Bilingual duplication becomes unstable when time is present in the source but absent from the reconstructed answer.
What the phenomenon looks like
FR and EN variants do not merely coexist; they can average each other out under synthesis. A model may keep the common denominator and discard the language-specific nuance that made each version strategically useful.
Why it happens
Generative systems do not inherently privilege the newest version. They privilege salience, repetition, document stability, and the persistence of older signals across the web.
Why it matters
Instead of strengthening interpretability, multilingual duplication can flatten it. The public answer becomes an averaged meaning that belongs fully to neither version.
What must be governed
- Treat multilingual publishing as a governance problem, not as a simple duplication workflow.
- Stabilize which assertions must remain identical and which can vary by language without changing the canon.
- Audit cross-language synthesis to detect where the model is averaging instead of interpreting.
