The story goes that German hospitals run patients’ medical reports through machine translation while cultural mediators catch the errors. Checking it, the project isn’t there. What is there: no enforceable right to professional interpreting in German healthcare, a cost estimate that swings by a factor of twenty, and a major welfare organisation saying AI isn’t ready for this job.
The premise was supposed to be a project: German hospitals feeding discharge letters and medical reports through machine translation for patients who don’t speak German, cultural mediators catching the mistakes about dosages, regional states rolling the tool out onto the wards. It is a well-shaped story, with a clean mechanism — the machine produces, the human checks. The trouble is that when you look for it, it isn’t there. There is no documented national or state-level programme of automated translation of Arztbriefe for migrant patients; no German Land is confirmed to have formally introduced translation software into its clinics; and there are no verifiable German studies quantifying error rates on treatment instructions produced by those apps.
What does exist, and is well documented, is a legal vacuum. In Germany there is no enforceable right (Rechtsanspruch) to professional translation in healthcare. Many clinics offer care in a patient’s own language on a voluntary basis — meaning, when somebody on shift happens to speak it. This is where artificial intelligence genuinely enters the story: not as a deliberate procurement decision, but as filler. Neural machine translation systems — machine learning models trained on vast parallel text corpora, producing a plausible string in the target language by computing, token by token, the most probable continuation — are on every phone, free, in thirty seconds. Nobody has to approve anything for them to be used. That is precisely what makes the topic worth writing about, and what makes the “pilot project” framing misleading.
Adopting a tool is not the same as letting one in
A project has an owner, a specification, a validation procedure, an inbox where complaints land. Filler has none of that. When a nurse facing a patient who can’t follow the dosage instructions opens a translation app on a personal phone, that isn’t policy execution: it’s improvising around a problem the legal system dropped on them without tools. Nobody owns that translation. There is no vendor under contract, no log, no quality control, not even a record that a translation happened at all.
This is why framing the issue as “hospitals introduced AI and mediators correct it” puts the problem in the wrong place. It assumes a chain of control exists. If there is no chain, the question isn’t how often the model is wrong. It’s who would notice, and what follows if nobody does.
What German organisations are actually asking for
The active front in Germany is legislative, not technological. There is a Bündnis für Sprachmittlung — an alliance for language mediation — whose members include Diakonie Deutschland, the social welfare arm of the Protestant Church and one of the country’s largest non-state providers of social services. Its demand is specific: anchor the right to language mediation in SGB V, Book Five of the German Social Code, which governs statutory health insurance, covering every service in the insurance catalogue; and extend it to asylum seekers via the AsylbLG, the Asylum Seekers’ Benefits Act. The position paper was published around a year before January 2025.
In parliament, in November 2024 the SPD and the Greens tabled an amendment to the Gesundheitsversorgungsstärkungsgesetz, the healthcare provision strengthening act, providing for language mediation — but not as an individually enforceable right. That distinction is not cosmetic. An individual right can be asserted before a court; a general provision cannot. On the available evidence, that remains a tabled amendment, not a law whose operative text can be quoted.
The most revealing number: €12.5m to €255m
The estimated cost of comprehensive language mediation across Germany ranges between €12.5 million and €255 million a year. That is a factor of twenty. It isn’t an estimate; it’s two incompatible pictures of the world, depending on how many interventions you assume need covering and how much you assume the people providing them should be paid.
That spread explains the pull towards automation better than any technological enthusiasm. A translation model has a marginal cost approaching zero: once trained, the millionth sentence costs what the first one did. A professional interpreter costs every single time. In a system where nobody is obliged to pay for the interpreter and the budget is a twenty-fold range, the machine doesn’t win because it’s better. It wins because it’s the only thing anyone can use without asking permission.
The verdict from the people doing the work
Maike Grube of Diakonie Deutschland has stated that in healthcare AI is “currently not usable” for sensitive communication, contextualisation and non-verbal cues, and that video and telephone interpreting will remain central. That judgement is worth taking seriously on technical grounds, not just political ones. A translation model works on text: it cannot see that the patient has stopped nodding, doesn’t know that in this family a diagnosis goes to the son first, can’t tell the “yes” that means I understood from the “yes” that means I’m embarrassed to ask again. And above all it always produces an output. It has no mechanism for saying “I didn’t understand that sentence.”
Grube also supplies a figure that upends a common assumption: in Switzerland, the twenty most frequent languages account for over 90% of mediation assignments. In practical terms, the language problem in European hospitals is not the long tail of a thousand world languages — the favourite argument of those who claim only automation can scale. It’s a narrow, predictable set, the kind you can staff with people.
What happens when there is no communication at all
A University of Hamburg study from 2017, cited in the German public debate, estimates that for 5% of patients in Berlin communication is absent or nearly so, and puts at roughly 34,000 the treatments and procedures performed annually in the city without adequate informed consent. It’s a second-hand figure, nearly a decade old, that ought to be re-checked against the original study. But it indicates the order of magnitude: this is not a communication inconvenience, it is a defect in the legal basis of the medical act itself.
Informed consent is exactly where machine translation becomes legally delicate. Consent is valid if the patient understood. An undocumented machine translation, with no professional attesting that the content was conveyed, produces no evidence of understanding. It produces a text nobody answers for.
The model that works isn’t a model
The one documented local scheme found here is in the Vogelsbergkreis, a rural district in the state of Hesse, which employs Sprachmittler — specially trained language mediators, used in healthcare settings among others. One of them, Nazila Afshar, is named in reporting on the scheme. It isn’t software; it’s trained people. And the obstacle they intercept, in the accounts available, is not a machine error: it’s the medical jargon of a flesh-and-blood doctor — the original problem, which no technology created and none has yet solved.
There is an open lead, and it’s honest to say it’s open. Springer’s journal HNO has published a paper titled “Webbasierte Übersetzungssoftware in der klinischen Praxis” — web-based translation software in clinical practice. Title and DOI check out; the content sits behind a paywall and has not been read. That is probably where the numbers missing from this story live. Until then, anyone writing about machine translation on the ward should resist lending those numbers a weight nobody has yet measured in German.
The least intuitive thing about the whole affair is this: the German debate about AI in healthcare is not a debate about how good it is. It’s a debate about who pays. Until the right to language mediation sits in SGB V, any technical assessment of the model is secondary, because the real comparison isn’t machine translation versus a human interpreter. It’s machine translation versus nothing. And against nothing, any model wins.