Language safety
Translation is a safety issue, not just a feature.
How Cancer Explained handles multilingual content, AI-assisted translation, bilingual feedback, disclaimers, and adversarial fact-checking.
The translation standard
A translation can be grammatically smooth and still unsafe if it shifts a medical meaning. That is why translation needs its own review path.
We use official translated sources wherever they exist. Where we translate or adapt something ourselves, the page says which it is: machine-assisted, checked by a bilingual human, checked by a qualified clinician, partly checked, or still awaiting review.
Every translated page links to the correction form, because bilingual readers catch what automated checks cannot — regional wording, false friends, a tone that lands wrong, or an ambiguity that matters for safety.
- Use official translated cancer sources first when they exist, especially NCI's Spanish cancer pages and MedlinePlus multilingual health pages.
- When AI assists with translation, label that plainly and keep the original source link visible.
- Never let a translation change diagnosis, treatment, side-effect, dosing, emergency, or eligibility meaning.
- Prefer simpler sentences over elegant wording when the medical meaning is safer.
Bilingual feedback loop
A public correction path lets bilingual readers help improve the site, especially for regional wording, tone, and safety-critical ambiguity.
Reports should be triaged by risk. A mistranslated emergency symptom, medication instruction, treatment warning, or screening eligibility statement should move faster than a style preference.
- Flag a translationReaders should be able to report a confusing phrase, unsafe wording, missing accent, regional mismatch, or machine-translation artifact.Open
- Track statusCorrection reports should be triaged by risk: emergency advice and treatment wording first; tone and style next.Open
- Credit reviewersBilingual clinicians, interpreters, advocates, and readers can be credited when they choose public recognition.Open
Adversarial AI checks
The useful idea is not to trust one AI. It is to ask independent systems to find unsupported claims, changed meaning, missing disclaimers, and unsafe next-step language.
A review pass like this raises questions; it does not decide anything. A person still has to judge whether a flagged issue is real, whether the source supports the wording, and whether the page needs editing, holding, or sending for medical review.
- Source-to-claim check: every medical statement must trace back to a cited source.
- Translation drift check: compare original and translated claims for changed medical meaning.
- Plain-language safety check: flag advice that sounds like diagnosis, treatment selection, or emergency triage.
- External-model challenge: run a skeptical review pass that looks for omissions, overclaims, and confusing language.