AI-generated discharge summaries could save junior doctors significant time — but only if the review process is genuinely clinical, not merely administrative. Digital Health reported in March 2026 that doctors had raised safety concerns about an AI discharge-summary tool, including calls for more clinical depth, regulatory assurance, and evidence of safety before further deployment. The report noted ongoing discussions with the MHRA about medical-device classification and the risk that over-reliance could allow mistakes to slip through.
These concerns are serious — because discharge summaries are not low-risk documents.
Why Discharge Summaries Are High-Risk Documents
A discharge summary is the primary clinical communication between hospital and GP. It must accurately convey: the reason for admission and final diagnosis (which may differ from the initial impression), all medication changes (new medications started, existing medications stopped, doses changed — and the rationale for each change), pending investigation results (blood tests, imaging, biopsies — with clear responsibility for who will follow up), diagnostic uncertainty (where the diagnosis remains unclear, this must be communicated explicitly rather than presented with false confidence), follow-up plans (who is responsible for what, by when), safety-netting advice (what should prompt the patient or GP to seek urgent re-assessment), and coding (which affects disease registers, future prescribing decisions, and insurance implications).
An error in any of these elements can cause real patient harm. A missed medication change can lead to a patient continuing a drug that was stopped for a clinical reason. A pending result that is not communicated may never be followed up. A false diagnostic certainty may lead the GP to manage the patient as if a definitive diagnosis has been reached when it has not.
What the Safety Concerns Tell Us
The concerns reported by Digital Health reflect a specific and important pattern: AI-generated clinical documentation can appear polished, comprehensive, and professionally formatted while being clinically incomplete, diagnostically overconfident, or factually incorrect in ways that a time-pressed junior doctor may not catch during rapid review.
The specific risks include: hallucinated clinical details (the AI generating examination findings or management decisions that did not occur), diagnostic overcertainty (converting a working diagnosis into a definitive one), medication inaccuracy (wrong drug names, doses, or omission of changes), missing pending results (the AI not capturing investigations ordered but not yet reported), generic safety-netting (producing boilerplate return-if-worse advice rather than presentation-specific red flags), and formatting polish that masks content gaps (a beautifully structured document that omits critical clinical information).
What Good Review Should Include
"Clinician reviews the output" must mean more than opening the document and clicking approve. Good review should include: comparing the AI draft against the source clinical record (not reviewing the AI output in isolation), verifying the diagnosis and all medication changes against the drug chart and prescribing record, checking pending investigations and confirming that responsibility for follow-up is clearly assigned, removing any hallucinated clinical details that the AI may have generated, ensuring diagnostic uncertainty is preserved where it genuinely exists, verifying that safety-netting advice is specific to the clinical presentation, and confirming that the document is understandable to the receiving GP and to the patient.
This review takes time. If the deployment model does not allocate sufficient time for genuine clinical review — if the expectation is that AI discharge summaries will save time by reducing review to a rubber-stamp — the safety concerns raised by doctors in March 2026 are likely to materialise.
The MHRA Classification Question
The discussions with the MHRA about medical-device classification are important. A tool that generates clinical documentation from patient records — extracting diagnoses, medication changes, and management plans — is arguably providing clinical information that influences subsequent clinical decisions. If the intended use meets the MHRA's definition of a medical device, the regulatory requirements (clinical safety case, post-market surveillance, conformity assessment) apply.
The regulatory classification question is not academic. It determines what governance, safety assurance, and oversight the tool requires before and after deployment.
Where iatroX's Trust Principles Apply
This is where iatroX's trust principles are relevant beyond search. Clinical AI should not rely on polish. It should support verification: source fidelity, provenance, uncertainty display, fail-safe behaviour, and feedback mechanisms. Those principles apply whether the output is a clinical answer, a consultation note, a referral letter, or a discharge summary.