In our technology adoption research, 177 orthopedic surgeons were surveyed about their use of generative AI, and a subset participated in in-depth interviews. Participants were optimistic and recognized meaningful benefits. Yet 71% never used generative AI in activities touching the operating room. When they did use it, it was backstage: for literature searches, preparing presentations, and supporting their own learning.
Now consider the classroom. In the 2025 Walton Family Foundation and Gallup survey of US public-school teachers, six in ten reported using AI during that school year, mostly for worksheet creation, lesson planning, and administrative work. Frontline activities involving direct interaction with students were far less common: grading at 16% and one on one instruction at 14%.
These two professions could hardly be more different, yet they show the same adoption curve: high use where the work is preparatory, and minimal where the professional directly meets the individuals they serve.
This is the taste problem
The usual explanation for this pattern of adoption is a gap in training or confidence. Improve professional development, the argument goes, and the adoption will follow. We no longer believe that is sufficient.
What is missing is not competence with the technology itself, but something harder to codify: professional judgment. In the technology sector, this is described as taste: the ability to discern what is right for a particular situation, a form of knowledge that resists being reduced to rules. Scott Young captured this sharply in a recent essay: taste is largely tacit, while model pretraining is a form of book learning that omits expert judgments that could not be put into words. A 2026 study on scientific judgment identified the same boundary: AI capability drops sharply when clear answers give way to discrimination under uncertainty.
A surgeon interpreting a fracture pattern that does not match the textbook is exercising taste. So is a fifth-grade teacher who hears a wrong answer and knows whether it is a slip to move past or a misconception worth stopping the lesson for. In both cases, the challenge is not the application of explicit rules but the interpretation of context. Every case is slightly different, and what experts know is tacit, acquired through experience, and often difficult to articulate clearly.
What is missing is not competence with the technology itself, but something harder to codify: professional judgment.
Why AI cannot skip the hard part
Before AI can teach, assess, or give feedback on a complex skill, someone must define what good performance looks like. That ground truth does not exist until it is elicited from experts, negotiated, formalized, and validated, and that process is slow and expensive.
In our own work at the AO Foundation, where we lead education design and education research, defining competence for a single surgical procedure typically requires an international expert-consensus study. We have undertaken such efforts for procedures including plating a wrist fracture, wiring an elbow, and femoral intramedullary nailing. Each required a global panel of experts and months of work, and none of the resulting tools transfers directly to the next procedure.
Schools know a version of this ritual as a norming session: teachers review student essays together until they agree what a proficient argument looks like, turning tacit judgment into shared ground truth. AI essay scorers work where such rubrics are explicit and broadly agreed upon. A 2026 study of automated essay scoring found exactly this pattern: the models matched human raters where the rubric was clear and the raters agreed, and lost the thread where raters disagreed. A second 2026 analysis found the same weakness from another angle: the models could not tell the best essays apart from the merely good ones. In other words, the machine inherits the agreement, not the judgment. Where experts have not yet made their standard explicit, there is nothing for the model to learn.
Frontier AI laboratories have recognized this challenge. In August 2026, Google agreed to pay 10 million dollars at a bankruptcy auction for the internal records of Spirit Airlines: 100 million emails and 500 million chat messages, with the links between them preserved so that a model could follow how a problem was handled from start to finish. The runner-up bidder was Mercor, a company that pays doctors, lawyers, and other experts by the hour to record how they reason, and whose founder says AI labs need such contractors because companies will not hand over data that would let models automate their workflows.
To learn judgment, a model needs recorded judgment, and most professions do not produce it. A surgeon’s reasoning unfolds in an operating room and a teacher’s in a classroom. Neither is routinely documented in a form that can be used for machine learning. The bottleneck is not only model capability. It is the scarcity of explicit representation of professional judgment itself.
To learn judgment, a model needs recorded judgment, and most professions do not produce it. A surgeon's reasoning unfolds in an operating room and a teacher's in a classroom. Neither is routinely documented in a form that can be used for machine learning. The bottleneck is not only model capability. It is the scarcity of explicit representation of professional judgment itself.
What the adoption data showed
Three findings from the surgeon AI adoption study are consistent with patterns observed in the teacher data and may therefore help explain teachers’ skepticism in adopting AI.
Attitude was not the barrier to adoption. Surgeons rated generative AI positively across the board, yet clinical use stayed low. Asked why, they did not cite lack of interest. They cited missing clinical evidence, unclear regulation, and management focused on cost rather than innovation. Persuading professionals that AI is valuable is the easy part, and the wrong lever.
The barriers were structural, not social. Nothing the surgeons described pointed to colleagues or peer pressure. Experienced professionals judge new tools on their own terms, against the standard of evidence their field demands. School districts that train a few enthusiastic teachers as AI champions or ambassadors, hoping they will bring colleagues along, are addressing a barrier the surgeons never named, as if peer influence were the obstacle. The surgeon data suggest it is not. Evidence, clear rules, and leadership support are what move experienced professionals, not enthusiastic colleagues.
Career stage changed the value proposition. In the surgeon study, early career surgeons wanted AI to help build their skills, while senior surgeons wanted it to cut administrative work and save time. A similar pattern exists in education: A first-year teacher is more likely than a veteran teacher to want help designing a lesson. A teacher with 20 years of experience is probably more interested in reclaiming time from routine tasks. One value proposition fits nobody.
Backstage first is the rational sequence
Professionals adopt AI where knowledge is explicit and stakes are low. Then they wait before extending its use into areas requiring greater professional judgment. That is not failure. It is a rational adoption pathway.
The 2026 Walton and Gallup data show why caution persists: 69% of teachers have received no guidance at all on using AI for one on one instruction, and fewer than one in ten had received formal guidance on any AI-supported teaching task. Teachers are being asked to carry the judgment risk alone.
Win the preparation and administrative layer, prove reliability there, and earn the right to move closer to the learner.
Professionals adopt AI where knowledge is explicit and stakes are low. Then they wait before extending its use into areas requiring greater professional judgment. That is not failure. It is a rational adoption pathway.
Three implications for learning engineers, education leaders, and AI development teams
- Make expert knowledge elicitation part of the product. Treat the elicitation and validation of expert knowledge as part of AI development, not as a data problem that someone else has already solved. Involve master teachers throughout the design process, just as master surgeons are involved in developing surgical assessment tools.
- Design a visible handoff. Make clear where AI assistance ends and professional judgment begins. The tool may draft feedback, but the teacher remains responsible for determining what a learner needs to hear. Transparency about this division of labor helps preserve trust and accountability.
- Evaluate adoption where it begins. Measure impact in the domains where professionals are already using AI: planning and administration. Too often, evaluations focus on ambitious use cases that remain far from everyday practice.
The practical question is not how to overcome professional skepticism. It is how to build systems that deserve professional trust. The skeptics are showing us where that work must begin.
Matthias Schmidt
1 thought on “From Operating Rooms to Classrooms: The Hidden Barrier to Frontline AI Adoption”
Dear Dr. Matthias Schmidt and Dr. Monica Ghidinelli:
Congratulations on this excellent and thought-provoking article. As a medical educator and AO Trauma Faculty member in Colombia, I am deeply interested in the integration of artificial intelligence into surgical education. Much of my work focuses on training orthopedic residents through clinical case discussions, surgical decision-making, reflective practice, and structured feedback.
Your insights into tacit knowledge and expert judgment are particularly inspiring. They have motivated me to reflect on how these concepts could be incorporated into my daily teaching activities and how AI might support the development of clinical reasoning in future surgeons.
Thank you both for sharing such a valuable perspective and for inspiring new ways of thinking about medical education.
Eduardo Carrillo
AO Faculty Colombia