Don't redact it, don't send it

Every system that sends health data to a foreign model has a slide about de-identification. Names stripped, phone numbers masked, identifiers removed before the data leaves. It is real work and it is worth doing, and it is not the same thing as safety.

A woman describing her symptoms in a voice note is not a row with a name field to blank out. Her voice is identifying. Her dialect places her. The details of her complaint, the village clinic she mentions, the relative who brought her, reassemble into a person no matter how many fields are masked. Free text and recorded speech are the hardest things to de-identify and the easiest to re-identify, and the most sensitive health disclosures arrive in exactly that form.

The Personal Data Protection Ordinance of 2025 treats health data as a specially protected category and points toward keeping it in the country. Read honestly, that is not a rule that better redaction satisfies. It is a rule about where the data goes.

So the honest fix is the blunt one. The most sensitive flows, the recorded voice, the free-text symptom story, the stigmatized consultation, do not get scrubbed harder and shipped. They get processed in the country, on infrastructure under the country's law, by a model that does not phone abroad. De-identification stays the floor for what must briefly leave. But the milestone KhaM-Med exists to reach is the one where the sensitive data does not leave at all, because the only data you can be certain was not mishandled abroad is the data that never went.