Human-led, AI-enabled: preventing another AI winter through workforce transformation

Artificial intelligence (AI) has played a role in life sciences research and healthcare delivery for decades, though its nature and applications have changed with each new technology. As early as the 1950s, symbolic logic models provided clinical decision support: a form of deterministic AI encoded human knowledge as rules that an AI logic model could follow and implement (Nawaz et al., 2025). These early rule-based systems highlighted both the promise and limitations of symbolic AI. Subsequent periods of over-expectation followed by underdelivery contributed to recurring "AI winters" in which investment and enthusiasm declined.After 2010, computational and algorithmic developments led to machine learning (ML) and deep learning models. While these performed extremely well in some use cases (Habehh and Gohel, 2021), notably in biomedical research and image analysis, their use remained largely concentrated among technical experts and specialized applications. In the 2020s, large language models (LLMs) have rapidly changed the AI landscape in life sciences and medicine, and we are on the cusp of even faster evolutions, with state space models and active inference-based concepts that could lead to non-linear, novel leaps.The adoption of LLMs has been unlike that of any other modern technology. Most innovations, such as cars, mobile phones, and the internet, have taken years (between 1 year and 75 years) to reach adoption by 50 million users globally (Desjardins, 2018), whereas ChatGPT took only 5 days to achieve this. This unprecedented uptake leads to a paradigm shift in how we think about the users, use cases, and impact of a new technology. LLMs could change working practices in nearly every industry, to varying degrees. In their lead article, Guni et al. articulate the potential influence of LLMs in life sciences research and healthcare delivery (Guni et al. 2026). They review the rise and fall of AI applications in biomedical research and clinical decision support, and robustly discuss the elements needed for the current wave of AI to succeed. Specifically, they comment that "multidisciplinary effort (is needed) across several key domains: i) algorithmic science and architectures, ii) datasets, iii) processing power and efficiency, iv) infrastructure, v) human factors through codesign, and vi) interoperability with both legacy and modern healthcare systems." These efforts will require collaboration between clinicians, researchers, technologists, patients, ethicists, policymakers and others, and Guni et al. discuss a thorough breadth and depth of key topics. However, a seventh critical domain underlies all the others: the transformation of biomedical education and workforce development. Without a workforce capable of effectively leading AI systems, even the most sophisticated algorithms, datasets, and infrastructure will fail to deliver lasting impact, and we risk another AI winter.With respect to education, the widespread yet heterogenous use of LLMs by students and professors alike is challenging current pedagogy and assessment frameworks. In some cases, students are using LLMs to study or complete homework assignments, yet their professors lack expertise and are unprepared to alter their syllabus, homework, or exams (Lelescu et al 2025). In other cases, faculty are trying to integrate LLMs but are being met by significant resistance by students (Farinosi and Melchior, 2025). LLMs are also changing the day-to-day activities of our workforce, where adoption is similarly mixed and there is limited infrastructure to support broad, comprehensive workforce development. Many proposed educational responses focus primarily on the development of AI competencies across the continuum of learners and professionals. However, an exclusive focus on AI competencies introduces two related risks: "cognitive de-skilling", in which professionals gradually lose critical expertise and skills due to over-reliance on AI and "never-skilling", where the next generation of learners never fully develops foundational competencies because AI systems perform those tasks from the outset.Currently, health systems plan to measure the immediate augmented performance of AI tools in terms of clinical outcomes or immediate workflow efficiencies; we have yet to comprehensively measure their impact on retained competency or the longer-term implications of their use as a performance standard. Over the next 50 years, AI will likely become a critical part of many decision-making processes, training programs, and assessments done in clinical settings. However, if we never measure the retained competency of clinical staff to understand the impact of AI on long-term skill erosion, we could unintentionally introduce a critical system failure in the future. Clinicians who can no longer spot errors in AI tools or perform tasks when tools fail may put patients at risk and create a new type of AI winter-a systemic failure due to over-reliance on AI and lack of human competency. This AI winter would be extremely damaging to the practice of medicine in multiple ways: 1) direct risk to patients, 2) time to re-skill a generation of medical practitioners, and 3) lack of medical leaders with expertise to train the clinicians.Life sciences and healthcare are not the first knowledge-intensive domains to confront the challenge of maintaining human expertise as automation becomes increasingly capable. As commercial aviation systems evolved to automate key functions such as takeoff, landing, and navigation, concerns emerged that pilots were becoming overly reliant on automation, potentially eroding critical manual flying skills and reducing resilience when automated systems failed. Recognizing this risk, the United States Federal Aviation Administration (FAA) issued a Safety Alert in 2017 (FAA, 2017) and further guidance in 2022 (FAA, 2022) to strengthen policies and training that preserved pilot proficiency in manual operations.This experience provides a compelling parallel to the adoption of AI-driven clinical decision support systems. In both aviation and healthcare, humans and technology function as components of a sociotechnical system (Trist, 1981), where safety depends on the effective interaction between people, processes, and tools. When technological capabilities advance without corresponding attention to human competencies, organizations risk losing the very redundancies that protect against failure. The FAA's approach demonstrates that successful integration of automation requires more than technical performance; it requires deliberate governance, training, monitoring, and ongoing investment in human expertise. Ultimately, the safety of any automated system depends not only on the technology itself, but also on how human competency is preserved and strengthened alongside it (Macnamara et al., 2024).In short, humans must be the redundancy when technology systems fail.To avoid creating a generation of professionals who cannot function effectively when AI systems are unavailable, unreliable, or incorrect, we need radical transformation in our education and workforce development frameworks to teach both AI competencies and domain-specific competencies in synchrony. We must develop these competency frameworks now to prepare the future workforce to thrive in an AI-enabled world.Perhaps the easier competency framework is that of AI competencies. For example, prompt engineering, AI literacy, validation and oversight, ethical and responsible use of AI, agentic AI, and the nature of human-AI collaborations are obvious topics that biomedical education for life sciences research and healthcare delivery must begin to learn. Many programs around the globe have begun offering didactic and practical-based educational programs to serve this need. Perhaps more challenging is the listing of domain-specific competencies. Not every fact that previous generations were expected to memorize will remain equally valuable in an AI-enabled environment. The challenge lies in determining which foundational knowledge remains essential for independent critical thinking and reasoning, and which tasks can appropriately be augmented by AI. Future domain-specific competency frameworks may place less emphasis on memorization and greater emphasis on reasoning, interpretation, critical evaluation, communication, judgement, and collaboration. Each domain needs to develop their list of domain-specific competencies and then design pedagogy, training, and learning assessment around these competencies, assuming that AI will be a partner for the learner. Neither AI competencies nor domain-specific competencies alone are sufficient; future professionals will require mastery of both.The goal should be to leverage AI to improve the efficiency and effectiveness of biomedical research and clinical practice, allowing humans to focus on uniquely human activities such as critical thinking, creativity, and interpersonal connection, rather than repetitive and mundane tasks.Our society has seen workforce transitions before, for example as result of the invention of electricity, the Industrial Revolution, and the emergence of the internet. For example, online streaming services have rendered many jobs unnecessary, such as cashiers in video rental or record stores, while creating thousands of jobs never imagined in the 1990s. Similarly, if the current AI era succeeds some workforce roles will vanish, but many new roles will emerge, some unimaginable today. Leaders, researchers, educators, and workforce members must proactively develop systems and frameworks that enable current and future generations to thrive with AI technology at their fingertips.The future of biomedical research and healthcare should not be defined by "humans in the loop", but by "humans in the lead". Our challenge is not simply to build more capable AI systems. It is to build educational and workforce development frameworks that preserve human expertise while enabling effective human-AI partnership, i.e., systems that are "human-led, AI-enabled." History suggests that AI winters occur when technological capabilities outpace the ability of institutions, professionals, and society to effectively adopt and sustain them. Workforce transformation may therefore be one of the most important safeguards against another cycle of overpromising and underdelivering. If we succeed, AI can become one of the most transformative technologies in the history of medicine. If we fail, another AI winter may be closer than we think.