Skip to main content
ATS Scholar logoLink to ATS Scholar
. 2026 Feb 21;7(2):103–106. doi: 10.1093/atsscholar/aapag009

AI won’t save us…yet: reimagining AI’s potential to prioritize people

Lekshmi Santhosh 1,✉, Leah J Witt 2,3
PMCID: PMC13220946  PMID: 41721776

For real transformational change in healthcare and medical education, AI efforts must address systemic flaws, innovate alongside medical educators, enhance workforce sustainability, and improve health equity. These changes must happen in concert with guidance from clinical experts and patients, instead of top-down approaches from corporate healthcare entities, insurance companies, and tech start-ups. We do hope that AI can save our broken and dysfunctional healthcare system, but until these people-focused shortcomings are addressed, AI won’t save us yet.

It’s late when we open our computers to start a “third shift.” It’s after our children have been tucked (and re-tucked) in bed that we begin a familiar after-hours routine to answer lingering patient messages, review endless cc’d charts and close the day’s encounters.1,2 At least we have each other as companions—in between sending parenting memes one of us inevitably texts “AI won’t save us.” Conversations about the transformational potential of artificial intelligence (AI) are inescapable, and in medicine we are already witnessing profound changes brought by AI to the way we work. Briefly defined, AI uses large amounts of data and machine learning to enable AI technologies to simulate complex tasks typically done by humans (eg, drive cars, solve problems, and generate intelligent text). In clinical medicine, AI’s potential to save time and reduce administrative burden has been highly touted, with programs like AI scribes for note writing, AI-generated text for responding to patients, and AI assistance in diagnostics.3,4 While AI does chip away at these task-oriented pain points, there are countless unknowns about how AI will affect the medical education of our next generation of physicians. It is critical that trainees understand AI tools should not substitute for the development of their own critical thinking and diagnostic skills. They must also learn that present-day AI tools do not meaningfully address structural health system infrastructure issues. AI technology is on track to assist healthcare at the margins by working in the same staid and dysfunctional ways that it has for decades, rather than completely reimagining and revolutionizing a broken system. As educators, we need to empower trainees to use these inescapable and powerful AI tools, while reminding them of the key limitations. In this Perspective piece, we explore how AI will not save our patients nor our healthcare system unless we explicitly address five people-focused areas of vulnerability: medical education, health equity, physician workload, health system dysfunction, and workforce attrition.

AI threatens medical education

ChatGPT (OpenAI, San Francisco, CA, USA) and OpenEvidence (Cambridge, MA) are powerful AI tools rapidly gaining in popularity throughout medical practice and education. While trainees and medical educators are increasingly using these tools, guidance around their use in medical education has not caught up. These tools have tremendous potential to advance medical education both as educational resources for trainees and curriculum and/or assessment assistants for medical educators, but they may also contribute to the deskilling of our most novice trainees and further devaluing already undervalued medical educators.5

We have witnessed trainees using these tools to shortcut critical thinking in diagnostic reasoning (eg, copying basic labs into OpenEvidence and asking for an interpretation). Such an experience deprives trainees of the rigorous iterative process of cognitively processing and building on foundational medical knowledge. In clinical training, the intention of clinical rotations is to continually test trainees to march through a stepwise reasoning process full of trial-and-error. AI tools may unnecessarily short-cut answers to these questions, so we need to ensure that we are teaching our trainees the fundamental knowledge skills and clinical reasoning steps, rather than further deskilling them. Decades of experience with medical resources such as UpToDate (Waltham, MA) have demonstrated that mindfully using educational electronic tools can be useful in adherence to evidence-based practice6—but we need to avoid short-cuts to critical reasoning. We recommend discussing appropriate use of these tools, as has been standard with other medical resources, and excluding their use in assessments.

For medical educators, faculty development in the capabilities of AI and existing resources is urgent. As discussed above, these tools have been deployed and are being used in the wild by trainees with little supervision. Institutions must urgently invest in faculty development resources and workshops and create standard guidance around their use.7 Medical educators should also be alert to how AI tools may influence specialty selection, as trainees may begin to select against specialties that they fear are most threatened by AI.

Finally, we worry that medical educators, already undervalued by health systems aligned with academic medicine, will promote AI to replace human-oriented medical education activities, such as tutoring, remediation, grading, and assessment. Although AI may make curriculum development and assessment more ­efficient in some ways, potential consequences include unintentional plagiarism, hallucinations of inaccurate medical information, and propagation of bias and misinformation. AI assistants could potentially standardize and reduce bias in grading, but they should not completely replace human educators in the assessment process because of the threats of hallucination. Future studies should rigorously explore the impact of AI on assessment/evaluation, the impact of curriculum development using AI, AI’s adaptations of clinical teaching, and more. The ideal state imagines a framework where human-AI interfaces can enhance medical education rather than deskill or devalue human educators.

AI doesn’t meaningfully address health equity

AI threatens to exacerbate inequalities in “techquity,” which may worsen health disparities.8 For example, rural areas with a limited healthcare workforce may try to address the problem by replacing providers with AI-human interfaces. This solution inadequately addresses the quality-of-care divide for people in lower resource settings. Further, people with limited tech literacy may struggle to decipher the bewildering array of new technology they encounter, intended to “solve” their healthcare woes. While AI may write our notes, it is unclear how AI will help people whose lives and wellness are threatened by structural determinants of health such as being unhoused, experiencing food insecurity or suffering from substance use disorders. For people with COPD living in rural areas, they have higher morbidity and obstacles to care such as not having seen a doctor due to cost.9 AI does not address the high medication costs that cause people to choose between food and chronic disease management, nor does it address the severe primary care and subspecialty shortage in rural and resource limited settings. Instead, AI and related technologies often end up helping tech-savvy higher-income patients further optimize their already excellent healthcare navigation, leaving others with fewer resources even further behind.

AI may paradoxically increase physician workload

In our view, AI threatens to increase physician workload, making it even less sustainable. For trainees in Pulmonary and Critical Care Medicine (PCCM), this problem is especially acute, given the high prevalence of burnout (32%) and depressive symptoms (41%).10 The uniquely human patient/provider relationship is built by time and trust to create behavior changes which improve health outcomes like hypertension or obesity management. However, most outpatient visits are too short for a real conversation, clinics and hospitals are understaffed, billing drives documentation, and a lopsided relative value unit (RVU) model incentivizes procedure-driven healthcare over preventive medicine. AI innovations such as AI scribes promise to return doctors to the bedside to spend less time on documentation, nevertheless we predict that physicians will be asked to use the “time saved” by AI to see more patients to drive more RVUs for increasingly corporate environments, with even less human capital resources. PCCM trainees report that clinical burden is already the number one determinant of their burnout, and exacerbations of this will certainly lead to the physician attrition that has been accelerated following the COVID-19 panedmic.11,12 AI interventions to address physician workload do not radically alter this broken system and instead may lead to even greater workload expectations.

AI doesn’t solve intrinsic U.S. healthcare system dysfunction

The dysfunction of the U.S. health system puts an overwhelming burden on the physicians who work within it. Physicians try to solve problems with individual overwork. Their professionalism leads them to call patients after hours, investigate alternative covered medications because of understaffed clinics, and decipher ever-changing insurance guidelines that create mounting paperwork. During the COVID-19 pandemic and in subsequent years, physicians experienced a loss of autonomy, increased burnout, and reduced morale due to hiring freezes and budget-tightening efforts, which left physicians to navigate patient care with dwindling support. When basic medications like albuterol inhalers require prior authorizations, physicians spend extraordinary uncompensated time with this unreasonable administrative burden, desperately working to avoid dire consequences to patient health.13 As a result, physicians end up feeling like “cogs of capitalism” caught in a tangled web of corporations including insurers, pharmaceutical companies, and private equity-owned health systems, each using their own AI bot to battle the other.14,15 Physicians’ status as professionals has atrophied in depersonalizing environments where providers are at the whim of increasingly corporatized health systems. As a result, physicians are searching for ways to regain autonomy, taking measures such as unionizing.16

AI doesn’t address the clinical workforce emergency

Given the previously described health inequities and tech inequities, U.S. healthcare system dysfunction, and unsustainable physician workload, it is no wonder that our healthcare workforce is hemorrhaging, particularly in critical fields like pediatrics and primary care.17 When physicians go to work, they experience understaffed clinics and wards, more inter-visit care, and rising patient expectations. As a result, physicians are turning to “quiet quitting” (or actual quitting).14 In conjunction with these larger trends in healthcare, trained physicians are leaving medicine for non-clinical careers after medical school or residency.

While the end of the workforce pipeline is leaking, the beginning is too narrow to let in the needed volume of prospective physicians. Medical schools and residencies are overflowing with interested applicants but paradoxically the physician shortage is increasing, with a deficit projected to be up to 86 000 physicians by 2036.18 AI does not significantly address this major rate-limiting step in creating a physician and we fear that health systems might presume that training an AI model is cheaper and easier than training a new healthcare worker. Successful matriculants to medical school quickly learn which specialties are undercompensated with high workloads and make rational economic choices that align with these facts. A common ­consequence of these choices is that many medical students would rather remain unmatched and reapply to “competitive” specialties than pursue careers in primary care. To date, we are not aware of any AI interventions that incentivize trainees to specialize in primary care or realign trainee choices with the needs of an aging population.

So how can AI save us?

Before accusations arise that we are anti-technology Luddites, we must share that we are enthusiastic early adopters who have signed up to pilot AI scribe programs at our institution. We know AI is not designed to solve every problem. However, the narrow focus on AI may lead to investments in AI at the expense of human capital or radical changes to a broken system. We assert that it is misguided that many digital health initiatives focus on solving “1% problems” for people with high tech literacy and adequate access to high quality care, instead of focusing on transformational changes.

Medical educators and schools should urgently consider potential effects on trainees, begin guarding against attrition of critical thinking skills at risk from AI tools, and alert medical educators to the benefits and potential risks of AI tools, both to medical education and curriculum development. Beyond scribing, patient messaging and diagnostics, there are numerous unmet opportunities where AI technology could impact healthcare more boldly.4 AI patient navigators could shepherd patients with limited health literacy or access to care in order to successfully navigate our incomprehensible healthcare system. Visit notes and instructions should be translated by AI into patients’ native languages and would allow people with limited English proficiency to have an improved healthcare experience which would translate to real quality-of-care outcomes. AI tools could make patient instructions interactive, not only to be re-written automatically at various health literacy levels but also to provide reminders of tasks and to give updates to their health team (which must still be appropriately staffed to manage the increase in bidirectional information sharing). At the federal level, legislators should consider how corporate AI technology may further obstruct patient care in an already gridlocked system, if patients and physicians fail to reach a human on the other end of the phone or chatbot to address their urgent need.

Health AI has immense potential, but tech companies and healthcare corporations must carefully harness the expertise of frontline clinicians, clinician-educators, and patient experiences to reimagine solutions. AI innovations must happen in parallel with efforts to recruit and retain a larger workforce, whose alarming attrition is threatening the health of the U.S. population. For the foundation of our health system to improve, we must align economic incentives at the highest levels by appropriately compensating primary care specialties, attracting physicians to rural and lower-resource settings, and addressing the high costs of care for patients. We hope AI will save us, but until its transformative abilities are refocused on people instead of solving corporate problems, AI won’t save us yet.

Supplementary Material

aapag009_Supplementary_Data

Contributor Information

Lekshmi Santhosh, Email: Lekshmi.Santhosh@ucsf.edu, Division of Pulmonary, Critical Care, Allergy & Sleep Medicine, University of California, San Francisco, San Francisco, CA, United States.

Leah J Witt, Division of Pulmonary, Critical Care, Allergy & Sleep Medicine, University of California, San Francisco, San Francisco, CA, United States; Division of Geriatrics, University of California, San Francisco, San Francisco, CA, United States.

Author contributions

All authors contributed to the manuscript.

Supplementary material

Supplementary material is available at ATS Scholar online.

Conflicts of interest

Please see the ICMJE disclosure forms, which have been provided as supplementary material. The authors report no conflict of interest.

Artificial intelligence disclaimer

No artificial intelligence tools were used in writing this manuscript.

References

  • 1. Saag HS, Shah K, Jones SA, Testa PA, Horwitz LI.  Pajama time: working after work in the electronic health record. J Gen Intern Med. 2019;34:1695-1696. 10.1007/s11606-019-05055-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Adler-Milstein J, Zhao W, Willard-Grace R, Knox M, Grumbach K.  Electronic health records and burnout: time spent on the electronic health record after hours and message volume associated with exhaustion but not with cynicism among primary care clinicians. J Am Med Inform Assoc. 2020;27:531-538. 10.1093/jamia/ocz220 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Rotenstein L, Melnick ER, Iannaccone C, et al.  Virtual scribes and physician time spent on electronic health records. JAMA Netw Open. 2024;7:e2413140. 10.1001/jamanetworkopen.2024.13140 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Sarkar U, Bates DW.  Using artificial intelligence to improve primary care for patients and clinicians. JAMA Intern Med. 2024;184:343-344. 10.1001/jamainternmed.2023.7965 [DOI] [PubMed] [Google Scholar]
  • 5. Budzyń K, Romańczyk M, Kitala D, et al.  Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10:896-903. [DOI] [PubMed] [Google Scholar]
  • 6. Isaac T, Zheng J, Jha A.  Use of UpToDate and outcomes in US hospitals. J Hosp Med. 2012;7:85-90. 10.1002/jhm.944 [DOI] [PubMed] [Google Scholar]
  • 7. Banerjee M, Chiew D, Patel KT, et al.  The impact of artificial intelligence on clinical education: perceptions of postgraduate trainee doctors in London (UK) and recommendations for trainers. BMC Med Educ. 2021;21:429. 10.1186/s12909-021-02870-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Clark CR, Akdas Y, Wilkins CH, et al.  TechQuity is an imperative for health and technology business: let’s work together to achieve it. J Am Med Inform Assoc. 2021;28:2013-2016. 10.1093/jamia/ocab103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Gaffney AW, Hawks L, White AC, et al.  Health care disparities across the urban‐rural divide: a national study of individuals with COPD. J Rural Health. 2022;38:207-216. 10.1111/jrh.12525 [DOI] [PubMed] [Google Scholar]
  • 10. Sharp M, Burkart KM, Adelman MH, et al.  A national survey of burnout and depression among fellows training in pulmonary and critical care medicine: a special report by the association of pulmonary and critical care medicine program directors. Chest. 2021;159:733-742. 10.1016/j.chest.2020.08.2117 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. O’Toole J, Zaeh S, Eakin MN, et al.  Balancing demands: determinants of burnout reported by fellows in pulmonary and critical care medicine. ATS Sch. 2021;2:108-123. 10.34197/ats-scholar.2020-0097OC [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Abbasi J.  Pushed to their limits, 1 in 5 physicians intends to leave practice. JAMA. 2022;327:1435-1437. 10.1001/jama.2022.5074 [DOI] [PubMed] [Google Scholar]
  • 13. Bernstein JA, Gadde D, Yassin M, et al.  The impact of prior authorization on clinical practice and patient care outcomes: a work group report of the AAAAI prior authorization task force. J Allergy Clin Immunol Pract. 2024;12:1719-1726. 10.1016/j.jaip.2024.01.051 [DOI] [PubMed] [Google Scholar]
  • 14. Rosenbaum L.  On calling-from privileged professionals to cogs of capitalism?  N Engl J Med. 2024;390:471-475. [DOI] [PubMed] [Google Scholar]
  • 15. Rosenbluth T. In constant battle with insurers, doctors reach for a Cudgel: A.I. Published online July 16, 2024. https://www.nytimes.com/2024/07/10/health/doctors-insurers-artificial-intelligence.html
  • 16. Schulman K, Richman B.  Hospital consolidation and physician unionization. N Engl J Med. 2024;390:1445-1447. 10.1056/NEJMp2400463 [DOI] [PubMed] [Google Scholar]
  • 17. Jordan KA, Rao P, Byerley JS.  Where are all the pediatricians?  JAMA. 2024;332:783-784. 10.1001/jama.2024.8485 [DOI] [PubMed] [Google Scholar]
  • 18. Prepared for the AAMC by GlobalData Plc. The Complexities of Physician Supply and Demand: Projections from 2021 to 2036, 2024.

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

aapag009_Supplementary_Data

Articles from ATS Scholar are provided here courtesy of American Thoracic Society

RESOURCES