The gap this article addresses is between two documents about the same medicine: the SmPC, written for professionals in the regulatory register, and the explanation a patient actually needs, in their language, at their health literacy, about their situation. AI is genuinely good at the transformation in the middle, register conversion is what language models do, and the risks are equally specific: fluency that shades into false reassurance, simplification that deletes material risk, and readability claims that outrun reality, the research on AI-generated patient information finding that specialist evidence tools can improve citation quality while readability frequently remains too technical. The method below keeps the transformation's benefits and staffs its risks.
What must survive the translation
Eight elements a patient explanation must carry across from the professional sources, whatever else gets simplified. The expected benefit, honestly sized: what this medicine is for and how much difference it typically makes, absolute framing where numbers appear, because relative-risk language misleads in both directions. Time to benefit: when to expect effect, the absence of which drives silent discontinuation. Common adverse effects: the frequent and manageable, framed with what to do. Serious adverse effects: the rare and urgent, framed with when to seek help immediately, the element false reassurance deletes first. Monitoring: what will be checked and why the checks matter, recruiting the patient into the plan per /blog/ai-medication-monitoring-plans-anp. Missed doses and practicalities: the operational questions patients actually have. Pregnancy and contraception implications where they exist, at the seriousness the product's requirements demand. And uncertainty itself: what is not known, said plainly, because preserved uncertainty is what separates information from persuasion, and the shared-decision standard requires it.
The workflow, and the checks that make it safe
Five steps. Identify the evidence and the approved documents first: the Ask-iatroX pass that surfaces the guideline position, the exact SmPC and the current PIL, sources the translation will be checked against, not decorated with. Draft with explicit constraints: plain language at a stated reading level, the eight elements required, absolute risk framing, uncertainty preserved, no reassurance beyond the evidence, and the output treated as a draft in exactly the way a scribe's note is a draft. Compare against the approved PIL and any regulator-issued risk-minimisation material: the step that catches deletion, because the PIL is the patient-facing document the licence actually produced, and material present there but absent from your draft is a finding, not a stylistic difference. Personalise clinically: this patient's renal function, interactions, priorities and concerns, the layer no template holds. And close with teach-back: the patient explains it back, key risks, what to do if, when to return, which is the only readability test that measures the actual reader, and five sentences of it convert counselling from broadcast into confirmation.
Language, literacy and the patients the defaults miss
The equity layer deserves explicit design rather than good intentions. Reading level: draft below where instinct suggests, then teach-back calibrates per person; the research's recurring finding is that AI output rates as too technical even when asked otherwise, so the constraint needs stating and the result needs checking. Language: translation of counselling content carries the modality and risk-language hazards this programme has mapped, graded recommendation and probability words flatten across languages, so translated materials get the same PIL comparison in the target language where one exists, and interpreter-mediated teach-back where it matters. Accessibility: format alternatives, large print, audio, easy-read structures, are generation-cheap now, which removes the excuse rather than the review step. And culture and context: examples, foods, routines and beliefs that make instructions followable are personalisation, not garnish, and the clinician holds them, which is the standing answer to where AI ends: it drafts the register shift, and the person who knows the patient makes it true.
Frequently asked questions
Should AI-drafted patient information be marked as such?
Follow local policy and lean transparent: the clinically reviewed and personalised version is the clinician's document in accountability terms, and openness about drafting tools costs nothing while trust is being built.
Can the same method produce clinic-wide leaflets?
With governance added: a leaflet leaving one consultation is counselling, a leaflet used across a service is controlled documentation, owned, dated, reviewed against PIL updates on a schedule, the recency discipline applied to your own materials.
What about consent conversations specifically?
The eight elements are the consent conversation's information half: material risks, benefits, alternatives and uncertainty, personalised and confirmed by teach-back, with the documentation recording what was actually discussed.
How long does the full workflow add to a consultation?
Front-loaded, minutes; in the room, it saves time: a reviewed draft plus teach-back is faster than improvised explanation followed by the call-back that confusion generates, and the draft is reusable for the next patient on the same medicine.
Should numbers always be included in patient explanations?
Where they exist and matter, in absolute form with a denominator patients can picture; where evidence gives no usable numbers, saying so plainly is the honest alternative to invented precision.