Disadvantages of AI in Healthcare: The Cost When It Works

Almost every page listing the disadvantages of AI in healthcare is listing risks instead. Bias, hallucination, privacy, opacity, job loss. Those are failures, and a failure can in principle be engineered away. A disadvantage is different. It is the bill that arrives while the system performs exactly to specification, and no better model retires it.

The distinction matters because the two categories are managed by different people with different budgets. If you want the failure side, sorted by who absorbs the harm, that is the risks of AI in healthcare, sorted by who carries them. What follows is the other column, and by 2026 there is enough measured deployment evidence to fill it without speculation.

The npj Digital Medicine framework that names seven recurring systemic failure modes in medical AI puts the idea in one line under its Misaligned Optimization heading: when proxies for success are poorly chosen, systems can perform exactly as designed and still fail their users. That framework was built by synthesis first and then tested against 914 stakeholders across 143 countries between July 2024 and March 2025, which found broad agreement on every one of the seven.

The benefit is real, and smaller than the business case

Start with the application that has the strongest evidence behind it and the least contested value: ambient documentation. If a disadvantage shows up here, it shows up everywhere.

A Dutch team ran the study the field had been missing, using continuous external observation rather than the electronic health record logs that most published evaluations rely on. Across 535 consultations with 12 general practitioners, the ambient scribe cut documentation time by 42.7 seconds per consultation, while total consultation time did not change at all. Patient throughput did not rise. Their own conclusion is that the tool is more effective at reducing cognitive burden than at improving operational efficiency, which is a defensible thing to buy and a different thing from what most procurement models assume.

The pattern repeats at scale. A retrospective study of 198,178 emergency department encounters at four hospitals found ambient AI scribes associated with a 1.6 minute reduction in median attending documentation time per note, against 3.3 minutes for a human scribe. Work relative value units per shift hour did not differ between the ambient AI group, the human scribe group and the encounters with no scribe at all.

And in a pediatric primary care network, 39 clinicians over 12 weeks used the scribe in 6,249 of 19,264 eligible encounters, with no significant change in documentation time, after-hours record time, total record time or visit closure rates. The reported benefits were real but self-perceived, and patient experience scores were identical between ambient and non-ambient visits.

Even the most favourable multi-site evaluation carries the caveat inside its own results. Across 120 clinicians at two academic medical centres the system usability score was 82, yet reduced burnout was reported by 55.0 percent and reduced frustration with the record system by 49.2 percent, and the authors state plainly that the underlying challenges were not eliminated.

None of this says the tools do not work. It says the measured gain is a fraction of the modelled one, and the gap between the two is the first disadvantage. Anyone building the case should read it next to what AI in healthcare actually costs, because the shortfall lands on the same spreadsheet.

The verification tax is permanent

The Dutch study also reports that the generated summaries required GP review in every single case and often missed non-verbal cues. That is not an error rate. It is a workflow step that exists for as long as the tool does.

How expensive that step is depends on the setting, and it can invert the whole proposition. A small Aalborg University study of two clinicians, a nurse and a psychologist, found documentation time rising from 7.5 to 21.5 minutes for the nurse once the scribe was introduced, with post-editing consuming the saving. Two participants is an anecdote with statistics attached and should be read as a warning rather than a finding. But it is a warning about the right thing: the review burden is not fixed, it scales with how badly the tool fits the encounter.

The safe use of general-purpose models adds its own tax. Clinicians are advised to ask the same clinical question two different ways to see whether the reasoning stays consistent, because phrasing alone moves the answer. That is sound practice and it is also a doubling of the interaction, on every query, forever.

The record itself changes shape

This one is almost invisible and compounds for years.

In the same Dutch cohort, documentation length increased during the intervention period, with more signs and plans recorded but fewer symptoms. Nothing malfunctioned. The note simply drifted toward what the model is good at capturing and away from what it is not.

The clinical record is the input to the next model, to audit, to research and to whoever sees the patient next. A systematic shift in what gets written down is not visible in any single note, does not appear in any time-and-motion evaluation, and is not reversible retrospectively. It is a structural cost paid by everyone downstream of the deployment.

The consultation is not the same room

The same study lists the drawbacks its authors observed alongside the time saving: inaccurate summaries, barriers to discussing sensitive information, and interference with the clinician's reasoning process. Neither patients nor GPs perceived an improvement in communication, and the authors flag reduced accessibility for patients with sensitive issues.

This is what the npj framework calls Dehumanization, and it is the purest example of a disadvantage rather than a risk. No incident report will ever be filed. The device did what it said. Something in the encounter got quieter, and the person who noticed was the patient who decided not to raise the thing they came in for.

The trust cost accrues whether or not the tool is accurate

Adoption is running ahead of consent, and the gap is measurable in three independent surveys.

In the European Patients Forum survey, the leading patient concerns were loss of human empathy and personal touch at 70 percent, biased AI-driven decisions at 68 percent and incorrect diagnoses at 59 percent, with 39 percent not confident in how AI is applied. Asked what they would do if AI and their doctor disagreed, 3 percent would trust the AI on its accuracy and 0.3 percent would trust it alone.

The direction of travel is the part that should worry anyone planning a five-year rollout. A Coalition for Health AI patient survey found 51 percent saying AI makes them trust healthcare less against 12 percent who say it increases trust, with 93 percent reporting at least one concern, and the concerns clustering on accountability and oversight rather than on capability. A poll of 1,007 US adults commissioned by Ohio State University's Wexner Medical Center found openness to AI in their own care had fallen to 42 percent from 52 percent in 2024. Among community health centre patients, 64 percent were uncomfortable with the general use of AI in their care and around 85 percent were concerned about AI accessing their personal health information.

Accuracy does not buy this back, because the objection is not about performance. It is about disclosure and accountability, which is the territory of the ethics of AI in healthcare rather than of model evaluation.

It is least available where it was supposed to help most

The strongest argument for clinical AI is reach: services in places without enough clinicians. The evidence says patients will not take it on those terms.

A preregistered conjoint study of 3,000 US adults found that the presence of a clinician raised the probability of choosing a given visit by 18.4 percent, and that every form of governance was preferred to none. The authors draw the conclusion themselves: strong patient preference for a clinician in the loop limits the potential of medical AI to expand service availability in exactly the low-resource settings, underserved populations and underserved specialties where trained clinicians are scarcest.

There is a route out, and it is not a communications strategy. In the same study, AI performance at or above specialist level moved patient choice more than any regulatory label did, roughly three times more than FDA approval. The constraint is current rather than permanent. It is also not one that today's deployments have cleared.

Validation never finishes, so the cost never stops

The seventh of the seven sins is Self-Referential Evaluation: judging a system on internal metrics and simulations without external audit or real-world testing. The framework's authors add the operational consequence, which is that patient and disease drift make one-time validation insufficient and require systematic, perpetual adjustment.

Underneath that sits a mechanism worth understanding before signing anything. Training labels frequently reflect historical clinician behaviour rather than objective truth, so a model can faithfully reproduce a prescribing habit that was never the best evidence-based choice, and score well doing it. Correcting for that is not a fix applied once. It is monitoring with no end date, which makes clinical AI an operating expense wearing the costume of a capital purchase. That is the practical substance of responsible AI in healthcare once the principles are set aside.

More correct alerts still cost attention

The last disadvantage is the one that predates AI and that AI makes worse by working.

A 2026 qualitative study of alert fatigue in hospitals found that raw alert volume was rarely identified as the cause on its own. What produced fatigue was clinically irrelevant, repetitive and poorly timed alerts, and the resulting mistrust generalised into broad suspicion of alerts as a category. Junior doctors learned dismissal by watching senior colleagues click through.

Map that onto a prediction system that surfaces more true signals. Each one may be correct. The attention available to act on any of them, including the pre-existing alerts that were already working, is finite and is now divided further. The npj framework files this under Overinforming and False Forecasting, and it is a cost that scales with how well the model performs.

What this changes about the decision

Look at the eight items again and note who owns each. Shortfall against the business case is finance. The verification tax is clinical operations. Record drift is informatics. The changed consultation is professional practice. Patient trust is communications and law. Reach is health policy. Perpetual revalidation is regulation. Alert load is safety.

Not one of them is answered by a better model, and no single department can write the against case alone. That is the reason the Health Tech track runs alongside Government, AI Ethics and Finance at the EX Future Summit rather than in its own room. The summit runs 18 to 20 November 2026 as a single continuous thirty-hour broadcast between Las Palmas and Bali, twelve hours apart, and online attendance is free for verified researchers and students. The adoption picture these costs attach to is set out in what the 2026 deployment evidence shows.

FAQ

What are the main disadvantages of AI in healthcare?

Ranked by how reliably they have been measured: a benefit smaller than the business case, a permanent review burden on the clinician, a systematic shift in what the medical record contains, a changed consultation, patient trust that falls as adoption rises, restricted reach into the low-resource settings that most need it, validation that never ends, and further division of clinical attention. Bias, hallucination and privacy breaches belong on the risk register instead, because they describe the system failing rather than succeeding.

What is the difference between the risks and the disadvantages of AI in healthcare?

A risk is a harm that occurs when the system fails. It is probabilistic, it can be engineered down, and better models reduce it. A disadvantage is a cost incurred while the system performs to specification, so improving the model does not touch it. Most published lists conflate the two, which is why they read as alarming and change nothing about a procurement decision.

Will AI replace doctors and nurses?

The productivity evidence does not currently support displacement. Across 198,178 emergency department encounters, work relative value units per shift hour did not differ between ambient AI, human scribe and no scribe, and in the general practice study consultation length was unchanged. Patients also actively price in the human, preferring a visit with a clinician present by 18.4 percentage points. The measured effect so far is on burden, not on headcount.

Do patients have to be told when AI is used in their care?

Legal duties vary by jurisdiction, but patient expectation does not. In the European Patients Forum survey, 93 percent said patients should be informed at the very beginning of a visit or treatment, and 82 percent wanted to hear it from their own clinician rather than from a website or a leaflet. Undisclosed use is a trust liability even where it is lawful.

Does AI in healthcare help underserved populations?

Less than the argument for it assumes, and the obstacle is patient preference rather than capability. The conjoint evidence shows a strong preference for a clinician in the loop, which the study's authors note limits AI's ability to expand services in precisely the settings where clinicians are scarcest. Performance at or above specialist level shifted patient choice more than any governance signal, so the barrier looks like a current condition rather than a permanent one.