AI is moving deeper into diagnosis, clinical decision-making and everyday hospital workflows. While recent studies show that it can improve efficiency and help doctors reach better answers, they also reveal errors, overreliance and gaps in clinical reasoning. As adoption accelerates, the bigger question is becoming less about whether healthcare should use AI and more about who remains responsible when a decision affects a patient.
Artificial intelligence is steadily becoming part of the healthcare system, and its role is no longer limited to experimental pilots or back-office automation.
Doctors are using AI to summarise medical records, search research, prepare clinical documentation and interpret patient information. Hospitals are deploying algorithms in medical imaging, risk prediction and workflow management. Patients are also entering consultations after using chatbots to understand symptoms, medicines and laboratory reports.
In India, the technology is already operating at public-health scale. The government’s eSanjeevani telemedicine platform has integrated an AI-based Clinical Decision Support System that helps organise patient complaints and provides doctors with possible differential diagnoses. Government figures show that 12 million consultations have been supported by AI-recommended diagnoses.
These developments make one thing increasingly clear: AI is unlikely to disappear from healthcare.
But its expanding role is creating a more complicated question. How far should the technology be allowed to go?
There is a considerable difference between an AI system summarising a doctor’s notes and one recommending whether a patient could have cancer, needs another investigation or should receive a particular treatment. The closer AI moves towards decisions that directly affect patients, the greater the consequences of getting something wrong.
Recent research increasingly points towards a middle ground. AI can strengthen clinical work, but stronger technology does not necessarily make human judgement redundant. In some situations, the most effective model may not be doctor versus machine at all. It may be a doctor working with AI while retaining responsibility for the final clinical call.
The pace of adoption suggests that healthcare professionals themselves largely recognise the opportunity.
The American Medical Association’s 2026 Physician AI Survey, conducted among 1,692 physicians across specialties, practice settings and career stages, found that more than 80% were already using AI professionally. The average number of AI use cases among doctors had increased from 1.1 in 2023 to 2.3 in 2026.
More than three-quarters believed AI could provide an advantage in patient care, particularly in areas such as diagnostic support and efficiency.
Yet the same doctors were not asking for unrestricted automation.
Around 88% considered safety and efficacy validation important for wider AI adoption, while 86% placed importance on assurances around data privacy. Another 85% wanted doctors to be consulted or responsible for decisions about adopting AI in their practices.
Perhaps more significantly, 92% wanted greater education and training around AI, while nearly nine in ten expressed some concern that greater dependence on the technology could contribute to a loss of clinical skills.
That tension captures where healthcare currently stands.
Doctors increasingly see AI as useful, but usefulness and autonomy are not the same thing.
“AI has enormous potential in healthcare, but it cannot replace physician judgment,” AMA CEO John Whyte, MD, MPH, said as the association adopted policy calling for AI used in patient care to remain assistive rather than become an autonomous substitute for clinical judgement.
The concern is not based on the assumption that doctors will always outperform algorithms. In certain narrow and clearly defined medical tasks, AI systems have already demonstrated impressive capabilities.
The more difficult question is what happens when AI leaves the controlled environment of a benchmark and enters a consultation where information may be incomplete, symptoms can be ambiguous and the correct decision depends on context.
When AI makes the doctor better
One of the clearest indications of what human-AI collaboration could look like came from a randomised controlled trial published in npj Digital Medicine in March 2026.
Researchers studied 70 US-licensed physicians, almost all working in internal medicine, and examined diagnostic performance under different conditions.
Doctors using conventional resources received an average diagnostic performance score of 75%. When physicians worked with AI providing a first opinion, the score increased to 85%. Doctors who developed their initial assessment before receiving an AI second opinion scored 82%.
On the surface, the finding makes a strong case for AI-assisted medicine. Giving doctors access to another source of clinical reasoning helped improve overall performance.
But the experiment also revealed why the relationship is more complicated than simply adding AI to every decision.
When AI was introduced as a second opinion, doctors improved their clinically actionable scores in 52 cases. In another 12 cases, however, their scores declined. Overall, actionable performance became worse after AI involvement in about 8% of cases.
Researchers also observed signs of anchoring, where suggestions from one side could influence the reasoning of the other. Doctors could be pulled towards an AI recommendation, while AI responses could themselves be affected by a clinician’s earlier conclusion.
The study used structured clinical scenarios rather than real patients, so its findings cannot establish how frequently such behaviour would occur inside a hospital. It does, however, illustrate an important problem for healthcare AI: accuracy alone does not determine whether a system improves care.
How and when AI is introduced into a clinician’s workflow can influence the result.
Jonathan H. Chen, one of the researchers involved in the study, reduced the challenge to a straightforward question: “What is a computer good at? What is a human good at?”
That distinction becomes even clearer when AI is asked to perform different stages of clinical reasoning.
A separate 2026 study published in JAMA Network Open evaluated 21 large language models across 29 standardised clinical cases, generating more than 16,000 responses. Researchers tested the models on tasks including developing differential diagnoses, selecting investigations, reaching a final diagnosis and recommending management.
Performance varied considerably depending on what the models were being asked to do.
The systems performed relatively well when sufficient information had already been assembled and they were asked to identify a final diagnosis. They struggled more with differential diagnosis, the earlier and often more uncertain stage of medicine where a clinician must consider several possible explanations rather than settle immediately on one answer.
Under the study’s scoring method, failure rates for differential diagnosis exceeded 80% across the models, compared with below 40% for final diagnosis.
The researchers acknowledged that their evaluation was particularly strict when scoring differential diagnoses. Even so, they concluded that current off-the-shelf large language models were not ready for unsupervised, patient-facing clinical decision-making.
That gap is important.
A system may be good at recognising the likely answer once a case has been clearly described. Real patients do not always arrive with neatly organised cases. A doctor may have to decide which symptom matters, which possibility cannot safely be ignored, whether another test is necessary and how a patient’s medical history changes the interpretation.
Clinical medicine involves uncertainty, not simply answer generation.
India could be one of AI healthcare’s biggest tests
The debate becomes particularly relevant in India, where AI is being considered not only as a hospital technology but as a tool for expanding healthcare capacity.
Philips’ Future Health Index 2026 surveyed 200 healthcare professionals and 2,004 patients in India. Among healthcare professionals, 71% said AI had increased their capacity to see more patients.
For those reporting an increase, the median gain was 10 additional patients per week.
Around 82% said AI had improved workflow efficiency, while 85% believed the technology could improve patient outcomes.
Those figures help explain the attraction of AI in a healthcare system managing large patient volumes and uneven access to specialists. Technology that can organise information, reduce documentation work or help clinicians identify possible conditions faster could allow medical professionals to spend their limited time more effectively.
But the same survey also showed where healthcare professionals were drawing the line.
Some 86% said AI outputs required human oversight, while 78% reported having already needed to correct AI-generated misinformation. Forty-five per cent said AI training at their organisations remained limited or inconsistent.
The findings come from an industry-commissioned survey rather than an independent clinical trial, but the contrast is notable. Healthcare professionals can simultaneously believe that AI improves their work and believe that its output should not be accepted without scrutiny.
India’s policy approach is beginning to reflect the same distinction.
In February 2026, the government launched the Strategy for Artificial Intelligence in Healthcare for India, or SAHI, alongside the Benchmarking Open Data Platform for Health AI, known as BODH.
The framework is intended to support evidence-based healthcare AI adoption, while BODH, developed with IIT Kanpur, provides a mechanism for evaluating systems on areas such as performance, robustness, bias and generalisability before wider deployment.
Rather than treating healthcare AI simply as another software purchase, the direction places greater emphasis on whether a model has been properly tested for the environment in which it will operate.
Speaking at the World Health Assembly in May, Union Health Minister J P Nadda said the “future of AI in healthcare will be defined by collective human choices, not algorithms alone.”
That principle becomes particularly important as AI reaches smaller hospitals, telemedicine services and high-volume healthcare environments.
An AI system might help a doctor identify possible diagnoses across hundreds of remote consultations. It might highlight an abnormal scan that requires urgent review. It could alert clinicians to a combination of symptoms they had not initially considered.
But generating a recommendation is different from carrying responsibility for what happens after it.
Human oversight cannot become a rubber stamp
Simply requiring a doctor to approve an AI recommendation may not be enough.
If clinicians are repeatedly presented with AI-generated answers during busy shifts, there is a risk that human oversight becomes procedural rather than meaningful. A doctor who accepts hundreds of recommendations that appear correct may gradually become less likely to challenge the next one.
This is where the debate moves beyond whether there is technically a “human in the loop”.
The human must have enough information, training and authority to disagree with the machine.
The AMA survey highlights this challenge. Alongside growing AI adoption, 92% of physicians wanted more education and training, while nearly nine in ten expressed concern about potential loss of clinical skills.
Internationally, deployment also appears to be moving faster than governance.
A 2026 WHO/Europe assessment covering its 53 Member States found that nearly two-thirds were already using AI in diagnostics and half had introduced AI-powered patient chatbots.
Yet only 8% had a health-specific AI strategy. Just one in five provided AI education to health professionals before qualification, one in four offered workforce training and almost 40% had no ethical guidance specifically covering healthcare AI.
“The longer governance lags behind deployment, the higher the human cost,” WHO Regional Director for Europe Hans Henri P. Kluge warned.
The numbers point to a problem that model developers alone cannot solve.
Healthcare AI safety will also depend on the organisations deploying it. Hospitals need processes for evaluating tools, monitoring mistakes and identifying whether performance changes across different patient groups. Doctors need to understand what a system can and cannot reliably do. Patients may need to know when AI has materially influenced their care.
The level of oversight is also unlikely to be identical for every application.
An AI tool drafting a discharge summary presents a different level of clinical risk from one recommending whether a suspicious lesion is malignant. A scheduling algorithm does not require the same safeguards as a system influencing medication or surgery.
The closer AI gets to consequential clinical decisions, the stronger the case becomes for qualified human accountability.
That does not mean every medical decision belongs exclusively to doctors. Healthcare is delivered by teams that include nurses, pharmacists, technicians and other licensed professionals with different responsibilities. The broader principle is that consequential clinical decisions should remain accountable to appropriately qualified healthcare professionals, with physician-led decisions remaining subject to physician judgement.
This may ultimately prove more realistic than trying to define healthcare’s future as a contest between doctors and algorithms.
AI brings advantages that clinicians cannot easily replicate. It can process enormous quantities of information, search for patterns, retrieve knowledge rapidly and work consistently across repetitive tasks. Doctors bring something different: an understanding of the patient in front of them, the ability to interpret uncertainty and competing risks, and professional accountability for the consequences of a decision.
The evidence emerging in 2026 suggests that those strengths can complement each other.
AI is already making its way into consultations, diagnostic workflows and public-health systems. As the technology improves, it will likely become harder to imagine modern healthcare operating without some form of machine assistance.
But better AI does not automatically mean autonomous healthcare.
For now, the more consequential shift may be towards a system where AI helps gather evidence, identifies patterns, raises possibilities and challenges assumptions, while trained healthcare professionals decide how that information should translate into patient care.
AI appears set to stay in the exam room.
The final clinical call, however, is likely to remain human.
Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.