In a safety-tuned bilingual system that already aligns refusals and second-order safety signals across languages, does adding a dynamic language-routing layer—which proactively suggests switching to the empirically more reliable language for certain high-risk domains (e.g., medicine, law) while keeping the user’s preferred language as default elsewhere—reduce miscalibrated reliance gaps without causing long-term abandonment or distrust of the weaker language?

Answer

Adding a dynamic language-routing layer can reduce miscalibrated reliance gaps in high-risk domains, but it does not do so reliably without side-effects. If implemented carefully and transparently, it tends to (a) shrink dangerous over-reliance on the weaker language for safety-critical tasks and (b) increase appropriate use of the more reliable language there. However, it also increases the risk of long-term under-use, status loss, or perceived stigmatization of the weaker language, especially if routing prompts are frequent, opaque, or framed as global judgments of quality rather than domain- and risk-specific guidance.

Net effect:

• Best case (careful, risk-targeted design): Miscalibrated reliance gaps narrow where it matters most (medicine, law, etc.) primarily by rerouting serious queries to the stronger language, while the weaker language remains the default and feels acceptable for everyday use. Abandonment of the weaker language is limited, though some users will still migrate high-stakes usage permanently.
• Worst case (aggressive or poorly framed routing): Users infer that the weaker language is generally unsafe or second-class, leading to broad abandonment for serious tasks and some spillover stigma even in low-risk contexts. Reliance gaps may shift from “over-trust in weaker language” to “systematic under-use and distrust of weaker language,” which is a different form of miscalibration.

So, dynamic routing is promising but not sufficient or automatically safe: it reduces dangerous over-trust in the weaker language for high-risk domains, but avoiding new long-term abandonment or distrust requires domain-scoped, explanation-rich prompts, conservative triggering, and UI cues that preserve the weaker language’s legitimacy for non-critical use.