Why Doctors Avoid AI When Patients Need It Most—and Why They Rely on It When It's Less Useful

Imagine two patients. One has a textbook condition, and the right treatment plan is obvious. The other has a more complex condition, in which even an experienced physician may be unsure of the right treatment plan.

Which patient is more likely to receive assistance from AI?

Most people would intuitively choose the one with the more complex condition. However, this is not always the case.

According to a recent Journal of Marketing Research study by Tinglong Dai and Shubhranshu Singh, physicians are more likely to use AI in textbook cases. In a more complex, high-uncertainty case, the physician may avoid using AI entirely.

We’ve been told that more accurate AI will transform medicine: more accurate algorithms, fewer errors, and better outcomes. But accuracy, it turns out, may not always be the bottleneck.

The Role of Legal and Financial Frameworks

The authors model a common clinical scenario: a physician choosing between a standard treatment plan (the default, low-liability option) and a nonstandard plan (higher reimbursement but also carries higher malpractice risk if something goes wrong). The physician can consult an AI tool that provides an informative but imperfect signal about which plan is right for the patient.

The physician makes the final call. But using AI changes the legal landscape. The study examines two malpractice liability schemes. Under the prevailing liability scheme, the standard of care is defined independently of AI: a physician who follows the standard plan is shielded from liability regardless of what the AI recommends. Under an emerging liability scheme, AI recommendations themselves become the standard of care: the physician is protected if they follow AI recommendations but may be held liable if they ignore them.

Three counterintuitive patterns emerge under both schemes:

1. Physicians overuse AI in low-uncertainty cases
   
   Even when physicians already know the right answer, they may still “use” AI to collect the reimbursement, satisfy a policy requirement, and then ignore its signal entirely. The AI adds no value, but the patient bears the cost.
   
   
2. Physicians may underuse AI when uncertainty is higher
   
   The physician may be less likely to use AI for patients who might benefit more from it. Why? A physician who has a strong preference for prescribing—or not prescribing—the nonstandard plan faces a higher likelihood of receiving an AI recommendation that conflicts with their treatment preference when the uncertainty about the right treatment plan is higher. In these situations, if the patient is harmed, the physician’s liability exposure is more severe. Therefore, the physician may avoid using AI altogether.
   
   
3. The accuracy paradox
   
   As AI becomes more accurate, certain physicians may use it less for some patients. Consider patients for whom the physician already believes the nonstandard plan is correct. If the physician uses AI, the AI might recommend the standard plan instead. As AI becomes more precise, that conflicting recommendation becomes more likely. Now the physician faces a dilemma: follow the AI rather than their own judgment or ignore it and risk liability for deviation. To avoid this position, the physician simply does not use AI with these patients. A better algorithm can therefore lead to less use.

The core insight is simple but easily overlooked: AI changes not just information but also incentives. Once those incentives change, adoption becomes distorted.

As the authors explain:

“We started discussing this idea back in 2018, when deep learning was beginning to move from hype to real clinical promise and people were starting to ask a harder question: What happens to physician liability when something went wrong and AI was part of the decision process? We kept coming back to that question and debated it nearly every week. Covid interrupted the work for a while, and we resumed it in 2021. What kept us interested throughout was that the real issue was never just whether AI is accurate. It was whether the people using it would have the right incentives to use it in the right way.”

We asked Dr. Dai and Dr. Singh to walk us through what they found and what it means for anyone trying to get AI adoption right in medicine.

Q: You’re affiliated with the Hopkins Business of Health Initiative, which puts you close to real healthcare decision making. Did you observe physicians navigating AI recommendations in practice? And if so, what did you see that made you think “the standard adoption story is missing something important here?”

A: What struck us early was that the standard adoption story around AI in healthcare was too simple. The usual framing is: better tools, better decisions, more adoption. But medicine does not work that way. Physicians practice under liability pressure, reimbursement rules, workflow constraints, and institutional norms. So even if AI improves decision quality, that does not mean it will be used when it should be, or followed when it is used.

The paper’s core idea is that AI changes not just information but incentives. And once incentives change, adoption can become distorted.

Q: Hospital adoption: Your model highlights that physicians’ behavior around AI is shaped by liability concerns and financial incentives, not just accuracy. For a hospital administrator rolling out an AI tool, what should they be thinking about beyond getting a high-accuracy algorithm?

A: For a hospital administrator, the key question is not just, “How accurate is the model?” It is also, “What behavior does this model induce?”

A highly accurate algorithm can still be deployed poorly if physicians fear being second-guessed or can bill for using the tool without meaningfully incorporating it into care. We really need to think about governance, accountability, and workflow design. Who is expected to use the tool? When? How will deviations be documented? What happens if a physician consults AI and then ignores it? Good accuracy is necessary, but it is nowhere near sufficient. The real implementation challenge is aligning incentives with appropriate use.

Q: Your model shows physicians sometimes use AI without intending to follow it and sometimes avoid it altogether when it would genuinely help. Do you think hospitals could realistically detect these patterns? And if so, who should own that monitoring responsibility? The quality control team, the insurer, or someone else?

A: Yes, we do think hospitals can detect at least some of these patterns, especially if they track when AI is opened, what it recommends, and whether the final clinical decision is consistent with that recommendation.

The question is how hospitals respond to such patterns. A physician should not be expected to follow AI blindly, so discordance by itself is not a problem. A more meaningful signal is systematic behavior, such as routinely calling up AI only in low-stakes cases, or using it performatively and then ignoring it. This sits at the intersection of quality, compliance, and clinical leadership.

Q: Your model frames the insurer as maximizing patient welfare, leading to the conclusion that reimbursing only when physicians follow AI is preferred. In practice, insurers also have strong cost-minimizing motives. Do you think they’d reach the same conclusion for different reasons? And does that distinction matter for policy?

A: Yes, insurers could very well arrive at similar reimbursement rules for different reasons. In the paper, we model the insurer as maximizing patient welfare, which provides a clean benchmark. In practice, a cost-focused insurer may also prefer to reimburse AI only when it meaningfully affects the decision, because paying simply for “clicking on AI” invites waste.

That said, motive still matters. A welfare-oriented insurer and a cost-minimizing insurer may endorse the same policy on paper, but they may differ in where they draw the line, how aggressively they audit, and whether they leave room for clinical judgment.

Sometimes the welfare-maximizing rule and the cost-minimizing rule point in the same direction, but they use different logic, and that difference shows up in implementation.

Q: Your paper shows that policies designed to increase AI adoption, such as paying physicians for using it, can backfire, creating both overuse and underuse. If you could change one thing about how policymakers currently think about AI in healthcare, what would it be?

A: If we could change one thing, it would be the instinct to treat more or less AI use as automatically better. That is the wrong objective. The goal should be appropriate use, not maximal or minimal use.

Our paper shows that policies meant to boost adoption can backfire. You can end up with physicians using AI when it adds little value, and avoiding it when it could help the most. That is a classic policy mistake: measuring success by uptake instead of decision quality.

Q: Your model is built around the U.S. fee-for-service system, where physicians earn more from nonstandard plans and face malpractice liability. What happens in a system like the UK’s NHS, where physicians are salaried and liability works differently? Do the overuse and underuse problems disappear, or do they show up in a different form?

A: The U.S. setting makes the incentive distortions especially visible because fee-for-service reimbursement and malpractice exposure are both strong. In a system like the NHS, the exact mechanism would look different, but we would not expect the problem to disappear.

Once you move away from fee-for-service, the financial motive may weaken, but other pressures remain: guideline adherence, organizational accountability, professional norms, resource constraints, and reputational concerns. So I expect overuse and underuse to persist, but in different forms.

The broader point is that this is not just an American billing story; it is a human incentives story. Change the system, and the distortions change shape, but they do not vanish.

Q: In your model, patients bear real costs from AI use (privacy concerns, potentially higher premiums) yet have no say in whether AI is used for their treatment plan. Where does the patient’s voice fit in? And would giving patients more agency change the dynamics you identify?

A: This is an important question. In many real-world settings, patients bear some of the costs of AI, whether that is privacy risks, added expenses, or simply discomfort with algorithmic involvement, yet they are not the ones deciding whether AI is used. That asymmetry matters.

In the paper, we model that cost explicitly, but not patient agency over the AI-use decision itself. We do think greater patient voice could change the dynamics, especially if patients are informed enough to ask when AI is being used, why it is being used, and whether the clinician is following it. At the same time, we’d be cautious about shifting too much of this burden onto patients. They should have a voice, but they should not have to police the incentive system.

Simply put: Patients should not be passive recipients of AI-mediated care, but neither should they be responsible for fixing the incentive problems built into the system.

Source: Tinglong Dai and Shubhranshu Singh (2025), “Artificial Intelligence on Call: The Physician’s Decision of Whether to Use AI in Clinical Practice,” Journal of Marketing Research, 62 (5), 854–75. doi:10.1177/00222437251332898

Go to the Journal of Marketing Research