When models explicitly encode true reliability asymmetries in localized meta-explanations, does also showing a per-language reliability indicator (e.g., a 3-level badge) systematically reduce miscalibrated reliance gaps, or does it instead push users to over-weight the badge and ignore nuanced textual framing, leading to new patterns of over-trust or under-use across languages?

Answer

Adding a per-language reliability indicator on top of localized meta-explanations that already encode true reliability asymmetries is likely to change miscalibrated reliance gaps rather than uniformly reduce them. In well-designed, low-friction interfaces, the badge can modestly improve calibration for some users and tasks by making asymmetries more salient. But it also creates a strong, easy-to-overweight visual cue that many users will treat as a primary signal and thus partially ignore nuanced textual framing, leading to new patterns of over-trust and under-use across languages unless the badge is carefully scoped and explained.

Net expectations:

• Calibration: Slight overall improvement in average calibration for high-risk tasks if badges align with real reliability differences and are paired with clear second-order safety signals; some users will better route critical queries to the safer language or seek external checks in weaker languages.
• Over-weighting risk: Many users will default to the badge as a heuristic, particularly in fast or mobile use, diminishing the incremental effect of nuanced meta-explanations and causing over-trust in high-badge contexts and under-use of low-badge languages even for low-risk, everyday tasks.
• Pattern shift, not simple fix: Miscalibrated reliance gaps will not vanish; they will be re-patterned—smaller for some high-risk decisions but potentially larger or less justified for routine use in weaker languages, and more sensitive to any mis-specification or over-simplification in the badge design.

Design implication: Treat per-language reliability indicators as a supporting, coarse cue that must be tightly aligned with underlying monitoring data, domain- and risk-specific where possible, and explicitly framed in the meta-explanations as approximate rather than definitive—otherwise users’ tendency to over-weight the badge will erode the benefits of nuanced textual framing and may introduce new fairness and dignity concerns for speakers of weaker languages.