Why Generic AI Translation Can't Enforce a Glossary
Pasting a glossary into a chat window and asking the model to "use these terms" is the most common terminology strategy in machine translation today. It is also the one that fails most predictably. The model complies for a few chunks, then drifts back to whatever rendering its training data prefers.
What goes wrong
The same character name appears as three different transliterations. A regulated term that must read identically in every clause appears as four near-synonyms. Brand-name capitalisation flips chapter by chapter. Each chunk is internally consistent. The whole document is not.
Why generic AI translation fails here
Two architectural reasons. First, a glossary placed in a system prompt is treated as background advice; local context can and does override it. Second, there is no verification step after translation, so drift is never caught, never repaired.
How TranslationAI solves it
Terms are extracted from the source, rendered into the target language once, then passed with every single chunk as binding data, not as advice. After each chunk is translated, an occurrence-aware verifier counts how many times each source term appears in that chunk's source and how many times the agreed target term appears in the output. Mismatches are repaired verbatim before the chunk is accepted. The reader never sees a drift, because no drifted chunk is ever delivered.
Further reading: glossary strategy for long documents, three root causes of consistency drift, how the workflow uses glossaries.