The global health workforce is facing a critical shortage, with projections indicating a deficit exceeding 11 million healthcare workers by 2030. At Google, we are researching how AI can transform education for health professions to help close this gap with studies exploring how Google’s AI models can serve as effective personalized learning tools in medical learning environments.
Two such studies are presented here today. First, in “Generative AI for Medical Education: Insights from a Case Study with Medical Students and an AI Tutor for Clinical Reasoning,” which was published at CHI 2025, we used interdisciplinary co-design workshops, rapid prototyping, and user studies to take a qualitative approach to comprehending and designing for medical students. Next, in our latest update of “LearnLM: Improving Gemini for Learning”, we quantitatively assessed LearnLM — our Gemini-based family of models fine-tuned for learning — on medical education scenarios through preference ratings from both medical students and physician educators. Both studies revealed a strong interest in AI tools that can adapt to learners and incorporate preceptor-like behaviors, such as providing constructive feedback and promoting critical thinking. When compared to base models, physician educators rated LearnLM as having superior pedagogy and acting “more like a very good human tutor.” These novel capabilities are now available with Gemini 2.5 Pro.
Understanding the medical learner
Employing a learner-centered approach has been critical in guiding our development of responsible AI tools that scale individualized learner pathways and augment competency-based approaches. Central to this approach, we first conducted formative user experience (UX) research to understand medical learner needs. Through a participatory design process, we began with a co-design workshop that convened an interdisciplinary panel of medical students, clinicians, medical educators, UX designers, and AI researchers to define opportunities for incorporating AI in this space. An AI tutor prototype was developed using the insights from this session to specifically guide students through clinical reasoning based on a fake clinical vignette. We then evaluated the AI tutor prototype’s helpfulness in a qualitative user study with eight participants (4 medical students and 4 residents). The study aimed to elicit participant learning needs and challenges as well as their attitudes toward AI assistance in education. Each participant engaged in a 1-hour session with a UX researcher involving semi-structured interviews and interactive sessions with the prototype. All sessions were remote and conducted through video conferencing software. Participants accessed the prototype through a web link and shared their screen while interacting with the prototype.
Through our thematic analysis of medical learner interviews, we discovered a number of obstacles to the development of clinical reasoning abilities and the potential for generative AI to overcome these obstacles. Medical students, for instance, expressed a strong interest in AI tools that can adapt to individual learning styles and knowledge gaps. Participants also highlighted the importance of preceptor-like behaviors, such as managing cognitive load, providing constructive feedback, and encouraging questions and reflection.
Getting to know medical students where they are Building on these insights, we conducted a blinded feasibility study with medical students and physician educators to quantitatively assess LearnLM’s pedagogical qualities in medical education settings compared with Gemini 1.5 Pro as the base model. In collaboration with experts, we designed a set of 50 synthetic evaluation scenarios across a range of medical education subjects, from pre-clinical topics, such as platelet activation, to clinical topics, like neonatal jaundice, reflecting the core competencies and standards in medical education.
We recruited medical students from both preclinical and clinical phases of training to engage in interactive conversations with both LearnLM and the base model, in a randomized and blinded manner. Students used the evaluation scenarios to role-play as different types of learners across a range of learning goals and personas, generating 290 conversations for analysis. Each scenario provided learners with context to standardize the interaction as much as possible between both models, including a learning goal, grounding materials, a learner persona, a conversation plan, and the initial query used by the learner to start the conversation.
Students then rated model behavior by comparing the two interactions for each scenario side-by-side across four criteria: (1) overall experience, (2) meeting learning needs, (3) enjoyability, and (4) understandability. Physician educators rated model behavior by reviewing conversation transcripts and scenario specifications. For each scenario, educators reviewed the transcripts from both learner-model conversations side-by-side, and provided preference ratings across five criteria: (1) demonstrating pedagogy, (2) behaving like a very good human tutor, (3) instruction following, (4) adapting to the learner, and (5) supporting the learning goal. We averaged three reviews from independent educators for each conversation pair. All preference ratings were done in a randomized and blinded manner using 7-point scales, which reflected a spectrum of preference strengths including the option to express no preference between the two models.
Physician educators consistently preferred LearnLM across all five of the comparison criteria. They judged LearnLM particularly positively in terms of demonstrating better pedagogy (on average, +6.1% on our rating scale) and for behaving “more like a very good human tutor” (+6.8%). When we simply look at whether educators expressed any preference one way or the other — regardless of its magnitude — LearnLM emerged as their choice in a clear majority of assessments across every criterion. On average, +9.9% on our rating scale, medical students expressed the strongest preference for LearnLM’s more enjoyable interaction. For the other three comparison criteria, student preferences were less pronounced, but LearnLM was also favored from a directional standpoint. This study points to LearnLM’s potential to transform education and learning paradigms and scale a competent health workforce. None of the data used for model development or evaluation in this study included real patient data. See the tech report for modeling details.
Education for health professions in a new light We recently presented this study at the Nobel Forum’s MedEd on the Edge conference, where we also organized a hands-on workshop for the international medical education community to investigate these options. We acknowledge that educators play both pedagogical experts and explorers in this rapidly expanding field of knowledge. A responsible future necessitates careful consideration of issues like ensuring accuracy, reducing bias, and preserving the essential role of human interaction and oversight. It underscores the need to re-evaluate competencies and entrustable professional activities, and for curricula that cultivate adaptive expertise, focusing not only on AI applications in education, but also on teaching foundational understanding of AI itself. At this convergence, generative AI can serve as a catalyst for the desired productive struggle to foster deeper understanding and critical thinking. As the journey has only just begun, below are a few examples of how Google’s AI can potentially transform health professions education.
