Key takeaways
- English accounts for close to half of all websites with an identifiable content language, and large language models inherit that imbalance.
- Controlled studies show measurable accuracy loss when clinical questions are asked in non-English languages, including sharply higher hallucination rates.
- Chronic autoimmune conditions expose the gap fastest, because they generate years of daily practical questions rather than a single diagnostic query.
- Independent native-language publishers are quietly becoming part of the retrieval layer that AI assistants draw on, which raises the bar on editorial standards for anyone entering the space.
Ask an AI assistant about a thyroid condition in English and you get a fluent, well-sourced answer in under three seconds. Ask the same assistant the same question in Serbian, Slovak or Lithuanian and you still get a fluent answer in under three seconds. That is the part worth paying attention to. The confidence stays constant while the evidence underneath it does not.
For health tech companies, insurers and publishers operating outside the Anglosphere, this is not a translation problem to be solved with a localization vendor. It is a supply problem in the underlying corpus, and it is creating an unusual opening for small, focused publishers who work in languages the major health platforms have never bothered to serve.
The corpus behind a health question is not the same in every language
The raw distribution of web content is lopsided in a way that most product teams underestimate. W3Techs reported on 5 June 2026 that English was the content language of 49.7 percent of websites it could identify, with Spanish and German next at 6.0 percent each, followed by Japanese at 5.0 percent and French at 4.6 percent. Serbian, Croatian, Bulgarian and their neighbours sit far below all of these, each accounting for well under one percent of indexed sites.
Language models are trained on roughly that distribution, and researchers building multilingual medical systems have been blunt about the consequences. Work on multilingual medical question answering identifies two compounding problems: mainstream models are trained predominantly on English-centric data, which limits their effectiveness in other languages, and high-quality external medical data for low-resource languages is extremely scarce. The common workaround, translating inputs into English for inference or translating English corpora into the target language, carries real cost and introduces medical semantic distortion.
In practice, a model answering a Serbian health question is doing something closer to translation plus inference than retrieval from a deep native corpus. It usually works. When it does not, there is no local source to catch the error.
The accuracy drop is measurable, not theoretical
The clearest evidence comes from studies that hold the clinical content constant and vary only the language.
A psychiatric evaluation of an open-source model tested identical interview notes in Korean and in English translation. Hallucinations appeared in 30.2 percent of outputs generated from Korean input compared with 13.4 percent from English, and top-1 diagnostic accuracy fell to 59 percent in Korean against 74.5 percent in English. Across 115 medical licensing examination questions, the same model scored 46.1 percent in English and 32.2 percent in Korean.
Arabic medical evaluation work points at the same pattern with an added warning about user behaviour. Researchers noted that the effect appears across languages but is particularly problematic in low-resource settings, where a lower baseline accuracy increases the risk of users over-trusting incorrect outputs, and cautioned that model-generated explanations should not be treated as a reliable remedy.
The picture is not uniformly grim, and it is worth resisting the easy conclusion. A large benchmarking study run with community health workers across four Rwandan districts found the opposite framing. Although model performance degraded slightly in Kinyarwanda, the models still outperformed local clinicians and cost over 500 times less per response. The honest reading is that AI health answers in smaller languages are simultaneously better than the alternative in some markets and less reliable than the English version of themselves. Both things are true, and both matter to anyone building in this space.
Why chronic conditions expose the gap fastest
Acute care questions tend to be one-shot. Chronic autoimmune conditions are the opposite, and they surface the language gap faster than almost anything else in consumer health.
Hashimoto’s thyroiditis is a useful case. It is the most common cause of hypothyroidism in regions with adequate dietary iodine intake, a meta-analysis across multiple countries put global prevalence at 7.5 percent and 11.4 percent in low- and middle-income areas, and prevalence in women runs about four times that in men. Its symptom profile is broad and slow-moving, covering fatigue, weight changes, cold intolerance, hair loss, cognitive changes and menstrual irregularities, which means patients do not ask one question. They ask a few hundred over several years, most of them practical rather than diagnostic.
That long tail is exactly where localized content either exists or does not. An English speaker researching daily management has decades of patient advocacy publishing, clinician blogs and structured medical content to draw on. A Serbian speaker typing hasimoto sindrom into a search bar is asking a clinically identical question against a far thinner pool: some clinic pages, scattered forum threads, and a small number of independent patient-run resources such as Living with Hashimoto, which publishes Serbian-language guidance and recipes for people managing the condition day to day.
That asymmetry is what the model is working with when it answers.
What the same query actually returns, by language
| English query | Major EU language | Smaller national language | |
| Depth of indexed native sources | Deep, decades of publishing | Moderate, mostly clinical | Thin, fragmented |
| Clinician-reviewed consumer content | Widely available | Partially available | Rare |
| Daily-management and dietary content | Extensive | Limited | Mostly forums and social |
| Typical model behaviour | Native retrieval | Mixed retrieval and translation | Translation plus inference |
| Where errors get caught | Multiple authoritative sources | Some | Usually nowhere |
The opening, and the obligation attached to it
The strategic point for operators is that retrieval-augmented systems and AI search surfaces do not invent authority. They select it. When a query arrives in a language where three credible sources exist, being one of those three is worth more than being the four-hundredth source in English.
That is a genuine market opening for niche publishers, patient advocacy organizations and health tech firms with regional footprints. It also comes with an obligation that the English-language content marketing playbook tends to skip. Content that AI systems cite in a health context needs identifiable authorship, visible credentials or lived-experience disclosure, explicit medical disclaimers, dated reviews, and structured markup that machines can parse. Publishing thin translated content into an underserved language does not fill the gap. It pollutes it, and in a corpus this small, one bad source carries disproportionate weight.
The companies that treat native-language health publishing as infrastructure rather than as a marketing channel are the ones likely to end up in the answer.
FAQ
Do AI assistants give worse health answers in less common languages? Often yes. Studies holding clinical content constant while varying language have found higher hallucination rates and lower diagnostic accuracy in the non-English version, though the size of the drop varies by model and language.
Why does this happen if the model speaks the language fluently? Fluency and knowledge are separate. A model can generate grammatical Serbian or Korean while drawing on a training and retrieval corpus that is overwhelmingly English, so the language layer is strong while the evidence layer behind it is thin.
Should patients outside English-speaking countries avoid AI health tools? Not necessarily. Benchmarking in low-resource health systems has found AI responses outperforming locally available clinical advice at a fraction of the cost. The practical guidance is to treat output as a starting point for a conversation with a clinician rather than as a substitute for one.
What can publishers do about it? Publish original, structured, clearly attributed content in the target language rather than machine translations of English pages, keep review dates and disclaimers visible, and mark up content so retrieval systems can identify what it is and who stands behind it.
Is this a temporary problem that better models will solve? Partly. Model multilingual capability keeps improving, but the underlying scarcity of high-quality native-language health content is a publishing problem, not a modelling one, and it will not resolve until someone writes the material.
Jana Radojcic is Founder and CEO of Adnen Enterprises, Director of Organic Growth at Alpha Market Flow and co-founder of GetStuffDigital, where she works on search and generative engine optimization for fintech and SaaS brands. She writes about authority building, digital PR and the shifting mechanics of discovery in AI-mediated search.
