Engineering note · Technical support engineering · by Ayman Sbeiti — I support high-trust software platforms · Hiring?
Engineering playbook · Investigation
Bilingual knowledge content and the AI that reads it
Core principle
A bilingual knowledge base isn't two translations of one truth. It drifts into two truths, and an AI assistant reading both will faithfully serve the drift.
Real-world example
A bilingual knowledge base sounds like one set of articles in two languages. In practice, the two versions age independently. An article gets corrected in one language and not the other, a new feature gets documented in whichever language its author wrote first, and terminology drifts as different people maintain different halves. None of this is anyone's failure. It's what happens when nobody owns the pair, only the individual articles.
Why it happens
Humans absorb this drift almost invisibly: a bilingual support agent reads both, notices the mismatch, and answers from the fresher version without ever filing the discrepancy. An AI assistant does the opposite. It answers from whatever it read, in whichever language the customer asked, with full confidence in both. The drift that human judgment had been quietly papering over becomes two different official answers.
What I now check
- Before connecting an AI assistant to any content set, ask not just "is this content good?" but "does this content agree with itself?"
- Reconcile the content pair by pair, and decide which version is canonical.
- Assign ownership so future changes touch both versions or neither.
Production takeaway
Knowledge-base cleanup, including bilingual reconciliation, belongs before the AI import, with clear content ownership. Not after.