform1212alignment of that data, right, to ask how you do the1213combination of the information, the debiasing considerations1214that were brought up earlier, and so on.1215    Mr. Barrett. Thank you, I appreciate that. How do you think1216balancing, you know, the access to this and the benefit that1217comes from it with keeping the paramount interest of, you know,1218veterans' consent and privacy and all of those things that we1219cannot miss the mark on as well? I would be interested in your1220thoughts on that.1221    Dr. Ghassemi. Yes, I think disclosure is really important1222transparency. You know, when we go to a supermarket and we turn1223around an item that is on the shelf? On the back is disclosed1224to us through nutrition label what are the contents inside of1225the food that we purchase. In a similar way, if you think of1226care that we receive as an item, then you need a similar way to1227inspect what components, which parts of the ingredients in that1228care came from which sources. Did they come from a model that1229Oracle trained on their Cerner ecosystem? Did they come from an1230academic paper? Did they come from a clinician's judgment? The1231traceability of that decision and making it transparent back to1232the end consumer of the care, which is the veteran, that is1233really important because they have a right to know how care is1234being derived prior to consenting to receive it. I think that1235transparency sits at the beating heart of doing this correctly.1236The reason there is trepidation, as far as I understand it,1237behind the use of AI--not just in healthcare by the way, but in1238a large number of industries, is because the transparency is an1239issue, right? It could tell you--hallucinations--I think maybe1240some of you have heard of this concept. If you have not, I will1241quickly define it--is when a model basically confidently tells1242you the wrong answer. There are ways to overcome this. They1243require some expertise, but it is solvable.1244    Mr. Barrett. Thank you. I appreciate it. I am out of time.1245    Ranking Member Budzinski, I will recognize for you 51246minutes.1247    Ms. Budzinski. Thank you very much. September is Suicide1248Prevention Month, and our full committee has not had a hearing1249for many years on suicide prevention, which I think is1250something that is a very big missed opportunity and something I1251am hoping we can be getting to.1252    I can use this opportunity at this subcommittee hearing to1253ask the VA some questions around suicide prevention and then1254the connection with AI and how AI might be a useful tool1255suicide in prevention, like the REACH VET algorithm model, in1256particular. My question is for actually Dr. Carey. Can you1257speak to how VA is planning to use its AI inventory to build on1258this success?1259    Dr. Carey. Absolutely. Thank you so much for the question.1260As you know, it is incredibly important that we take care of1261our veterans, especially in this context of mental health1262needs. We have been operating the REACH VET model for a number1263of years, as Mr. Worthington noted, since 2017 successfully. We1264have updated that model recently to ensure it has ongoing high1265performance of identifying identification of veterans at the1266highest risk core tiles. Then we implement that model as part1267of a multipronged strategy to ensure veterans get the care they1268need. Their receipt of the care they need does not depend only1269on identification of an AI tool or being flagged as being at1270high risk. It is just one of many strategies we use to ensure1271that veterans are regularly screened, and, as you noted, in the1272opening statement, if anybody falls through the cracks, that1273they have an opportunity to still receive the care they need.1274    Ms. Budzinski. One of my concerns is just we do not want to1275prevent human involvement from being a part of suicide1276prevention. We can use AI as a tool. How does the VA look at-1277you know, working to ensure that human involvement is not1278eliminated as a part of the critical nature of the care that we1279want to be able to provide to a veteran with suicide prevention1280efforts?1281    Dr. Carey. Thank you. That is a fantastic question, and we1282completely agree. I want to make it absolutely clear that VA1283clinicians deliver care to veterans. VA clinicians are in1284control of the care that veterans receive. While we do use AI1285tools to surface risks and ensure that all veterans are flagged1286to get the care they need, what happens next is that a human at1287the VA reaches out to that veteran, where it first reviews the1288information and decides if outreach is necessary.1289    Ms. Budzinski. Okay. Could you commit for me that the VA1290will never use AI, including chatbots, as a substitute for1291frontline staff responders for mental health crisis1292intervention?1293    Dr. Carey. We do not currently have any plans that I am1294aware of to use AI as a treatment device instead of providers.1295I personally have been a part of many conversations where we1296ensure that continues to be the case.1297    Ms. Budzinski. Okay. Thank you.1298    Ms. Harris, could I ask, what risks are posed by using AI1299tools for use cases other than their intended purpose, like the1300use of chatbots that were developed for programs like Veterans1301Readiness and Employment (VR&E) or home loans and crisis1302intervention support?1303    Ms. Harris. Well, I think that there would be significant1304risks in a tool that is not being performed as intended.1305    For example, if you are using an AI chatbot for one1306program, but, you know, obviously if you use that same bot for1307another program, it is going to produce poor results. That is1308because the data that was used to teach that tool would not be1309relevant to the expected role for that other program. We would1310certainly think that there is significant risks in dealing with1311what you have asked.1312    Ms. Budzinski. Okay. Then I guess the VA's Office of1313Inspector General reported in April that Veterans Benefits1314Administrations (VBA) automated decision support tool was1315ineffective in helping claims processes assign the correct1316effective date for PACT Act claims. This resulted in at least1317$7 billion in improper payments. I worry that VHA's rushed to1318expand automation will lead to similar errors that could put1319patients at risk.1320    Shifting gears, Mr. Worthington, how do you plan to measure1321accuracy of implemented and piloted AI tools?1322    Mr. Worthington. That is a great question. I think by1323having all of the use cases documented, along with the owner of1324each AI use case, we will have the consistency plans available1325to us so that then our colleagues and VHA can be regularly1326following up to see what they found. We agree that continuous1327monitoring of AI in production is very important.1328    I do think our healthcare system is particularly well1329designed to monitor for those sorts of things because that is1330part of what they do in a non-AI context as well.1331    Ms. Budzinski. Okay. Just a quick follow up. At the hearing1332on this topic, Mr. Worthington, last year, you mentioned that1333the key to understanding how any particular AI may introduce1334biases is to understand the data that it was trained on and the1335outputs it provides.1336    Considering the efforts of this administration to limit1337what kind of data may be available in research data sets or in1338a veteran's medical file, do you believe that this will impact1339the efficacy of VA's AI tools?1340    Mr. Worthington. I would have to get into the specifics of1341any given case. I think, at a high level, it is very important1342to understand what data went into the training and do pre-1343deployment testing before we use something in production.1344    Ms. Budzinski. Okay.1345    I yield back.1346    Mr. Barrett. Thank you. I will now recognize Mr. Luttrell.1347    Mr. Luttrell. Thank you, Mr. Chairman.1348    Dr. Ghassemi, I am fascinated with your previous statement.1349Clean data, dirty data, retrospective, prospective data, the1350transfer of information is very challenging. I am not going to1351say impossible. I will never say that.1352    Currently, the VA does not house all of veterans' data. It1353sits in the different silos of the different hospitals. I think1354the death records lives in one spot, but everyone else is1355assimilated, right? Correct?1356    Mr. Worthington. There is definitely siloed systems,1357although our health data is pretty consolidated.1358    Mr. Luttrell. Consolidated. It make senses to me--and I do1359not know the price tag on this, if this is even possible, that1360if all the data lived in one enclave, the entire veterans space1361lived under just say the VA data center--which I do not even1362know what that would look like--but then the VA could control1363access to anybody, including all the sites, plus every single1364university and research student, whoever it wants to touch it,1365and they could prevent the ability for data theft. Is that a1366fair statement? Anybody?1367    Mr. Worthington. I do think that consolidating data into1368secure platforms can be a good enabler of this sort of1369technology for sure.1370    Mr. Luttrell. Are we even having that discussion inside the1371VA? You can say no.1372    Mr. Worthington. Yes, we were actively working and, in1373fact, have done a number of data consolidations to make that1374possible.1375    Mr. Luttrell. I have been here for about 3 years now, and1376the word ``activity working,'' it does not really resonate in1377this place.1378    Are we really wanting to do this, or is this just something1379that is just something you are throwing at me?1380    Mr. Worthington. No, I think like an example, like the1381REACH Vet model that we just tried is a model that was created1382based on that consolidated data set that draws on data from all1383the different medical centers as well as other data into one1384central data warehouse.1385    Mr. Luttrell. Everybody can touch it. If somebody in1386Conroe, Texas, a VA facility that I have says, ``Hey, look I1387have a veteran here that has this,'' they can reach out to that1388data center, populate from tens of trillions of data points,1389and send back, ``Hey, most likely this is what we are looking1390at''?1391    Mr. Worthington. Well, when you are using it--it gets1392complicated quickly, as you know.1393    Mr. Luttrell. I know.1394    Mr. Worthington. Different use cases have different degrees1395of connectedness. In terms of building places where we can1396create those models that we just went through like REACH VET,1397we do already have investments that help with that.1398    Mr. Luttrell. Okay. If we do have the willingness to do1399this, somebody is going to have to have the software in place1400to do it. Mr. Ghatak, I am not going to let you out of here1401without saying something. Okay.1402    Who can handle something like this? Company-wise, industry,1403whoever? Do not say Michigan State because you are sitting in1404the room with me.1405    Mr. Ghatak. No, sir. I would say University of Michigan1406where I went to school, they could probably take----1407    Mr. Luttrell. They are pretty good, too? Okay. Yes.1408    Mr. Ghatak. Sir, I spent 4 years in the Federal Government,1409I have worked in the General Services Administration under1410Technology Transformation Services (TTS), and I had the1411opportunity to work with a lot of different agencies in that1412capacity. What I saw there was what I had seen throughout my1413commercial career, which is, as I put in my written statement,1414organizations have way more data than they realize. That data1415exists in more locations than they are aware of. That data1416means different things in different places at the fundamental1417root level in terms of where the data exist. The number one1418reason that projects fail--if it is an AI project or if it is1419any other technology project, it is because of the data. If the1420data is not there, then no matter what position, what solution1421you have, it will never really work. It is sort of like what we1422call lipstick on a pig, in other words. You have to solve that1423problem.1424    Now, who solves that problem? That is an enterprise wide1425problem. That is an enterprise wide acknowledgment that the1426problem exists, and then an enterprise wide effort to make the1427investment in solving that problem from a----1428    Mr. Luttrell. Multiple agencies are going to have to come1429in on top of this.1430    Mr. Ghatak. I would say multiple departments within an1431agency would report up through a business leader, a chief1432officer, reporting up at the highest level to make that1433investment and to solve that problem at the fundamental level.1434Because if it is not solved fundamentally, then the underlying1435structure of any solution will not work.1436    Mr. Luttrell. I am going to make the assumption, which I1437probably should not. This is what is going to have to happen.1438Yes?1439    Mr. Worthington. I think we need to find ways to get the1440exact right piece of data from everything that VA and U.S.1441Department of Defense (DOD) have access to the person that1442needs it at the right time. I actually think that search-and-1443summarization capability is actually one of the things that we1444are excited about AI may be being able to help with.1445    Mr. Luttrell. This is what AI will do for us.1446    Mr. Worthington. I think it could help with those sorts of1447things to sift through all----1448    Mr. Luttrell. I do not think the human brain could process1449that many data sets.1450    Mr. Worthington. That is right. This is one of the areas we1451are actively investing in.1452    Mr. Luttrell. I should not say that. The human brain could1453absolutely do anything; the human being cannot.1454    Mr. Worthington. I think it gives an opportunity to empower1455people to act on more information than they would be able to do1456manually.1457    Mr. Luttrell. That is something that--the kind of1458downstream I would like to--you know, I would like to see the--1459how we are laying this out. At the end of the day, as1460appropriators, in Congress, we are going to have to put a1461dollar sign on that. Since Electronic Health Record (EHR) is1462really giving us a great time, you kind of see where I am going1463with this?1464    Thank you, Mr. Chairman. I yield back.1465    Mr. Barrett. Yes. Thank you, Mr. Luttrell.1466    I will recognize myself for 5 minutes.1467    Mr. Ghatak, you mentioned in your testimony, kind of1468compared AI to the early days of the railroad, right? You know,1469this was a great advancement, but it was fraught with all these1470problems and challenges, and, you know, over time was1471perfected--and I guess never truly perfected, but certainly1472perfected to the degree that we can reasonably get to.1473    I think when it comes to artificial intelligence, there is1474a greater risk than the occupants of a train rolling down a1475railroad track. This could have catastrophic outcome if left,1476you know, unguarded or breach of information or, you know, who1477knows what. It could be truly problematic.1478    What are the guardrails that you think are appropriate and1479necessary right now to make sure that that does not happen with1480AI? Like how are we going to look over the horizon of what1481could happen and prevent it from happening on the front end?1482    Mr. Ghatak. Thank you. It is a great question. I think it1483is a very--it is sort of a fundamental question in terms of AI1484and what it is and what it is not. As I said in my statement,1485when you interact with AI tools today, it feels like you are1486talking to a human being, but it is not a human being. It has1487no moral conscience. It does not really understand the words1488that is actually being given to it or the words that it is1489producing.1490    There are a number of ways to really address this issue.1491One of those is really understanding the difference between1492correlation and causation without getting into great1493statistical detail.1494    There is nearly a perfect correlation--as I put in my1495testimony--in terms of the number of Google searches for the1496word ``Nintendo'' and the number of librarians in the State of1497Michigan. Most statistical models will rely on this relation--1498since I am using correlation--to identify patterns and then1499reproduce those patterns in its output. What is really needed1500is an emphasis on causation, understanding the inputs that a1501model uses, how those inputs relate to each other, and how1502those relate to the outputs.1503    There is very little effort being placed on that type of1504technology and that type of investment because the dollars are1505already chasing correlation. Correlation is a lot easier to do1506than causation. That is where a lot of the investment goes.1507    I would say one of fundamental areas is--and I do not know1508if it can be mandated, but I would think--I would hope that the1509scientific and research community would realize that is the1510power of AI, is to unlock the true potential of it, is to1511really mimic how a human mind works, which is it sees something1512and reacts to it, and then produces something else. To mimic1513that with other technology would be great.1514    The other thing that I did want to say is going back to the1515data itself. A model is trained. I think the question was1516around bias, right? The data that the model is given, if it is1517not inherently debiased, if a lot of thought is not given to1518the data itself that the model receives, then the output will1519be inherently biased. It could be biased because of the way it1520is engineered. It could be biased because of the data that it1521is given. Because these models are so complex and so little1522work has been done to understand how they work, we will never1523know if it is the model that is biased or the data that was1524biased.1525    Again, a principle, a development principle, a research1526principle, a standardization that is adopted by industry to1527address all of those would be very helpful.1528    Mr. Barrett. Yes, thank you. That correlation-causation1529thing is really important. I would bet or guess that a lot of1530information at the beginning is correlation information. Over1531time, maybe it can be perfected or improved into the causative1532and non-causative, you know, parts of that. At the beginning,1533it is ``if this, then that'' correlation. We may not know why1534or how, but these things, especially when you are dealing with1535medical information over a long period of time, and, you know,1536if enough people come in with a correlating condition, enough1537times we begin to believe it is causative for a risk factor for1538something else.1539    I guess how do we--like how do we make good decisions based1540on that? You know, because we may not even understand the1541causative nature of it, but if it is enough correlation data1542there, maybe it does tell us something.1543    Mr. Ghatak. Absolutely. I think correlation has a purpose1544in terms of identifying patterns or identifying things that are1545outside of the norm, absolutely. It is a wonderful tool, and it1546is a critical tool. My position would be that it just cannot be1547used in a vacuum. That coupled with understanding causation and1548investing more in those types of tools to help understand the1549true relations between these things and why one is causing the1550other.1551    As I mentioned, there is no obvious relationship between1552Google searches and the number of librarians. The problem is1553correlation models do not know that.1554    Mr. Barrett. Right.1555    Mr. Ghatak. They just Google that number and run with it.1556There are a lot of crazy examples that I can give, but that is1557a good one and relevant. That emphasis on causation I think is1558really one that has not been invested in as much as it should1559be. It is something that we found that is really helpful and1560powerful that helps us understand our models and why they came1561that way. Also, when they fail, we understand why they fail. It1562is likely because something broke in that relationship or did1563not work in that relationship.1564    Mr. Barrett. Thank you. I am out of time.1565    I am going to yield to Ranking Member Budzinski for 51566minutes.1567    Ms. Budzinski. Thank you. Thank you, Mr. Chairman.1568    I wanted to ask Ms. Harris some follow-up questions. Just1569as ranking member, I have spent now a lot of time asking the VA1570how it plans to juggle all of these different numerous1571modernization efforts the Department is pursuing, like1572Electronic Health Record Modernization (EHRM), of course,1573supply chain, HR modernization, and now AI.1574    I was wondering if you could speak to the types of1575resources that the VA will need to consider having at its1576disposal as it deploys these systems.1577    Ms. Harris. Yes, thank you for the question. I mean, first1578and foremost, I think it is hugely problematic that VA does not1579have a permanent CIO in place. I know you mentioned it in your1580opening statement. That is because, under his or her1581leadership, that is where these, you know, various IT1582modernizations get prioritized, you know. Plus our work has1583shown that, you know, when you have that steady leadership1584over, you know, 3-to 4-year time period, that is essential for1585any successful major IT initiative, including all the AI1586initiatives, Zero Trust, EHRM, all those things.1587    The second point, OIT is obviously going through a major1588restructuring right now. They have requested almost $3001589million less in Fiscal Year 2026 than the previous year. They1590have also reduced staff by 931 staff.1591    Now more than ever, VA needs to fully understand--have the1592comprehensive grasp on the skills and inventories that they1593have in their IT workforce, and at this time, they do not know1594that. They are not in a position to effectively assess what1595they need if they do not know what they have. That is an open1596recommendation that we have. That is first and foremost1597something that they need to do in order to answer your1598question.1599    Ms. Budzinski. Okay. Thank you.1600    Mr. Worthington, in GAO's review from July, VA noted that1601it faced challenges with implementing generative AI use cases1602due to a lack of sufficient technical resources and budget.1603Your testimony highlights this issue of cost as well.1604    As it is currently funded and staffed, do you believe the1605VA is capable of implementing additional AI-use cases on top of1606these other modernization efforts that I have mentioned.1607    Mr. Worthington. Thank you for the question, Ranking1608Member. I do think that we have the resources to implement1609high-impact AI, but it is a tough environment. Everything is1610competing for resources with each other. It is a matter of1611prioritizing those things that are going to have the most1612amount of veteran impact with the resources that we have.1613    Ms. Budzinski. Okay. I guess I just go back to what Ms.1614Harris' recommendation, getting a CIO I think is really1615critical to helping to prioritize all of these different really1616important initiatives.1617    Mr. Worthington, do you believe the VA's challenges with1618retaining AI experts and other technical employees may impact1619VA's ability to scale AI tools and other modernization efforts?1620    Mr. Worthington. I definitely think having AI experts on1621the VA side will help make us a better purchaser of these1622solutions, and it is an important thing for us to do. We have1623invested a lot in trying to build this team, especially through1624partnerships with things like the United States Digital Corps1625and the Presidential Innovation Fellows Program. You want to1626lean into those sorts of partnerships to help us bring AI1627experts in, in addition to those that we can recruit ourselves.1628    Ms. Budzinski. Okay. Great. Mr. Worthington, we are hearing1629reports of VA's ambient listening pilot will be rolled out1630across ten facilities by the end of this year. What is the1631Department determining a success for this pilot?1632    Mr. Worthington. Thank you for the question. I will let Dr.1633Carey give you some details on that.1634    Dr. Carey. Thank you for the question. We have established1635a series of criteria and evaluation as we roll this out that is1636focused on user acceptance testing, veterans' perceptions of1637the tool as its used in their ongoing trust, and the care they1638receive, and just overall performance of the tool. We will1639continue to monitor that during the pilot.1640    Ms. Budzinski. Okay. Are you measuring clinician burden,1641and what are your targets?1642    Dr. Carey. We are--I can take that for the record to get1643back to you with the specifics. In general, we are measuring1644clinician burden and getting clinician feedback both1645synchronously and through survey mechanisms to understand the1646impacts.1647    Mr. Worthington. One thing I would love to add is the users1648of our generative AI tool that is deployed to the workforce as1649a whole, in a survey, 73 percent of the users of that tool1650reported that they were able to spend more time fully using1651their professional skills, and 68 percent reported increased1652job satisfaction. I do think that these tools are going to be1653value adds to our workforce to help them do more to serve1654veterans.1655    Ms. Budzinski. Well, it seems to me that we are placing a1656massive burden on providers. That is a concern. From being an1657ambassador to the tool for veterans, ensuring the tools'1658accuracy, and then reporting and mediating issues as they1659arise, how is the Department working to be proactive about1660receiving feedback from providers on issues with this tool?1661    Dr. Carey. Thank you. That is a great question. Just1662briefly, I want to recognize, it is so important to balance1663that survey response burden and burden on the clinicians that1664are also providing care. We have been partnering with1665clinicians on day one, designing this as they are the end1666users. We just have ongoing conversations with them about the1667best way to balance those competing things.1668    Ms. Budzinski. Okay. Thank you. I yield back.1669    Mr. Barrett. Thank you. I will--we are going to close here1670momentarily. I just have one quick question.1671    On that listening and automation transcribing, is that file1672of that recording, is that deleted after it is transcribed? Is1673there some protection there to make sure that it is not1674archived or held someplace?1675    Mr. Worthington. We do have procedures on that and would be1676happy to get that back to you for the record. I do not have the1677details in front of me, but, yes, we have got that accounted1678for.1679    Mr. Barrett. Thank you. I will now yield to Ranking Member1680Budzinski for her closing statement.1681    Ms. Budzinski. Okay. Thank you. I just want to thank the1682panelists for being here today to have this conversation. I do1683very much appreciate it.1684    I do want to go back, though, Mr. Worthington, to a1685conversation we had earlier about the six VA employees that had1686been working with DOGE and a letter that Ranking Member Takano1687had written to the VA back in June. We have not gotten a1688response. We just want similar transparency around access to1689the data that those six employees had. That is veterans' data.1690I just want transparency and some additional information on1691that.1692    Anything you can do to help us get a response back for1693Ranking Member Takano would be very appreciated. Thank you.1694    Mr. Barrett. Thank you, Ranking Member Budzinski. I1695appreciate it.1696    I want to thank our panelists and the members today for1697joining us for this important hearing. This hearing has made1698clear that VA has both made a tremendous--we have had both a1699tremendous opportunity as well as a serious responsibility when1700it comes do using artificial intelligence within the VA.1701    VA has access to some of the best data and research assets1702in the world. I know Mr. Luttrell pointed that out in some of1703the questioning too.1704    If used the right way, AI could help doctors detect cancer1705earlier, prevent heart disease, cut down on paperwork, and,1706most importantly, save veterans' lives and hopefully prevent1707veteran suicides in the process.1708    Programs like REACH Vet show us it is possible when1709technology is focused on the mission, and we can improve1710outcomes. Let us be clear, AI is a tool, not a replacement for1711doctors, nurses, and care teams. I appreciate the VA1712stipulating that we are not trying to replace practitioners1713with AI tools.1714    It can help identify risks earlier and provide clinical1715pathways, but it cannot and must not replace treatment or human1716judgment. That is the reason we send doctors to college, right,1717because we want them to be experts on what they are doing.1718    Veterans deserve both cutting-edge technology and a strong1719medical team working together on their behalf. That means1720vigilance and self-responsibility--and a sense of1721responsibility are still required. If VA fails to safeguard1722veterans' data or to maintain transparency, trust will be lost,1723and progress is going to stall.1724    This subcommittee will continue to hold the VA accountable1725to ensure that AI enhances care, reduces red tape, and1726strengthens--not substitutes--the human touch needed in1727medicine.1728    I ask unanimous consent that all members have 5 legislative1729days to revise and extend their remarks and include extraneous1730material.1731    Without objection, that is so ordered, and this hearing is1732adjourned.1733    [Whereupon, at 4:23 p.m., the subcommittee was adjourned.]17341735=======================================================================17361737                         A  P  P  E  N  D  I  X17381739=======================================================================17401741                    Prepared Statements of Witnesses17421743                              ----------17441745               Prepared Statement of Charles Worthington17461747    Chairman Barrett, Ranking Member Budzinski, and distinguished1748Members of the Subcommittee, thank you for the opportunity to testify1749regarding VA's opportunity to use Artificial Intelligence (AI) to1750improve health care and services to Veterans. Your longstanding support1751of Veterans and their families is greatly appreciated. I am accompanied1752today by Dr. Evan Carey, Acting Director of the National Artificial1753Intelligence Institute, Digital Health Office, Veterans Health1754Administration.1755    While the use of AI at VA is not new, recent advances in the1756capabilities of AI systems represent a significant opportunity for VA.1757Many of the most time-consuming tasks VA employees and Veterans must1758now complete manually could, in the future, be made dramatically faster1759and more accurate when assisted by effective AI-enabled software. VA,1760in partnership with industry, academia, and other Federal agencies, is1761working rapidly to seize this opportunity.1762    VA's strategic vision is to make the Department an industry leader1763in AI that improves Veterans' lives by delivering faster, higher1764quality, and more cost-efficient services, with strong governance and1765trust.1766    We have distilled this strategy into five execution priorities: (1)1767expanding AI access across the VA workforce; (2) reimagining high-1768impact workflows with AI and automation; (3) ensuring the most1769promising AI projects receive prioritized investment; (4) building an1770AI-ready workforce; and (5) running transparent and effective AI1771governance.1772    To realize this strategy, VA is increasingly investing in AI-driven1773tools that enhance productivity, reduce manual burden, and improve1774service delivery to Veterans. In VA's 2024 inventory, we reported 2271775AI use cases, representing nearly 100 more use cases than in the 20231776report. We expect this increase to continue in our 2025 report.1777    These investments are yielding tangible results. In one highly1778anticipated use case, VA now offers an on-network generative AI tool1779known as VA GPT. Over 85,000 users are engaged with the tool which1780assists with basic administrative tasks such as drafting emails and1781summarizing documents and meetings notes. A survey of VA GPT users1782found that the tool saves its users an average of 2.5 hours per week,1783with more than 80 percent agreeing that it has made them more1784efficient.1785    Furthermore, we have successfully piloted and scaled an AI-assisted1786software development tool called GitHub Copilot, now used by over 2,0001787developers within OIT and our contract partners. These software1788developers indicate this AI-assisted software development tool is1789helping them deliver capabilities faster and saving them over 8 hours a1790week. This includes faster development of Veteran-facing features on1791VA.gov, making it easier to refill prescriptions and apply for1792benefits, and the improvement of backend systems that accelerate claims1793processing.1794    AI-augmented tools are also driving improvements in clinical care,1795with 82 percent of the over 200 use cases in VA's inventory coming from1796the Veterans Health Administration (VHA). VA's Stratification Tool for1797Opioid Risk Mitigation (STORM) uses machine learning to identify and1798mitigate the risk of overdose and suicide among Veterans prescribed1799opioids or with opioid use disorder. By summarizing patient risk1800factors, STORM identifies high risk Veterans for review by expert1801health teams. Health care teams reviewed the care of over 28,7001802Veterans identified by STORM in the past year alone, decreasing1803mortality in high-risk patients by 22 percent. Since its launch in18042017, the REACH VET program has used tools like STORM to identify and1805bring clinical attention to nearly 135,500 Veterans, improve outpatient1806care, reduce suicide attempts, and decrease the number of mental health1807emergencies.1808    Additionally, VHA has deployed 84 AI-assisted devices that have1809been authorized by the , including one that uses computer vision to1810enhance clinical outcomes such as early tumor detection. One VA study1811showed that using AI-assisted colonoscopy devices increased adenoma1812detection rates by 21 percent, which is associated with lower late-1813stage cancer incidence and reduced mortality.1814    VA is committed to implementing innovative, AI-powered tools that1815advance health care for Veterans, improve the experience of care teams,1816and optimize VA's workforce. As part of this commitment, VA will pilot1817ambient scribe technology at 10 sites beginning this fall. Ambient1818scribe is an AI technology that listens to and documents the1819conversations between health care providers and patients. AI processes1820a transcript of the encounter to generate secondary outputs like1821clinical encounter notes and coding recommendations. It has the1822potential to transform health care by reducing clinician burdens,1823enhancing efficacy, improving patient care quality and experience, and1824engaging with clinical decision support services. Ultimately, it allows1825the provider to spend more time face-to-face with Veterans.1826    As we progress, protecting Veterans' data privacy while responsibly1827leveraging AI's potential is a top priority for the Department. Like1828all software approved for use at VA, AI systems must meet VA's rigorous1829security and privacy standards before they receive an Authority to1830Operate. Additionally, consistent with the Office of Management and1831Budget memorandum M-25-21, our team is facilitating an agency-level1832review of each AI use case to ensure the tool meets the Government's1833standards for innovation, governance, and public trust. Each use case1834undergoes an AI Impact Assessment to identify and mitigate risks.1835    Further, VA has established and is committed to maintaining an1836annual AI use case inventory. First released in December 2024, we are1837on track to provide an update to this inventory in December 2025. This1838inventory positions VA among the most transparent health care systems1839in the country regarding AI.1840    Looking ahead, our focus over the next 12 months will be1841implementing our strategic execution priorities by expanding employee1842access to generative AI to 100 percent of VA staff, reimagining high-1843impact workflows, prioritizing investment strategy to high return-on-1844investment AI solutions, releasing new AI training opportunities for1845employees, and maintaining transparent and effective AI governance by1846ensuring 100 percent of VA's high-impact AI use cases meet the1847Administration's standards.1848    Despite our industry-leading progress, VA acknowledges the adoption1849of AI tools presents significant challenges. Among them is integrating1850new AI solutions within VA's highly complex existing system1851architecture, and as a Government entity entrusted with Veterans'1852private information, balancing adoption of new and emerging tools and1853vendors with the Government's strict security compliance standards is1854crucial. Retention of AI experts is a challenge. Finally, scaling1855commercial AI tools will incur additional costs, making it an ongoing1856effort to align these costs with available technology funding. Cost is1857one of many reasons the Department encourages Congress to fully fund VA1858next year in lieu of another continuing resolution.1859    In conclusion, VA remains steadfast in its commitment to harnessing1860the power of AI to improve the lives of Veterans. By strategically1861investing in AI tools and enhancing our workforce's capabilities, we1862aim to deliver faster, higher quality, and more cost-efficient1863services. While we acknowledge the complexities and challenges inherent1864in this transformation, we are dedicated to maintaining the highest1865standards of governance, transparency, and ethical use of AI. With your1866continued support, we can ensure that VA leads in AI innovation and1867sets a benchmark for responsible AI use in public service. Thank you1868for the opportunity to testify before you today. I look forward to your1869questions.18701871                    Prepared Statement of Sid Ghatak18721873    Chairman Barrett and distinguished Members of the Subcommittee:18741875    Thank you for the opportunity to testify. My name is Sid Ghatak,1876and for almost three decades, I have designed and deployed artificial1877intelligence and forecasting systems across finance, healthcare,1878pharmaceuticals, media, and government.1879    I currently serve as the Chief Technical Advisor for the National1880Artificial Intelligence Association, the premier organization1881representing over 1,500 businesses in the advancement of AI, and am1882also the founder and Chief Executive Officer of Increase Alpha, LLC,1883where we use artificial intelligence to predict stock prices and1884license these predictions to hedge funds.1885    In the Federal Government, I served in the General Services1886Administration for almost 4 years, where I was a Director of the Data &1887Analytics Center of Excellence. In that role, I co-authored the Federal1888AI Maturity Model 3 years before AI took the world by storm and1889contributed to previous Executive Orders on AI, specifically on the1890critical issues of data security and privacy.1891    At Increase Alpha, I architected a predictive AI model that1892generates alpha once thought impossible--a deep learning system1893exceptionally accurate at predicting equity prices. Increase Alpha far1894exceeds multiple industry benchmarks, including accuracy, Sharpe ratio,1895and alpha generation. This solution is not based on Large Language1896Models but is a purpose-built predictive engine designed for a very1897specific need.1898    I want to emphasize that it is entirely unrelated to the Department1899of Veterans Affairs and has no bearing on today's testimony. I mention1900it only as an example of how AI, when carefully designed with a clear1901purpose, can achieve exceptional effectiveness.1902    Taken together, this diverse background--spanning academia,1903government, and industry--has given me the rare opportunity to actually1904build AI systems that work well in the real world. Because I have spent1905my career outside the orthodox worlds of academia, venture capital, and1906big tech, I am also not beholden to herd mentality. Instead, I bring an1907expert, independent perspective which is especially valuable now, when1908much of the world is caught up in the `art of the possible' with AI,1909when what is most urgently needed is a sober understanding of what is1910safe, practical, and ready to serve the public.1911    LLMs like ChatGPT, Claude, and Gemini are a powerful subset of AI,1912but they come with their own set of problems, specifically in1913healthcare, where hallucinations and sycophancy on the part of chatbots1914can lead susceptible users down psychological rabbit holes. Which is1915why it's important to clarify that AI is bigger than just ChatGPT and1916its competitors.1917    To use an analogy: the steam engine transformed society, fueling1918the Industrial Revolution. While steam power still exists today, it1919gave way to other forms of power over time. Until steam engines were1920used to create the first railroads, no human had ever traveled faster1921than a horse. This new form of transportation opened the world's eyes1922to what was possible, just as ChatGPT has shown the world the art of1923the possible with AI. But early train travel was dangerously1924unreliable. Accidents were frequent, derailments common, and thousands1925of lives were lost before rail systems matured into the safe networks1926we know today.1927    The lesson is clear: revolutionary technologies will evolve and1928improve over time when the private sector and government work in1929collaboration. The same applies to AI.1930    As the Committee gathers information on how to modernize technology1931at the VA, I would like to offer three pieces of advice from my decades1932at the front lines of building and implementing advanced analytical1933solutions:19341935        1. Expand the playing field: For the last several years, the1936        world has been consumed with Large Language Models to the point1937        where AI has become synonymous with it; however, that is not1938        the case. Many other types of AI may have similarities to these1939        models, but function very differently. Technologies that1940        specialize in interpreting and understanding images, video, and1941        audio, for example. Or technologies that are better suited to1942        working with numbers and symbols instead of words. And new tech1943        that has yet to be invented.19441945        There is an old adage that when you are a hammer, everything1946        looks like a nail. The world has become so enamored with LLMs,1947        and rightfully so. Interacting with them can feel magical,1948        giving you the sense that they are real people, though they are1949        not. This may be why little to no investment is being made into1950        these other areas.19511952        At Increase Alpha, we have demonstrated clearly what can be1953        done with other forms of Artificial Intelligence. I began1954        building our models at the same time as the research underlying1955        ChatGPT was published. I had also encountered the same compute,1956        cost, energy, and reliance on Nvidia GPUs issues we still see1957        today. I also took a different approach to conserve resources1958        and focus on simplification using Predictive Intelligence,1959        which led to lean AI models that use a minuscule amount of data1960        compared to LLMs, and which are small enough to run on a cell1961        phone.19621963        Over 4 years, the success of my models directly contradicts the1964        notion that massive amounts of data--along with their1965        associated infrastructural and operational costs--are needed to1966        build AI solutions that are extremely accurate, innovative, and1967        reliable. Not to mention that they also consume ever-increasing1968        amounts of energy and utilize models that produce outputs that1969        are often incomprehensible and unexplainable. I have proven, in1970        one of the most competitive and challenging tech arenas, that1971        modern AI does not require all this if it is designed correctly1972        from the outset. The Administration, in its recent AI Action1973        Plan, does not limit AI to the narrow definition of LLM and1974        provides support for numerous types of technologies to be1975        developed.19761977        2. Correlation is not Causation: The difference between1978        correlation and causation is best understood through an1979        example. There is a near-perfect correlation between the number1980        of Google searches for the word `Nintendo' and the number of1981        librarians in Michigan. It doesn't take a rocket scientist to1982        understand that there is no actual relationship between the two1983        trends.19841985        Why is this so important? Because AI solutions today, such as1986        ChatGPT, are based on correlations, even if those correlations1987        are nonsensical. It is why they hallucinate (make up answers1988        based on nothing), and why they have an inherent bias. While1989        they give the impression of understanding and reasoning through1990        their rapid generation of coherent text, they have no idea what1991        the words themselves actually mean. They excel at predicting1992        the next best word based on a vast network of correlations and1993        are even better at providing the user with the answer they want1994        to hear, even if it's not accurate.19951996        To achieve true artificial intelligence, these systems would1997        also have to know why the next word was predicted, which cannot1998        currently be explained. They would need to know the truth1999        behind every output. This is causation. That is how the human2000        mind works. Until AI systems can understand and explain the2001        `why' of their inner workings and outputs, and become reliable2002        sources of truth, they will never be truly intelligent. I2003        remain hopeful that I will experience this in my lifetime, but2004        it has not happened yet, nor is it likely to happen soon.20052006        3. Data, Data, Data: AI is an engine that requires data. But2007        not just any data. Accurate, functional AI systems that produce2008        explainable and auditable outputs require vetted and cleaned2009        data, which we feel 100 percent confident using. By some2010        estimates, the Federal Government has more data than any other2011        organization in the world.20122013        As a former Federal employee, I had the opportunity to work on2014        projects that required this type of clean data to achieve their2015        envisioned solutions. What I saw firsthand was the same thing I2016        had seen in every other large organization. There was always2017        more data than anyone realized. No one really knew where all of2018        it was located or what it meant, and the sheer effort to2019        gather, clean, and organize that data for proper use would have2020        been enormous and cost-prohibitive.20212022        This is one of the key reasons many AI and data analytics2023        projects fail. Unless an organization is willing to make the2024        investments in organizing and cleaning their data, these2025        solutions--to put it bluntly--will be like lipstick on a pig.2026        They will not work over the long run, and even when they do,2027        they will not be reliable because they are not explainable.20282029        We can see this already in current versions of Large Language2030        Models, which are aptly named because they are built on2031        unfathomably large amounts of language data. Some of it is2032        factually correct. Some of it is factually wrong. Some of it2033        has good intentions. Some of it is prejudiced, with built-in2034        hate, discrimination, and the bias of their very human authors.2035        As the old saying goes, garbage in, garbage out.20362037    What does this all mean for the VA and the well-being and care of2038our veterans? I can't claim to know. No one does. But I want to leave2039you with a prediction of my own. I believe we truly are on the verge of2040a revolution on the scale of the Industrial Revolution. So, if I could2041leave you with one idea today, it would be this: AI is actually much2042bigger than today's LLMs. And it is these technologies, many of which2043have yet to be invented, that will enable the VA to execute its mission2044``To fulfill President Lincoln's promise to care for those who have2045served in our Nation's military and for their families, caregivers, and2046survivors.''2047    In light of this, the Committee's work is vitally important to2048ensure that the investments the Federal Government makes into AI2049solutions will actually fulfill its mission. Our Veterans have given2050their bodies, minds, and very lives so that we all can enjoy ours, and2051we owe them more than our thanks and gratitude. We owe them the help2052and services they need when and how they need them.2053    I am privileged to be here today, amongst my esteemed colleagues,2054and I look forward to answering your questions. Thank you for this2055opportunity.20562057                Prepared Statement of Mohammad Ghassemi2058[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]20592060                   Prepared Statement of Carol Harris2061[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]20622063                                 [all]