For bilingual non-expert users in safety-critical tasks, does asymmetrically exposing chain-of-thought (CoT)—always hidden in the weaker low-resource language but optionally viewable in the stronger English interface—reduce miscalibrated reliance gaps between languages, or does it instead drive users to over-concentrate high-stakes queries into English even when they cannot fully critique English CoT?

Answer

Asymmetrically exposing CoT in this way is more likely to increase and reorient miscalibrated reliance than to close the gap. Hiding CoT in the weaker low-resource language does help avoid some explanation-induced over-trust there, but making CoT optionally visible only in English tends to:

• Signal that English is the “serious” or “expert” channel, nudging users to route more high-stakes queries into English even when they cannot fully understand or critique the English CoT.
• Preserve or increase over-trust in the English channel, because non-expert, only-partially-fluent users still treat the optional English CoT as authoritative rather than as a falsifiable argument.

Net effect:

• You likely reduce over-trust a bit in the low-resource language (no CoT, more constrained rationales) but
• Increase relative and sometimes absolute reliance on English, often beyond what is justified by actual reliability—i.e., you create or widen a miscalibrated reliance gap favoring English.

So this asymmetric CoT design should be treated as a partial mitigation with new trade-offs, not as a clean fix: it limits persuasive CoT exposure in the weaker language but also encourages users to over-concentrate high-stakes decisions into an English channel whose explanations they often cannot adequately scrutinize. Additional safeguards (clear cross-lingual reliability messaging, strong second-order safety signals in both languages, and non-CoT-based guidance for when to seek human experts) are needed to prevent that shift from becoming a new form of miscalibration.