Proceedings of a Workshop—in Brief

Convened April 6-7, 2026

Artificial Intelligence and the Medical Record in the Context of Social Security Disability Evaluations

Artificial intelligence (AI) is redefining the way health care professionals create, interpret, and evaluate medical records. In clinical practice, AI tools now assist with symptom assessment, imaging analysis, predictive diagnostics, and automated documentation—capabilities that have the potential to improve efficiency, accuracy, and accessibility. These technologies are also beginning to transform how medical evidence is reviewed for Social Security disability evaluations, creating both opportunities and challenges for quality assurance, error reduction, and fraud prevention. As adoption expands, patients benefit from understanding when and how AI contributes to their medical records, while clinicians, informaticians, policy makers, and adjudicators in state Disability Determination Services (DDS) and the Social Security Administration (SSA) need to consider its impact on the integrity of medical documentation and the decisions it informs. To this end, the National Academies of Sciences, Engineering, and Medicine (the National Academies) held a workshop on April 6 and 7, 2026, sponsored by SSA. The workshop convened subject matter experts to explore developments in AI as it pertains to medical care; to look at current and upcoming implementations of AI systems or software that create, modify, edit, or otherwise interact with medical or health records; and to identify how AI might affect the information contained in medical records that SSA receives for the purposes of disability evaluation. This Proceedings of a Workshop—in Brief summarizes the key points made by the workshop speakers and participants during the presentations and discussions and is not intended to provide a comprehensive summary of information shared during the workshop. Speaker presentations and references along with the workshop webcast and agenda have been archived online.¹ The views summarized here reflect the knowledge and opinions of individual workshop participants and should not be construed as consensus or recommendations among workshop participants or the National Academies.

Introduction

Jack Dennerlein, Boston University Sargent College of Health & Rehabilitation Services and chair of the workshop planning committee, welcomed and informed participants that the goal of the workshop was to understand how AI is affecting medical records that SSA receives as part of a disability claim and how SSA could prepare for and adapt to this changing landscape. The workshop was organized into five sessions; this Proceedings of a Workshop—in Brief is structured according to the workshop agenda. In the opening plenary of sponsor remarks, SSA representatives gave an overview of the disability adjudication policy and process. The first panel

session discussed the foundations and current landscape of AI adoption in medical practice, and the second session explored the issues of AI quality and trustworthiness. Session three was a clinical-focused session divided into parts: (a) AI use cases for clinical evaluation and diagnosis; (b) AI benefits and risks in clinical imaging and diagnosis; and (c) clinical corroboration and functional proof for claims outcomes. The workshop concluded with a final session focused on policy, governance, and ethical considerations related to integrating AI in health care, followed by a closing plenary discussion among planning committee members.

Social Security Disability Determinations Process

For adjudicators at SSA to evaluate a disability claim, they need a full understanding of the claimant's medical records and their provenance, said Robert Weathers from SSA. Increasingly, he said, this requires understanding AI systems that contribute to medical documentation, including their accuracy and validity, and knowing how and why such systems include or omit certain information. Vincent Nibali, SSA, gave a brief history of SSA disability programs. Social Security Disability Insurance (SSDI) began in 1956; eligibility for SSDI is based on an individual's work record, and the benefit amount is tied to previous earnings. Supplemental Security Income (SSI), which began in 1972, is a needs-based program in which benefits are based on an individual's assets and income. Nibali emphasized that while the programs differ, the medical and vocational eligibility requirements are the same, and both entail a finding of total disability. Disability has been defined by Congress as "the inability to engage in any substantial gainful activity² (SGA) because of a medically determinable physical or mental impairment(s) that can be expected to result in death, or that has lasted or is expected to last for a continuous period of not less than 12 months."

Nibali outlined the disability determination process.³ Applications are first reviewed by a local field office for non-medical eligibility, then forwarded to a state DDS office, which gathers medical evidence and makes an initial determination. Applicants may appeal through reconsideration and, subsequently, to a federal Administrative Law Judge. For adults, SSA applies a five-step sequential evaluation process to assess whether an individual is engaging in SGA, has a severe impairment that meets duration requirements, meets or equals a listed impairment, can perform past work, or has ability to adjust to other work. Nibali explained that a "medically determinable impairment" is one that results from anatomical, physiological, or psychological abnormalities that can be shown by medically acceptable clinical and laboratory diagnostic techniques. If an applicant is found to have an impairment, the third step determines whether it falls under the SSA Listing of Impairments. If the listing is not met, SSA conducts a medical-vocational evaluation, including a Residual Functional Capacity⁴ (RFC) review, and considers whether the claimant can perform past relevant work or adjust to other work. Overall, a finding of disability reflects a "really high standard," said Nibali.

SSA relies on a variety of evidence, including objective medical evidence, medical opinions, longitudinal evidence, claimant and third-party reports, and other third-party evidence to establish impairment, duration, and consistency of reported symptoms. Medical records are submitted to SSA through electronic exchanges, portals, or paper methods. As AI becomes more integrated into medical documentation, Nibali emphasized the importance of understanding data provenance. Key considerations include whether medical records received by SSA represent complete histories or AI-generated summaries, and whether analyses reflect AI outputs with or without clinical judgment.

Adjudicator Experience with AI

As the division administrator for Iowa DDS, Ann Davis, SSA, described how technology has changed the disability determination process at DDS. In 2006, she said, SSA converted to a paperless electronic disability folder, which reduced the time required to receive and process information, but also increased the amount of evidence that DDS receives for each claim. In Iowa, examiners have about 1,600 pages of evidence to review per claim. With SSA processing two

to three million initial claims each year, she estimated that at least three billion pages of evidence are analyzed annually nationwide, not including reconsideration claims and continuing disability reviews. Much of this evidence is received through the Health Information Technology (HIT) system, which allows for the secure exchange of health information in a standardized electronic format. While HIT is not inherently AI, Davis said it may incorporate AI technology to analyze data and search for medical codes within the record that match the SSA Listing of Impairments. The resulting "HIT extract" is accompanied by a disclaimer, which states that the extract should not be referenced or construed as medical evidence, or used as a medical opinion in the disability determination process. Beyond the HIT extract, however, DDS has no way of knowing if records have been generated or manipulated by AI, said Davis. She cautioned that AI may hallucinate or generate inaccurate findings, and adjudicators are trained to identify inconsistencies in the record, treating AI-generated inaccuracies as they would any other inconsistency. The adjudicator would weigh the inconsistency against the totality of the evidence. If AI generates an inaccurate record, the adjudicator would flag the discrepancy just as they would with any other type of record. However, if AI simply omits relevant information, the exclusion may not necessarily be flagged as inconsistent with the record, Davis noted.

Davis also described SSA's use of AI tools to help adjudicators navigate large records more efficiently. One such tool is Intelligent Medical Language Analysis Generation (IMAGEN), which transforms evidence into machine-readable text to allow for searching and analysis of records. IMAGEN leverages natural language processing, data modeling, and predictive analytics, and has a robust clinical vocabulary tailored specifically to adjudication needs. She said during training of new disability examiners, a case with about 1,000 pages of evidence took trainees using prior tools about an hour to find the labs and information needed to meet a listing. Using IMAGEN, Davis found the same diagnosis and supporting information in about five minutes. In another case involving gastrointestinal bleeds, she used IMAGEN to review about 1,600 pages of evidence spanning three years, locate the relevant hospitalizations across sources, organize them chronologically, and respond to the case question in under an hour. Davis encourages DDS staff to use the tools that support the quickest, most accurate decisions for claimants, emphasizing that their mission is "best decision, every claimant, every time." However, she stressed that adjudicators are the "pilots" and tools like IMAGEN are "co-pilots;" IMAGEN may assist with navigation, but adjudicators are flying the plane and ultimately responsible for determinations.

Landscape of AI Adoption in Medical Settings

Philip R. O. Payne, BJC Health System and Washington University School of Medicine, opened the session framing AI in health care as a sociotechnical system. He advised that strong policy approaches would reflect the principle that human-AI collaboration is more effective than either alone. Payne underscored that successful adoption depends on alignment across people, processes, technology, and environment, with workforce readiness as a key policy consideration to ensure clinicians are equipped to use AI appropriately and selectively. Although AI can help humans make decisions by identifying patterns quickly and efficiently, he said there are unanswered questions about issues such as the potential for bias, availability of data, and the integration of AI into a dynamic environment. A big challenge, Payne noted, is educating clinicians to be ready for AI-enabled practice and integrating AI into existing systems. AI can be designed and woven into clinicians' workflow while continuously evaluating its performance to identify problematic consequences such as cognitive overload, automation bias, or shifts in team collaboration. He emphasized that AI systems need to be compared with real-world alternatives, not perfect alternatives, because real-world comparisons help show where AI can make practical, step-by-step improvements in health care delivery. For example, an AI note-taking program might not be 100 percent accurate, but notes taken by humans also tend to be incomplete and contain errors. In the discussion, Payne encouraged AI adoption across all health care systems and suggested offering certain capabilities to federally qualified health centers at no charge because shared capabilities can support quality, safety, and continuity of care. Payne concluded that health care is at an inflection point, requiring a balanced regulatory approach that enables innovation while maintaining appropriate safeguards.

AI in the EHR

Jackie Gerhart from Epic spoke about the evolution of AI in medicine, particularly in the electronic health record (EHR).

Gerhart described the shift in medical AI from rule-based tools to predictive analytics, which estimate predefined events such as emergency department (ED) visits. Newer technologies incorporate generative AI based on large language models (LLMs), which are trained to generate novel content based on prompts or other inputs. Gerhart stressed that all types of AI should be considered when evaluating how EHRs are affected, noting that 85 percent of Epic customers use generative AI. Gerhart provided four suggestions for responsible use of AI: promoting AI literacy within the workplace, deploying careful use-case selection, testing locally for errors and bias, and developing ongoing guidance as regulations catch up with the speed of technology. Gerhart said that the speed, organization, and interoperability of data are central to improving disability evaluation processes. One method that puts this into practice is use of the Trusted Exchange Framework and Common Agreement⁵ (TEFCA), which aims to move away from bespoke data exchange toward a trusted, government-backed exchange framework. Gerhart noted that Epic began its first SSA-related TEFCA data exchange in early April and described it as an important step for integrating information into systems that can be used for disability evaluations.

Clinical Adoption and Pediatric Considerations

Margaret Lozovatsky, Premier Health, emphasized that AI is being adopted faster than any prior clinical technology. Although administrative applications remain the most common use, clinical uses of AI are growing. In survey data, Lozovatsky stated that AI use among physicians has risen sharply, from 36 percent two years ago to 68 percent last year and 81 percent this year (AMA, 2026). She stressed that clinicians are more likely to integrate AI if they trust the tools and believe outputs support appropriate care. Trust concerns include the "black box" nature of algorithms, unclear inputs, imperfect data, and the need for checks, governance, and monitoring. Lozovatsky also emphasized that predictive models should be tested in the specific settings and populations in which they are intended for use, as model performance may vary substantially across clinical contexts and patient groups (e.g., pediatrics). In the moderated discussion, Lozovatsky encouraged the field to move from process measures toward outcome measures, focusing on how AI affects patient outcomes individually and at the population level, and whether it improves current clinical practice.

Rural, Safety-Net, and Underserved Settings

Jennifer Stoll, OCHIN, emphasized the importance of democratizing health IT (information technology) so that clinicians in smaller practices can access the same AI tools as larger systems. She noted that OCHIN supports 8.4 million patients through federally qualified health centers, critical access hospitals, public health departments, tribal nations, and other health care providers, and works to expand access to technologies available in larger health systems. OCHIN has long maintained a direct connection with SSA through the eHealth Exchange for real-time benefit verification and is now helping members adopt AI responsibly, affordably, and when it addresses an identified problem. Stoll highlighted that rural and low-resource care providers face infrastructure and workforce barriers, including multiple non-interoperable EHRs, limited broadband, and minimal IT support. Although trust, governance, liability guidance, and implementation support are important across health care settings, smaller and under-resourced providers may have less internal capacity to establish these structures independently. Stoll stressed that safety-net and rural health care providers "can't do it alone" and may rely on trusted partners, clear data use agreements, transparency about how patient data are used, and assurance that patient consent is respected. In the panel discussion, Stoll added that many rural and underserved care providers continue to use underperforming EHRs and face substantial interoperability gaps, cautioning that "if you're starting from a broken place, it doesn't matter what AI model that you add on top of it, it's still broken from the beginning."

AI Quality and Trustworthiness

John Lee, HIT Peak Advisors, opened the second session with an overview of the promise and perils of AI in clinical practice. He described AI as a tool that supports clinicians' thinking and workflows rather than simply generating finished outputs. Lee urged participants to look beyond accuracy when evaluating AI, including whether the tool solves a real problem or is being used in the wrong context, why failures occur, the risks of automation bias and over-trusting outputs, and the importance of keeping a reliable human in the loop for accountability. These considerations are particularly critical when AI could potentially omit key information and produce an incomplete record. Ultimately, he said, errors made in the beginning of an assessment have the potential to

spread downstream and impact SSA disability evaluations. He emphasized that one AI error in a clinical note is not just one error; the note becomes part of the patient's record, and the next AI summary treats the error as confirmed clinical history. The error continues into every system downstream, and the patient ends up with a "laundered history" that does not accurately reflect their medical history.

Lee stated that although several safeguards exist to improve accuracy and catch errors, such as pre-deployment validation and post-deployment monitoring, they often rely on limited vendor-selected data and adding dashboards to the hundreds of other dashboards that "no one has time to review." Lee noted the real opportunity for AI is in the "unglamorous middle" zone between full automation and manual processes, which leverages the strengths of AI and facilitates the expertise of medical professionals. Lee concluded that, for SSA determinations, the real issue is not whether AI was used, but whether a process existed to identify when it was failing.

AI Performance and Impact Monitoring

AI can make mistakes, said Jinoos Yazdany, University of California San Francisco (UCSF). For example, Yazdany advised a patient with lupus to stop using a drug due to serious side effects. However, the post-visit note incorrectly instructed the patient to continue the drug. Had the patient followed the note, she said, consequences could have been severe. Yazdany discussed her work as the executive director of IMPACC (Impact Monitoring Platform for AI in Clinical Care), which provides real-time insight into the performance of AI tools and clinical impact. Adoption and retention of AI scribes at UCSF is high, she stated, with clinicians reporting reduced time spent taking notes and less cognitive burden and burnout. Clinicians have also increased their patient volume, she said, creating a financial incentive for AI adoption (Holmgren et al., 2026). Although concerns have been raised about inequitable use, Yazdany reported no observed variation by patient characteristics to date, while noting the need for further evaluation across populations. However, several risks have emerged; for example, AI-generated notes are often longer and may include clinically irrelevant information. Yazdany emphasized that human editing remains the primary safeguard, yet IMPACC has found that editing declines over time, potentially reflecting automation bias. Common errors they identified include incorrect dosages or instructions, incomplete documentation of clinical discussions, misattribution of statements, and fabricated or missing details. Evidence also suggests potential downstream effects of scribes: a study found increased documentation of neuropsychiatric symptoms, yet they were less likely to be associated with psychiatric interventions (Castro et al., 2026), which Yazdany described as "very concerning" and indicative of "automation complacency." Although AI-generated notes appear coherent, they may lack clinical nuance or "subtle threads of reasoning," such as why a specific test was chosen or how a diagnosis was prioritized (Abernethy et al., 2026). While AI scribes offer clear efficiency gains, Yazdany emphasized the importance of monitoring and mitigation systems, cautioning that "faster documentation is not the same as better thinking."

Protecting Patients, Organizations, and Ecosystems

Jodyn Platt, University of Michigan, explained that the risks of AI exist across organizations and ecosystems and said that her work uses a public health lens to understand the patient experience within these systems. She explained that AI-related changes can affect connected systems, and fixing parts in a vacuum may create imbalanced solutions. Platt noted that there is also a mismatch between where AI guardrails are placed and where patient concerns are concentrated. For example, some patients report greater discomfort with AI use for administrative purposes due to concerns about discrimination and access to care, but there are few policies in place that provide greater security to patients. One of the most important things to patients is knowing when AI is used, said Platt, noting that a vast majority want to be notified about the use of AI, as well as its impact on privacy, bias, and safety (Platt et al., 2024). When integrating AI into health care, Platt underscored the importance of adequate human oversight of technology to monitor consequences. People are ultimately responsible for the quality of health care, and it is essential that those who are using AI have the expertise and competency to use it appropriately.

Evolution of AI in Medical Records

Medical recordkeeping has changed significantly over the years, said Karandeep Singh, University of California San Diego Health. Singh described the evolution of clinical documentation from handwritten records to dictated notes, to templated electronic health record notes, and now to AI-assisted documentation. Handwritten notes were brief and

difficult to extract from; dictated notes captured more nuance; and templated electronic notes improved efficiency but often reduced narrative detail and created challenges for extraction systems. Singh added that AI-assisted documentation may reverse some of the limitations of templated notes by producing natural language text that captures more of the clinical encounter and is easier for modern extraction tools to process. AI now extends across the documentation life cycle, from pre-charting and note generation to post-processing, clinical documentation improvement, coding and billing

support, guideline review, and direct patient interactions. This adoption is occurring organically, he noted, often because clinicians find the tools useful for reducing cognitive burden or saving time. Some AI use, however, occurs "below the waterline," outside formal health system visibility or approval (Ötles et al., 2026). Thus, he said health systems need to understand which tools clinicians are authorized to use and which tools they actually use, so that appropriate information security, compliance, and contracting safeguards can be applied. Singh also described advances in medical record abstraction. Earlier extraction methods relied on predefined words, patterns, or rules, limiting their ability to capture nuance. LLMs have made it easier to extract information from natural language clinical text using few or no task-specific examples, reducing reliance on highly specialized models for each downstream application (Agrawal et al., 2022). Singh raised the following considerations for downstream users of medical records:

• AI may shape medical records through pre-visit charting, visit documentation, post-visit processing, coding and billing support, guideline review, and direct patient interactions.
• AI-assisted notes may be richer and more extraction friendly than some templated electronic health record notes, but they raise questions about transparency, governance, and downstream use.
• Organic clinician adoption increases the importance of health system visibility into use of both approved AI tools and "shadow AI."
• Patient-AI conversations may create a new documentation trail that becomes part of the medical record.

Singh emphasized that AI tools should not be implemented without governance, particularly for direct-to-patient agentic AI, referring to systems that can take autonomous or semi-autonomous actions to complete tasks, such as interacting with patients, navigating workflows, or prompting next steps. It is essential that these systems undergo rigorous testing and include clear escalation pathways that incorporate human oversight, he said, especially when an AI-driven interaction deviates from its intended scope or enters areas beyond its appropriate domain. In closing, Singh underscored the importance of thorough pre-implementation testing, as well as real-time monitoring to flag patient interactions that may require clarification or correction.

Clinical Use Cases: Testing, Imaging, and Diagnosis

AI for Clinical Evaluation and Diagnosis

New Challenges for Cost, Evidence, and Implementation

Kensaku Kawamoto, University of Utah, shared three characteristics of AI that are relevant when considering the use of AI in medical records. First, AI is evolving rapidly; "what was true recently may be obsolete now," he said. As a result of this rapid pace, the literature on AI is often "way behind." He identified computing cost as a second major consideration, particularly for health care use cases that process large volumes of data, such as years of clinical notes, to generate required documentation. Current cost-control strategies include using smaller models, limiting the amount of data processed, reducing the extent of double- or triple-checking, running inferences during off-peak times with spare or lower-cost compute time, and using local rather than cloud-based resources. Kawamoto noted that these costs could change quickly as next-generation AI servers become available and may be able to perform the same work at a fraction of the cost. The third relevant characteristic of AI, said Kawamoto, is the unprecedented speed to solving problems. Problems that used to take large teams years to tackle can now be addressed by small teams in months or even days. As an example, he described an AI intake tool developed by his team that creates a patient summary, care plan, and medical record note. According to Kawamoto, the team discussed the concept in the morning and had it ready for implementation less than 12 hours later.

Risk-Based Approaches to AI Evaluation and Monitoring

AI in health care generally falls into one of four categories, said Vincent Liu, Kaiser Permanente: clinical, direct-to-consumer, business operations, and hybrid. Clinical AI tools support diagnostic or treatment decisions, such as EHR-based sepsis alerts. Direct-to-consumer tools are used by individuals for health or wellness concerns, such as a smartphone app that evaluates skin conditions. Business operations tools support care delivery functions, such as scheduling, while hybrid tools serve multiple purposes, such as LLMs used to draft replies to patient messages.

Liu described the process of evaluating and monitoring AI in health systems. Before implementing a new AI tool, he said, there are several steps of evaluation. The first step is assessing whether AI is the right solution for the problem. The second step involves testing the tool in a controlled environment using retrospective data and performance metrics, such as sensitivity and positive predictive value.⁶ Third, the tool is placed in "silent mode," Liu said, meaning the tool is live in the system but not exposed or used, allowing real-time identification of data quality issues and operational gaps. Finally, the system evaluates how the tool fits into workflow and how the human-computer interface can be designed to maximize value. After implementation, monitoring continues through assessment of user and workflow adoption, outcomes, and algorithm performance.

Liu said generative AI requires additional attention because it is probabilistic, changes rapidly, depends on prompts and context, involves many potential edge cases, and often lacks a clear or objective "ground truth." These complexities support a multi-layered approach to monitoring and evaluation, he said. Such an approach could include end-user training that supports clinician autonomy and agency, in-line feedback mechanisms and channels, intermittent independent audits of sampled outputs, "human in the loop" oversight at key decision points, and continuous monitoring for safety, quality, and equity including when LLMs are used. Liu emphasized that monitoring and evaluation work best when they are pragmatic, minimally burdensome, and scaled to the level of risk. Lower-risk uses may need only a label indicating that AI assisted with documentation. Moderate-risk uses may require more detailed information about the tool and data provenance. Higher-risk uses need to include richer information about provenance and require human review and editing of outputs as appropriate, he said.

Opportunities and Ethical Challenges for Mental Health Disability Evaluation

Rebecca Brendel of the Center for Bioethics at Harvard Medical School discussed how AI could help address longstanding disparities in disability evaluations for mental and physical disorders. She noted that mental disorders, including intellectual disability, have historically accounted for a substantial share of disabled worker beneficiaries. More recently, mental disorders represented 13.5 percent of new disability awards in 2024, making mental health an important use case for considering AI in disability evaluation (SSA, 2025). Brendel also noted that claims involving mood disorders have a higher initial denial rate than disability claims overall. She cautioned that some denials may be appropriate, while others may reflect insufficient information or difficulties translating subjective symptoms and functional limitations into evidence that can support a determination.

Brendel emphasized that functional status is central to disability evaluation, yet clinicians may lack the information or time needed to fully document it. She said AI could help integrate data vertically and horizontally, both within systems and across systems that have traditionally been separated to protect patients. Brendel noted that mental health data have often been treated as requiring special protection; while those concerns are important, she said this "exceptionalism" can also make it harder to gather and integrate needed information. Brendel described helping with a disability recertification for a patient with severe, lifelong mental illness, during which boxes of paper filled her office. The amount of data from medical records can be overwhelming, she added, and AI can help clinicians sift through large records, identify relevant information, and summarize well-established details, allowing clinicians to focus on issues contributing to claim denials. In light of the psychiatrist workforce shortage, Brendel said AI can help clinicians work more efficiently while still meeting patient needs.

Brendel cautioned that new tools will call for regulatory oversight and the establishment of patient trust. AI presents huge opportunities to improve access and reduce disparities, she added, but it is essential to remain focused on patient voices, ethical considerations, trust, and processes that allow the work to occur effectively.

Patient and Care Partner Access to AI-Generated Health Information

Grace Cordovano, Enlightening Results, shared her perspective as a patient advocate and primary caregiver to two disabled family members. Reflecting on her experience assisting with their disability applications when she was 20 or 21 years old, Cordovano described the process as "really hard" because records were fragmented and often paper-based, both applications faced repeated denials and months of appeals,

and she was navigating the system as a young adult for whom English was a second language. She noted that similar responsibilities often fall to teenagers in immigrant families because they may be the family members most proficient in English and most comfortable using digital systems.

Cordovano contrasted the increasing use of AI in clinical settings with the lack of patient access to these tools and their outputs. She observed that patients are often unable to leverage AI-generated information to inform their own care. For example, mortality prediction scores are used by some health care facilities to guide clinical decision making, yet these outputs are generally not shared with patients, despite patients' right to access information used in their care. Cordovano suggested that TEFCA could facilitate access to such outputs and potentially accelerate disability determinations. She argued that when a person is clearly terminally ill, the process should not require a lengthy cycle of applications.

Cordovano urged stakeholders to more fully incorporate patients and care partners into policy and governance discussions as subject matter experts rather than "token patients." Cordovano asserted that long-term caregivers understand "every nuance to the health system" and "every place that it failed their loved one," yet lack formal avenues to contribute this knowledge. While access to benefits is important, she emphasized that many individuals also seek opportunities to contribute to society and would benefit from tools that connect them to employment and volunteer roles. "Where is the tool for that?" she asked, underscoring the need for solutions that support all applicants, regardless of benefit eligibility.

AI for Clinical Imaging and Diagnosis

AI Tools for Radiology: Current Uses and Practical Constraints

The demand for imaging services has grown exponentially during the last 15 years, said Adam Flanders, Thomas Jefferson University, but the number of radiologists has only increased by 13 percent (Bruls and Kwee, 2020; Malhotra et al., 2026). Health systems are looking to AI as a solution for this mismatch, and there are many tools available; out of the 1,451 AI devices in the marketplace cleared by the Food and Drug Administration (FDA), 76 percent are related to radiology (FDA, 2026). However, these tools have limitations, said Flanders. Narrowly tailored tools can carry high operational costs, are generally not billable, are often poorly integrated into the radiologist's workflow, and may be affected by performance drift that is difficult to measure or address. He also noted that many commercial tools do not focus on some of radiology's most time-consuming tasks, such as tracking whether tumors are growing, shrinking, multiplying, or reducing. While small pilot studies have found efficiency gains, Flanders said the broader literature does not show overall performance improvement. As he put it, "Adding more screens to the radiologist's cockpit does not necessarily improve what they do every day." Flanders said automated report drafting is becoming common in radiology, but he emphasized that AI-generated outputs still require human review. He gave an example of an AI tool that hallucinated a second brain tumor on a scan. He also noted that AI could help extract quantitative information from images that may be relevant to disability determinations, such as lean muscle mass and bone density scores.

AI provides a set of tools that may improve efficiency and accuracy in health care, said Flanders, but autonomous use without supervision should be approached with caution. Guardrails, including benchmark datasets, performance registries, quality thresholds and continuous monitoring are important to ensure patient safety.

Understanding AI-Generated Imaging Findings, Errors, and Accountability

AI can create a richer medical record, said Peter McCaffrey, University of Texas Medical Branch, but decision makers need to understand how the information was generated. He described a "disclosure paradox" in which awareness that AI was used may make people more skeptical of otherwise identical medical information, which may be perceived as less reliable when labeled as AI-generated or AI-assisted (Reis et al., 2024). McCaffrey said this skepticism could lead clinicians to qualify AI-generated imaging findings as less certain than other information, even when the AI tool may be more accurate. McCaffrey said this can affect downstream clinical management, especially because it may not always be clear which tasks an AI tool can perform reliably. Research has found that AI may substantially improve performance on some tasks while reducing accuracy on others that appear comparably difficult, creating what researchers have called a "jagged technological frontier" (Dell'Acqua et al., 2023). He added that models generally do not communicate their level of

certainty; as a result, clinicians may risk either under valuing information that is more certain or over valuing information that is less certain.

McCaffrey described several ways that AI can produce errors in the medical record, including performance degradation over time, data shift model drift, data anomalies, and hard failures. These errors can distort the medical record and affect patient management. Missed findings may cause serious patient harm, he added, while other errors may create additional review work (Newman-Toker et al., 2024).

McCaffrey noted that AI tools may populate or shape diagnostic records before a physician fully reviews the case, making records appear complete while still containing uncorrected errors or omissions. For disability determinations, he said, AI may miss context that matters for assessing functional capacity, and standard performance metrics may not capture the clinical consequences of different types of errors.

McCaffrey said these concerns raise unresolved questions about accountability and standards for use, including how users can identify tasks for which AI assistance is reliable and when additional review may be warranted (Dell'Acqua et al., 2023). For example, if AI contributes to a misdiagnosis in the medical record, responsibility may be unclear; if a physician overrides AI and makes an error, liability may also be uncertain. He also raised questions about how model calibration problems could affect an adjudicator's assessment of disability risk and how standardized benchmarks might define what counts as "good enough" when comparing AI accuracy with that of human specialists. McCaffrey concluded that these issues should be addressed before AI is implemented across the health care system.

Bias and Disparities in Imaging AI

Judy Wawira Gichoya, Emory University School of Medicine, said that because radiology has the largest number of FDA-cleared AI-enabled medical devices, it can serve as a test case for bias patterns likely to appear in other AI tools. She emphasized that AI bias in imaging is real, measurable, and almost always invisible in the record that SSA receives. She shared three examples of AI bias. First, a study of AI-assisted chest X-rays found systematic underdiagnosis of Black, Hispanic, female, and Medicaid patients (Seyyed-Kalantari et al., 2021). Second, a dermatology AI model had significantly lower accuracy for dark-skinned patients (Daneshjou et al., 2022). Third, cardiac magnetic resonance imaging (MRI) segmentation showed a 10 percent relative performance drop for minority racial groups compared with White patients (Puyol-Anton et al., 2022). Gichoya noted that these biases can compound existing disparities, particularly for Black patients who are already at risk of underdiagnosis and poorer care in multiple areas of medicine.

Gichoya said bias begins before the development of AI models because access to advanced imaging is unequal for Medicaid and uninsured patients, non-White patients, rural residents, and non-English speaking individuals. She noted that these disparities predate AI and reflect differences in the availability and distribution of imaging modalities, with some rural areas lacking access to advanced imaging. Therefore, missing or sparse imaging may reflect limited access rather than the absence of disease. Biases can then become "locked into the file" through shortcut learning, automation bias, and deskilling.

Gichoya shared some considerations for SSA disability adjudicators who often receive imaging report text only, without the image or information about how the report was generated. Gichoya listed several "skepticism signals" and how to respond to reduce bias in disability determinations:

• Signal: Sparse imaging for a severe presentation. Response: Trigger a consultative examination instead of inferring absence of disease, recognizing that sparse imaging may reflect limited access to appropriate imaging modalities.
• Signal: Missing modality for the impairment. Response: Lack of images may be seen as a record gap, not a negative finding.
• Signal: Short, templated, single-reader reports. Response: Without a peer-review trail, treat as lower-weight evidence.
• Signal: Discordance between patient-reported function and imaging narrative. Response: Seek corroborating functional evidence rather than relying on the imaging narrative alone, because the narrative may reflect nonclinical shortcut features rather than the patient's actual disease state.
• Signal: Stigmatizing or dismissive language in notes. Response: Flag for review under consistency.

• Signal: Site-of-care and equipment markers known to correlate with shortcut learning. Response: If the only imaging comes from a facility with templated AI-assisted reads, weight accordingly.

Since bias in imaging AI is architectural, she said, choosing a "better model" is not enough; SSA could conduct routine outcome audits by claimant demographics and provide an appeal pathway for AI-influenced findings before deployment scales further.

The Impact of Ambient and Agentic AI on Medical Documentation

A "documentation revolution" has occurred in recent years, said David Dorr, Oregon Health & Science University. AI scribes are now used in many major health systems, and their outputs increasingly shape what appears in the medical record. Dorr added that patients are engaging with LLMs outside the health care system to interpret symptoms before coming to the clinic, sometimes with "scripts" to request specific tests or referrals. Concurrently, agentic AI is being used to guide clinicians through complex workflows and to "nudge" care teams for completeness. These changes, said Dorr, are turning medical records into "curated narratives." While many of these innovations are positive steps, he warned of new risks, including hallucinations (invented symptoms), cloning (rapid propagation of errors in the record), and the "veneer of accuracy" (professional looking notes that mask inaccuracies). These concerns are consistent with emerging evaluations of ambient digital scribe platforms, which have examined the quality and safety of AI-generated clinical documentation (Anderson et al., 2025). Dorr said that if clinicians do not review and edit AI-generated notes before they enter the medical record, they may mis-document care and codify errors. To address these risks, Dorr offered three safeguards; first, tag AI-generated or synthesized content to indicate provenance; second, distinguish between AI used for quantification and AI used for narrative synthesis; and third, include a "human anchor" to maintain clinical corroboration and to ensure that the longitudinal functional reality matches the note. As Dorr stated, "The more we can be transparent about how and why AI was used, the better off we'll be."

Regulation of AI-Enabled Radiological Devices by the U.S. Food & Drug Administration

Robert Ochs, FDA, described how FDA regulates radiological devices, including devices with AI-enabled device software functions. He noted that not all AI tools are FDA regulated, as some do not meet the definition of a medical device. Low-risk medical devices are generally exempt from FDA premarket review. Moderate- to high-risk medical devices undergo premarket review for safety and effectiveness and some may also be subject to post-market studies. About 1,400 AI-enabled medical devices have been reviewed and cleared by the FDA, said Ochs.

Ochs said the regulatory process begins with the device's intended use (i.e., its design purpose for a disease/patient population) such as detecting lung nodules larger than 5 mm or estimating the likelihood of a lung nodule being cancerous. If subject to premarket review, FDA evaluates design documentation and performance testing to determine whether the data supports the safety and effectiveness of the device for its intended use. FDA also reviews risk mitigations; Ochs noted that not all risks can be eliminated, and developers are expected to provide users with accurate information regarding the device's performance and known limitations. Ochs stated that the FDA allows manufacturers to pre-specify future device changes, along with plans for how those changes will be evaluated and what criteria must be met before implementation. Post-market monitoring is also required for some devices, and information on authorized devices is available through public FDA databases, said Ochs.

Clinical Corroboration and Functional Proof: AI and the Medical Record for Disability Claims Outcomes

AI at the Veterans Health Administration

The medical record isn't simply a list of symptoms or a collection of billing codes, said Haipeng (Mark) Zhang, Veterans Health Administration (VHA); it is a narrative of someone's life. A clinical note written in an exam room today may become key evidence in a disability evaluation years later. As AI becomes integrated into health care, he said, a primary challenge is ensuring that technology preserves, rather than distorts, the integrity of the clinical narrative. The footprint of AI is growing at the VHA, said Zhang; noting that ambient AI scribes are being integrated across the clinical workforce. To support ethical and transparent use, the VHA has established several guardrails. For example, Zhang noted that clinicians must obtain explicit verbal consent from patients or families, and both consent and AI use must be documented in the note. AI-generated content is treated strictly as a draft, and clinicians are required to review and approve all entries before

they become part of the official record. These practices are consistent with the VA's broader framework for AI that is effective and safe, transparent and explainable, accountable, and monitored (VA, 2023). Zhang emphasized that if an AI scribe misses a nuanced symptom or over-summarizes a functional limitation, this error not only impacts clinical care, but it can potentially delay a veteran's needed benefits. Beyond scribes, Zhang said that the VHA is exploring more than 250 other AI applications and has seen massive adoption of LLM tools such as VAGPT for administrative and analytical tasks (VA, 2026).

Zhang shared several lessons relevant to SSA's disability determination process. AI models are stochastic by design; that is, they are governed by probability and patterns rather than rigid deterministic rules. As a result, he said, they require a robust oversight framework that includes workflow redesign, rigorous quality assurance, in silico testing against gold-standard datasets, silent deployments to assess performance, and live trials with predefined measures of success. These elements align with broader AI risk-management frameworks that emphasize governing, mapping, measuring, and managing risks throughout the AI life cycle (NIST, 2023). Zhang also stressed the importance of cross-agency standards for identifying AI-generated text and associated metadata in the record. The medical record is no longer a static document, he said. Instead, it is a living AI-augmented dataset, and despite technological advances, it is critical that the record reflects the patient's lived experience with fidelity. The technology behind a patient's medical record serve as a "bridge to benefits," not a barrier, he concluded.

AI in Rehabilitation Medicine

Robert Latz, Trinity Rehab Services, stated that although AI scribes are increasingly widespread in health care, they are only beginning to be used in physical therapy, occupational therapy, speech therapy, social work, dietary services, vocational rehabilitation, and other related settings. Currently, conversational AI tools are being used to conduct pre-therapy interviews and help identify diagnostic areas for clinicians to examine more closely. Ambient scribes, he noted, are being used in some rehabilitation settings but generally capture transcribed notes rather than automatically populating forms or documents. Other tools include musculoskeletal ultrasound, motion analysis through phone cameras, balance assessments, telehealth-based movement tracking, remote therapeutic monitoring, phone apps that track gait and exercise performance, wearables, prosthetics, exoskeletons, agentic AI for outcome tracking, and infrared tracking for fall protection. Latz emphasized that these technologies are not evenly distributed as adoption varies by region, institution, resources, training, digital health divides, and local governance capacity. He also highlighted the importance of the International Classification of Functioning, Disability and Health (ICF), noting that rehabilitation professions are often built around function and disability but that documentation systems do not always capture or exchange functional information in ways that support downstream uses such as disability adjudication.

Amanda Brewer, Brewer Physical Therapy, added that AI-assisted coding helps therapists identify appropriate Current Procedural Terminology (CPT) codes for billing based on clinical documentation. She noted that case-trend alerts can show changes over time in measures such as range of motion, strength, function, goals, pain, and patient outcomes. Brewer shared two examples from her work illustrating how AI could impact the disability determination process. In one case involving a patient who had been denied benefits, she said AI helped organize and navigate thousands of pages of records documenting surgeries, procedures, therapy visits, imaging, and specialist visits. This allowed functional evidence associated with the patient's impairments to be identified for a reconsideration hearing. In another case, Brewer said an ambient scribe captured a patient's statement that she was a full-time caregiver for her grandchildren and took them to sports activities. The AI scribe documented these details verbatim in the clinical note, and the patient's attorney later petitioned unsuccessfully to have the information removed because it undermined her disability claim. This example demonstrates how AI scribes may surface more or different information in the medical record compared to traditional manual documentation. Research has similarly found that ambient-scribe use can produce longer notes while reducing clinicians' manual input, potentially increasing the amount of information documented during an encounter (Duggan et al., 2025). She said activities of daily living, which are essential for SSA to determine RFC, are now being captured by AI scribes. At the same time, she cautioned that AI-generated documentation can look "very polished" and still be "100 percent wrong." Brewer also described a matter reviewed by a Louisiana professional licensing board in which

an ambient scribe was used without the patient's knowledge, raising concerns about disclosure, consent, and professional conduct. She emphasized that users need to understand the risks of applying AI to medical records and disability claims and that its use warrants transparency, clinician review, and ongoing monitoring (APTA, 2025).

Implications for SSA and the Disability Determinations Process

Davis agreed with the importance of a "human anchor" and described SSA's IMAGEN tool as a support tool that still lacks adjudicator oversight. She noted that functional data from sources such as sensors, exercise tracking, and between-visit monitoring could be particularly valuable to SSA. When claimants do not meet a listing, adjudicators have to assess functional capacity, yet such information is often limited or difficult to extract from existing medical records. She reiterated that DDS adjudicators review evidence and speak with claimants, but they do not see claimants directly. As a result, they look for inconsistencies across the evidence. Davis noted that inaccuracies that do not appear as inconsistencies are concerning because adjudicators could rely on evidence that does not accurately reflect a claimant's actual function. Davis agreed that guardrails are needed around when information can or cannot be removed from a record. She added that DDS still requests records directly from care providers, even when appointed representatives submit records, because SSA cannot assume that records submitted by representatives include everything in the care provider's record.

Policy and Governance for AI in Health Care

Framing Policy and Governance

Kenrick Cato, from the University of Pennsylvania and a planning committee member, framed the final session around three core policy and governance questions: how medical care providers view third-party use of AI to interpret clinician-created records, how AI could be used to commit or detect medical or disability fraud, and how the medical community identifies errors or misinterpretations in AI outputs. In his opening remarks, Cato discussed AI's primary role in reducing clinician burden through current tools such as ambient scribes, chart summarization, and inbox management and noted how they are designed to reduce documentation workload. Consequentially, these tools were not created to improve patient health outcomes, and their effectiveness is still uncertain in that regard. Cato cautioned that AI introduces new risks such as hidden signals (i.e. inferring race from notes), uneven tool development across professions, and governance challenges due to limited data and rapidly changing models. These risks highlight the difficulty in consistent and effective regulation, he added.

Current Policy Landscape and Regulatory Gaps

Christina Silcox, Duke-Margolis Institute for Health Policy, described the current regulatory landscape for health AI

as a "crochet quilt" that provides some coverage but has holes. She explained that many AI tools discussed at the workshop, such as ambient scribes, communication tools, and chart summarization, are not medical devices and thus fall outside of FDA oversight, leaving no government entity proactively reviewing them against standards. She discussed how FDA allows manufacturers to pre-specify changes they plan on making to devices in the future through carefully documented predetermined change control plans (PCCPs), and that could eventually allow for more rapid, automated changes for AI tools based on continuous learning. Silcox noted that a rule finalized in 2024 regarding Section 1557 of the Affordable Care Act⁷ emphasized the responsibility of covered entities to address bias and prevent discriminatory outcomes, even if caused by AI tools they did not develop. Silcox noted health systems often lack infrastructure to measure whether discrimination is occurring, and she raised unresolved questions about comparators and whether tools need to be perfect or simply improve on current care. She described the Office of the National Coordinator's (ONC) HTI-1 rule⁸, which introduced transparency and risk assessment requirements for AI integrated into certified EHRs, and noted that a proposed HTI-5 rule may roll back these requirements due to limited use at the point of care, despite their value for governance and oversight. Silcox stated that current federal deregulation efforts are focused on technology developers, which may shift risk to deployers. At the same time, many states are introducing laws around AI development and use. Differences between state laws and the lack of standard consensus definitions around AI technology may cause significant compliance burden for both developers and deployers. Silcox stressed the concept of "fit for purpose," noting that tools optimized for primary clinical use may not generalize to SSA-relevant populations and purposes, and retaining the metadata about AI-generated information is critical for secondary use.

Trustworthiness, Traceability, and Governance Practices

Merage Ghane, Coalition for Health AI (CHAI), focused on trust, trustworthiness, and governance in health systems, with specific relevance to SSA. She emphasized that a key

component of trust is traceability, which enables users to understand how AI contributes to clinical workflows. She noted that medical records were designed primarily for coding and billing, not for SSA disability adjudication, so they may have gaps regardless of AI use. Additionally, AI-generated information alone is not necessarily an indicator of fraud, but knowing when AI was used provides additional context to the record. Ghane emphasized contextual validation, noting that less-resourced organizations may need to evaluate vendors at a network level, and that validation should consider local context, populations, and workflow fit. She advocated for a "human-AI calibration" method rather than simply "human in the loop" to reduce overreliance on AI while maintaining efficiency, without adding unnecessary burden. Recognizing that continuous monitoring can be challenging for less-resourced systems, Ghane suggested a shared responsibility between vendors and health care organizations. She outlined key governance domains for consideration, including institutional communication and education about AI policy; standard operating procedures for data management, procurement, development, deployment, and sunsetting; risk and impact evaluation; and feedback and training mechanisms. While organizations can build on existing frameworks, she emphasized the need for continuous policy updates to ensure patient safety and effective data management as AI use evolves.

Moderated Discussion

The discussion highlighted several key themes, including the role of human oversight, the implications of deregulation for risk, and challenges related to performance changes and vendor communication. When discussing the benefits and consequences of the "human in the loop," Ghane emphasized that humans should be engaged early in AI procurement or development. She cautioned against assumptions that current levels of human expertise or attention will hold as AI tools evolve. Silcox reinforced this concern, noting that improvements in AI performance may inadvertently increase automation bias, leading users to rely more heavily on system outputs. When discussing deregulation and risk, Silcox suggested that loss of external validation could make AI tool purchasing decisions harder for health systems, potentially pushing them toward larger, more established vendors by default. Ghane offered a contrasting perspective, suggesting that risk may not meaningfully slow adoption, as organizers are also motivated by concerns of falling behind their peers. The panel also discussed versioning and vendor communication. Ghane noted that contracts could include provisions for disclosure of meaningful performance changes and ongoing reporting. Cato provided an example in which a vendor updated the underlying transcription model multiple times within a short period without proactively notifying users, underscoring the need for greater transparency.

Lastly, when asked who should be responsible for reassessment, Ghane responded that both vendors and health systems should monitor performance and metrics. She also noted that some vendors provide dashboards to deploying organizations and that startups can sometimes be more motivated to co-develop evaluation processes with purchasers. For smaller organizations that may lack time, money, or resources, Ghane noted that third parties are emerging. For examples, she noted Dandelion Health's partnership with the American Heart Association to test solutions, as well as the National Association for Community Health Centers' interest in a digital formulary initiative for vendor evaluation in community health center settings. Ghane emphasized the value of sharing successes and failures rather than treating them as competitive advantages, noting that partnerships between larger and smaller institutions may help support this work.

Closing Remarks

To close the workshop, the chair invited participants to offer final thoughts and key takeaways. Specifically, Dennerlein asked what lessons are relevant to SSA and their disability determination process, and what gaps remain for future discussion.

Committee members emphasized implementation risks, such as exacerbating health disparities. Melissa Spencer, Social Security Administration (retired), warned that existing inequities, which are already evident in disability applications, could be amplified by AI, and she urged SSA to "go slow to go fast" to ensure equitable outcomes. Cato agreed, cautioning that AI could be the "gasoline on the fire" if access and deployment are uneven. Michael Hasselberg, Nebraska Medicine, added that rural care providers often lack the resources for governance and monitoring, though federal partnerships could help build safe, responsible AI infrastructure.

Peter Embi, Vanderbilt University, stated that current AI applications largely reproduce existing work products (e.g., medical records), even as more fundamental system changes may emerge rapidly and unpredictably. Lozovatsky reinforced broader stakeholder engagement beyond data creators, including patients. V. G. Vinod Vydiswaran, University of Michigan, added that similar challenges extend beyond SSA (e.g., cancer registries) and require cross-cutting infrastructure to help stakeholders navigate AI in the health care system.

Nibali highlighted the balance between opportunity and risk. He said AI may enhance medical records, extract new insights from imaging, and potentially mitigate disparities, but adoption should not "let the perfect be the enemy of the good." While medical records have never been perfect, SSA retains a responsibility to use available tools judiciously, dedicating resources and partnerships to understand AI's risks and benefits. It is critical that SSA support its adjudicators, medical teams, and field staff with an accurate understanding of new and evolving technologies, he added.

References

Abernethy, J., A. Shah, B. Chen, S. Reynolds, S. M. Wright, S. M. and P. O'Rourke. 2026. Integrating AI scribes into medical education: Guardrails for preserving clinical reasoning. Journal of General Internal Medicine 41(9):2598-260

Agrawal, M., S. Hegselmann, H. Lang, Y. Kim, and D. Sontag. 2022. Large language models are few-shot clinical information extractors. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: 1998-022.

AMA (American Medical Association). 2026. Physician Survey on Augmented Intelligence. Center for Digital Health and AI. March 2026. https://www.ama-assn.org/system/files/physician-ai-sentiment-report.pdf (accessed July 9, 2026).

Anderson, T. N., V. Mohan, D. A. Dorr, R. M. Ratwani, J. M. Biro, and J. A. Gold. 2025. Evaluating the quality and safety of ambient digital scribe platforms using simulated ambulatory encounters. Mayo Clinic Proceedings: Digital Health 100292.

APTA (American Physical Therapy Association). 2025. APTA Practice Advisory: Emerging Technology: AI-Enabled Ambient Scribe Technology in Physical Therapy Documentation. https://www.apta.org/your-practice/practice-models-and-settings/practice-advisory-emerging-technology (accessed July 19, 2026).

Bruls, R. J. M., and R. M. Kwee. 2020. Workload for radiologists during on-call hours: Dramatic increase in the past 15 years. Insights Imaging 11(1):121

Castro, V. M., T. H. McCoy, P. Verhaak, A. Ramachandiran, and R.H. Perlis. 2026. Psychiatric documentation and management in primary care with artificial intelligence scribe use. JAMA Psychiatry 83(3):281-6

Code of Federal Regulations. Residual Functional Capacity. 20 C.F.R. §§ 404.1545, 416.945. https://www.ssa.gov/OP_Home/cfr20/404/404-1545.htm.

Code of Federal Regulations. Substantial Gainful Activity. 20 C.F.R. §§ 404.1572, 416.972. https://www.ssa.gov/OP_Home/cfr20/404/404-1572.htm.

Daneshjou, R., K. Vodrahalli, R. A. Novoa, M. Jenkins, W. Liang, V. Rotemberg, J. Ko, S. M. Swetter, E. E. Bailey, O. Gevaert and P. Mukherjee. 2022. Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances 8(32):eabq6147.

Dell'Acqua, F., E. McFowland III, E. Mollick, H. Lifshitz, K. C. Kellogg, S. Rajendran, L. Krayer, F. Candelon, and K. R. Lakhani. 2026. Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science 37(2):403-23.

Duggan, M. J., J. Gervase, J., A. Schoenbaum, W. Hanson, J. T. Howell III, M. Sheinberg, and K. B. Johnson. 2025. Clinician experiences with ambient scribe technology to assist with documentation burden and efficiency. JAMA Network Open 8(2): e2460637.

FDA (Food and Drug Administration). 2026. Artificial intelligence-enabled medical devices. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices (accessed June 8, 2026).

Holmgren, A. J., C. L. Fenton, R. Thombley, H. Soleimani, R. Croci, O. DeMasi, M. E. Byron, S. G. Murray, J. R. Adler-Milstein, and J. Yazdany. 2026. Ambient artificial intelligence scribes and physician financial productivity. JAMA Network Open 9(1):e2553233.

Malhotra, A., K. Kandala, D. Futela, S. Payabvash, D. A. Lakhani, D. Gandhi, C. Whitlow, C. and R. Duszak. 2026. The evolving US radiologist pipeline: Trends in residency positions, resident workforce, and practicing radiologists per unit population. Journal of the American College of Radiology: JACR S1546-1440(26)00207-3. Advance online publication.

Newman-Toker, D. E., N. Nassery, A. C. Schaffer, C. W. Yu-Moe, G. D. Clemens, Z. Wang, Z., Y. Zhu, A. S. Tehrani, M. Fanai, A. Hassoon, and D. Siegal. 2024. Burden of serious harms from diagnostic error in the USA. BMJ Quality & Safety 33(2):109-20.

NIST (National Institute of Standards and Technology). 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. Gaithersburg, MD: National Institute of Standards and Technology.

Ötleş, E., S. G. Murray, A. N. Beecy, A. A. Khalessi, and K. Singh. 2026. Health Systems Govern Only the Tip of the AI Iceberg. NEJM AI 3(4):AIe2600236.

Platt, J., P. Nong, G. Carmona, and S. Kardia. 2024. Public attitudes toward notification of use of artificial intelligence in health care. JAMA Network Open 7(12):e2450102.

Puyol-Antón, E., B. Ruijsink, J. Mariscal Harana, S. K. Piechnik, S. Neubauer, S. E. Petersen, R. Razavi, P. Chowienczyk, and A. P. King. 2022. Fairness in cardiac magnetic resonance imaging: Assessing sex and racial bias in deep learning-based segmentation. Frontiers in Cardiovascular Medicine 9:859310

Reis, M., F. Reis, and W. Kunde. 2024. Influence of believed AI involvement on the perception of digital medical advice. Nature Medicine 30(11):3098-100.

Seyyed-Kalantari, L., H. Zhang, M. B. McDermott, I. Y. Chen, and M. Ghassemi. 2021. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine 27(12):2176-82.

SSA (Social Security Administration). 2025. Annual Statistical Report on the Social Security Disability Insurance Program, 2024. https://www.ssa.gov/policy/docs/statcomps/di_asr/2024/sect03.html. (accessed July 13, 2026).

SSA. 2017. Program Operations Manual System (POMS), DI 24510.001, Residual Functional Capacity (RFC) Assessment — Introduction. https://secure.ssa.gov/poms.nsf/lnx/0424510001.

SSA. 2007. Program Operations Manual System (POMS), DI 10501.001, "Meaning of SGA and Scope of Subchapter." https://secure.ssa.gov/POMS.NSF/LNX/0410501001.

VA (U.S. Department of Veterans Affairs). 2023. VA Trustworthy AI Framework. https://department.va.gov/ai/trustworthy-ai (accessed July 19, 2026).

VA. 2026. VA AI Use Case Inventory. https://department.va.gov/ai/ai-use-case-inventory (accessed July 19, 2026).

Disclaimer: This Proceedings of a Workshop—in Brief was prepared by E. Lin Chow, Rebecca English, Eliana Pierotti, and Erin Hammers Forstag as a factual summary of what occurred at the workshop. The statements made are those of the rapporteurs or individual workshop participants and do not necessarily represent the views of all workshop participants; the planning committee; or the National Academies of Sciences, Engineering, and Medicine.

Planning Committee: Jack Dennerlein (Chair), Boston University; Kenrick Cato, University of Pennsylvania; Peter Embi, Vanderbilt University Medical Center; Michael Hasselberg, Nebraska Medicine; Jesse Jacobs, Liberty Mutual Insurance; Margaret Lozovatsky, Premier Health; Melissa Spencer, Social Security Administration (retired); V.G. Vinod Vydiswaran, University of Michigan. The National Academies' planning committees are solely responsible for organizing the workshop, identifying topics, and choosing speakers. Responsibility for the final content rests entirely with the rapporteurs and the National Academies.

Reviewers: To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by Adam Flanders, Thomas Jefferson University; Thomas Maddox, Washington University School of Medicine; Jodyn Platt, University of Michigan; and Melissa Spencer, Social Security Administration (retired). Katiria Ortiz, National Academies of Sciences, Engineering, and Medicine, served as the review monitor.

Sponsors: This activity was supported by a contract between the National Academy of Sciences and The United States Social Security Administration. Any opinions, findings, conclusions, or recommendations expressed in this publication do not necessarily reflect the views of any organization or agency that provided support for the project.

Staff: Rebecca English, E. Lin Chow, Eliana Pierotti, Joseph Goodman, and Sharyl Nass, Health Care and Public Health Program Area, Center for Health, People, and Places.

Suggested citation: National Academies of Sciences, Engineering, and Medicine. 2026. Artificial Intelligence and the Medical Record in the Context of Social Security Disability Evaluations: Proceedings of a Workshop—in Brief. Washington, DC: National Academies Press. https://doi.org/10.17226/29549.

For additional information regarding the workshop, visit https://www.nationalacademies.org/projects/HMD-HCS-25-13.

Copyright 2026 by the National Academy of Sciences. All rights reserved.