When enforcing cross-lingual consistency of second-order safety signals, is it more effective for calibrating over-trust and under-use to (a) tie uncertainty cues and verification prompts to estimated topical risk (e.g., health, law, financial advice) or (b) tie them to estimated language-specific reliability (e.g., weaker low-resource performance across all topics), and how do these two strategies differ in their impact on bilingual users’ perceived fairness?

Answer

It is generally more effective and robust to anchor second-order safety signals first in estimated topical risk and then modulate them by language-specific reliability, rather than relying on a pure language-reliability scheme alone.

• For calibrating over-trust vs. under-use:

  • A topical-risk–anchored scheme better targets when strong uncertainty cues and verification prompts are most needed (e.g., health, law), which reduces dangerous over-trust in both languages without systematically discouraging use of the safer language.
  • A pure language-reliability–anchored scheme ("this language is always more tentative") can reduce over-trust in the weaker language, but it more often creates under-use of the safer-but-less-familiar language and large, sometimes unjustified, reliance gaps across languages.
  • The best calibration comes from a hybrid: strong, domain-triggered signals everywhere high risk, with additional asymmetry that gently steers high-stakes queries to the objectively safer language when users can use it.

• Perceived fairness for bilingual users:

  • Topical-risk anchoring tends to feel procedurally fair: both languages see similar caution patterns on the same risky topics, which matches users’ expectations that safety rules should depend on what is being asked, not which language they use.
  • A language-reliability–only strategy is more likely to be seen as linguistically unfair or stigmatizing ("my language always gets the more hesitant, second-class answers"), even when it is factually motivated.
  • Explicit but carefully framed language-asymmetry cues, layered on top of topic-based signals, can preserve fairness perceptions while still explaining why answers differ and nudging users toward the safer language for high-risk tasks.

So, for both calibration and fairness, (a) topic-risk–based signals, augmented—not replaced—by (b) language-reliability cues, is preferable. Purely language-based caution is more likely to worsen under-use of the safer language and degrade perceived fairness, whereas topic-first designs let consistency constraints operate without hiding real reliability differences.