A review paper on large language models in healthcare lists the opportunities — diagnostic precision, patient engagement, clinical documentation, medical research, tailored treatment planning — and then lists the problems: privacy, data security, algorithmic bias, explainability, misinformation, accountability, and reliability across different patient groups.

Both lists are correct. What’s worth examining is why the second list is so much harder to act on than the first, and which item on it actually decides whether any of this is safe.

The opportunity list is the easy part

The case for language models in medicine is genuinely strong in one specific place: text. Healthcare produces staggering volumes of unstructured writing — clinical notes, discharge summaries, referral letters, prior authorisations, research literature no clinician has time to read. Summarising, searching and drafting that material is precisely what these systems do well.

Clinical documentation is the clearest win, and not a marginal one. Documentation burden is a leading driver of clinician burnout, consuming hours that could go to patients. Reducing it is valuable and comparatively low-risk, because a clinician reviews the output before it becomes a decision.

Diagnostic precision and treatment planning are a different proposition. Here the model isn’t organising information a clinician already has — it’s influencing a judgment. The risk profile changes completely, and so should the standard of evidence.

The bias problem is not the one people expect

The review’s concern about reliability “for different patient groups” is the item that deserves the most attention, because it’s structural rather than incidental.

Medical AI learns from medical data, and medical data encodes the history of who received good care. If a condition has been historically underdiagnosed in women, the training data contains fewer diagnosed women. If a population had less access to specialists, their records are thinner. A model trained on that corpus doesn’t just inherit the disparity — it can launder it, converting a historical inequity into an algorithmic output that carries the authority of a computed result.

This is harder than a data-quality bug because the bias is not an error in the data. It is a faithful record of what happened. Fixing it requires deciding what should have happened, which is a clinical and ethical judgment rather than an engineering one.

It’s also why the review’s call for “ongoing monitoring of performance for different patient groups” is the most important sentence in it. Not one-time validation — continuous, disaggregated monitoring. A model can perform well in aggregate while failing a subgroup badly, and an aggregate accuracy figure will never show it.

Explainability is where the real tension sits

The demand that medical AI explain itself is reasonable and, in current systems, largely unmet.

A clinician acting on a recommendation needs to know why, for several reasons at once: to exercise professional judgment about whether the reasoning applies to this patient, to catch the model’s errors, to explain the decision to the patient, and to be accountable for it afterwards. “The system said so” satisfies none of those.

The uncomfortable part is that language models can produce fluent explanations that are reconstructions rather than accounts of their actual processing. An explanation that sounds medically reasonable but doesn’t describe what the system did may be worse than no explanation, because it invites trust it hasn’t earned. Plausible-sounding justification is the failure mode that most efficiently defeats human oversight.

Accountability is the unresolved one

Privacy and security have known, if difficult, technical answers: encryption, access control, de-identification, governance. They’re hard engineering problems with established practice.

Accountability doesn’t have an equivalent. When an AI-influenced clinical decision harms a patient, responsibility is genuinely unsettled — between the clinician who accepted the recommendation, the institution that deployed the tool, and the developer who built it. Existing medical liability assumes a human decision-maker; existing product liability assumes a device that doesn’t learn. A system that is neither sits in the gap.

That gap has a practical consequence today. Clinicians are being asked to use tools whose recommendations they cannot fully audit while retaining full responsibility for the outcome. That’s an unstable arrangement, and it will get resolved — by courts and regulators rather than by developers.

What this means if you’re a patient

Two things are worth knowing without alarm.

First, these systems are already in use — most heavily in the administrative and documentation layer, which is where they’re least risky and most useful. If a summary of your visit was drafted with AI assistance and reviewed by your clinician, that’s a reasonable use of the technology.

Second, you are entitled to ask. If a recommendation about your care was influenced by an algorithmic tool, asking your clinician what informed the decision is a legitimate question, not an awkward one. Clinician oversight is the safety mechanism that all of this currently depends on — and it only works if the clinician is genuinely evaluating the output rather than deferring to it.

The read

The review’s framing — real opportunities, serious ethical challenges — is accurate but symmetrical in a way the situation isn’t. The opportunities are largely available now, concentrated in documentation and information retrieval, and mostly low-risk. The challenges are not evenly distributed either: privacy and security are hard but tractable, while subgroup reliability, genuine explainability and accountability remain substantially unsolved.

The sensible position is neither rejection nor enthusiasm but sequencing. Deploy aggressively where a human reviews every output and the failure mode is a bad draft. Deploy cautiously, with disaggregated monitoring and clear liability, where the failure mode is a bad diagnosis. The distinction between those two categories is the most important governance decision in medical AI, and it is the one most often blurred by describing everything as “AI in healthcare.”

Commentary on a published academic review of large language models in healthcare, as indexed on 22 July 2026. The source is a review paper; specific claims about clinical performance would require the underlying primary studies. General information, not medical advice.