Why Doctors Still Resist Better AI: Identity, Not Accuracy

I kept running into the same contradiction this semester. Paper after paper showed AI systems outperforming clinicians at specific diagnostic tasks. DeepMind\u2019s mammography model matched or beat radiologists. Google Health\u2019s retinal scan AI hit 94 percent accuracy on diabetic retinopathy detection. And then I would read a completely separate set of papers showing that most of these tools have not been adopted at scale. The two facts sit next to each other in the literature and nobody seems surprised by the gap anymore. IBM Watson Health raised hundreds of millions of dollars, promised to transform oncology, and was shut down in 2022 after years of underperformance. The pattern is not a technology problem. It is an identity problem.

I have been reading identity theory in the tradition of Stryker (1980, 2001) and Burke (1991) for my comps work, and I wrote about it more broadly in a previous post on identity theory and technology resistance. The core argument is simple. People hold identities, sets of meanings about who they are in a particular role. When events or technologies generate perceptions that mismatch those meanings, the person experiences distress and acts to restore verification. The mechanism is not cognitive override. You cannot show a radiologist a better AUC score and expect her to suddenly embrace a tool that redefines what it means to do her job.

Strich et al. (2021) studied this exact mechanism in their JAIS paper on the CleverLoan case. A bank deployed a substitutive decision-making AI system for loan consulting. The AI took over the core activity of evaluating loan applications. Experienced loan consultants, the ones whose professional role identity was built on expertise, discretion, and the judgment that comes from years of reading cases, experienced the AI as an identity threat. They felt deskilled. Their autonomy was restricted. Their professional role identity, the answer to the question \u201cwho am I at work?\u201d, was under attack. Less experienced service employees who were promoted into the same role through the AI system experienced the exact same technology as empowerment. The technology was constant. The identity was the variable.

Now apply this to healthcare. A radiologist\u2019s professional identity is built on multiple things. Clinical expertise, obviously. But also discretion, the authority to make the final call on what a scan shows. And the relationship with the referring physician and the patient, the human interaction that gives the work its meaning. An AI that flags suspicious nodules on a CT scan, and that does it with higher sensitivity than the radiologist, does not just add efficiency. It changes the meaning of the work. If the AI catches everything the radiologist would catch and a few things the radiologist might miss, the radiologist\u2019s role shifts from primary reader to AI supervisor. The identity standard, \u201cI am the expert who sees what others miss\u201d, no longer fits the situation.

The reason IBM Watson Health failed, I think, goes beyond bad technology or overhyped promises. Watson was positioned from the start as a system that would replace physician judgment, not support it. The branding was explicit. IBM marketed it as a cognitive system that could outthink doctors. When the tool could not deliver on that promise, the failure was spectacular. But even if it had worked, the positioning would have generated the same identity resistance that Strich et al. documented. The experienced oncologists whose identity depended on treatment planning expertise were not going to embrace a system that treated their judgment as optional.

This is where the distinction between augmentation and replacement becomes concrete in a way that the IS field has not fully theorized. Baird and Maruping (2021) showed that for agentic IS, the foundational construct should shift from use to delegation. I wrote about what that shift means in practice in an earlier post on delegation and agentic systems. The human does not simply use an AI tool. They delegate tasks to it, appraise its outputs, distribute work between human and system, and coordinate outcomes over time. I think about this as a design principle in identity terms. The AI systems that succeed in healthcare will be the ones that let clinicians delegate detection tasks while keeping the identity-relevant work, the judgment that follows detection, the conversation with the patient, the differential diagnosis reasoning, the contextual adjustment of a finding based on clinical history. In identity theory language, the system must verify the professional identity standard, not disrupt it.

Jussupow et al. (2024), in their MISQ paper on algorithm aversion, make a closely related point. People\u2019s reactions to algorithms are shaped not just by accuracy but by interface design, task context, and identity implications. A financial advisor may reject an algorithmic portfolio recommendation not because the algorithm performs poorly but because relying on it threatens their professional self-concept. A radiologist may resist diagnostic AI not because it is inaccurate but because the interface presents the AI\u2019s finding as the conclusion rather than a recommendation. The design of the interaction, not just the accuracy of the model, determines whether identity threat occurs.

Berente et al. (2021) established that AI is qualitatively different from traditional IT because of its autonomy, learning, and inscrutability. Those three facets make identity threat worse. A traditional system is a tool. You can ignore it, override it, or understand its logic. A substitutive AI takes the decision and does not always allow override. When the system learns and changes its behavior over time, the professional cannot even predict when their expertise will be relevant. And when the system is inscrutable, the professional cannot explain why a different decision might be better, which undermines the core experience of being the expert.

The practical implication is straightforward but hard to implement. \u201cAugment not replace\u201d has been a slogan for years, but it has not been treated as a design principle grounded in theory. If I were advising a healthcare technology company, I would start with identity analysis, not feature analysis. What is the identity standard for the clinician who will use this tool? What parts of that identity does the tool verify, and what parts does it threaten? The tools that succeed will be the ones that automate the parts of the work clinicians are willing to delegate and leave the identity-relevant parts under human control.

DeepMind\u2019s mammography model is still not widely deployed in clinical practice. Neither are most of the high-performing diagnostic AIs in the literature. The accuracy was never the bottleneck. The identity threat was always the real barrier. The models that finally do get adopted will not be the most accurate ones. They will be the ones designed around what it means to be a doctor.