If we deliberately tune second-order safety signals so that the lower-reliability language is slightly more cautious than the higher-reliability anchor language on the same topical risk band, do bilingual users perceive this as unfair linguistic discrimination or as appropriate risk-weighting—and under what conditions (e.g., explicit reliability disclosures, UI framing, or prior expectations) does this asymmetry improve vs worsen miscalibrated reliance gaps and feelings of procedural fairness?

Answer

Bilingual users are likely to see a small extra caution layer in the lower‑reliability language as appropriate risk‑weighting when it matches disclosed reliability differences and stays stable and respectful. It feels like linguistic discrimination when it is opaque, large, or conflicts with observed behavior.

Perceived as appropriate risk‑weighting (more likely when):

• Explicit reliability signals: Clear, simple reliability indicator plus short text (“This language is less tested; answers may be less accurate, so we add extra cautions here”).
• Topical symmetry: On a given risk band, both languages show similar patterns (same topics flagged, similar refusals); the weaker language only adds slightly more hedging or verification prompts, not extra bans.
• Mild, consistent asymmetry: Differences are small, predictable (e.g., +1 verification prompt, slightly stronger uncertainty words), not dramatic style shifts.
• User priors align: Users already believe the low‑resource language model is less capable; extra caution confirms, rather than surprises, them.
• Respectful tone: Wording frames this as model/coverage limits, not as blaming the language or its speakers.

Perceived as unfair linguistic discrimination (more likely when):

• No or vague disclosure: Users just see their language acting more nervous or refusing more with no explanation.
• Policy divergence: Same query, same risk band, but the weaker language gets extra refusals or noticeably harsher treatment, not just more careful wording.
• Incoherent with reality: Users can see the model perform similarly across languages, yet one is systematically more hedged; or the “weaker” label seems outdated.
• Identity‑laden domains: Extra caution appears mainly on culturally central or politically sensitive topics in one language, especially if anchor language responses are looser.
• UI framing is clumsy: Disclaimers implicitly stigmatize the language (“this language is not reliable”) instead of the system (“our model is less tested in this language”).

Impact on miscalibrated reliance gaps:

• Improves gaps when:

  • The low‑reliability language currently suffers over‑trust; modest extra hedging and verification prompts plus explicit reliability labeling nudge users toward more checking.
  • High‑stakes domains are treated similarly in both languages, with small additional nudges steering users to the more reliable language for those domains.
  • The high‑reliability language is not down‑tuned in a way that discourages its use; it remains the clearly “best” option where users can access it.

• Worsens gaps when:

  • Asymmetry is large enough that users avoid the weaker language even for low‑stakes tasks where it is adequate, or abandon the system in that language.
  • Users respond to heavier hedging by ignoring all warnings in that language (“it always panics”), worsening calibration precisely where extra caution was intended.
  • The stronger language is made artificially cautious to hide differences, leading to under‑use of the genuinely safer channel and confusion about which to trust.

Conditions that especially matter:

• Reliability disclosures: Simple badges or short per‑language meta‑explanations make asymmetry feel more legitimate and reduce discrimination perceptions.
• UI framing: Neutral, system‑centric wording (“our testing coverage is lower here”) and consistent visuals across languages softens fairness concerns.
• Stakes banding: If users see the same risk strata and similar strong signals on high‑stakes topics in both languages, they are more likely to accept slight between‑language calibration tweaks.
• User history: Long‑term bilingual users are more sensitive to subtle asymmetries and contradictions; for them, even small, unexplained differences can erode perceived procedural fairness.

Net: Slightly more cautious second‑order signals in the lower‑reliability language can help alignment and fairness perceptions when clearly tied to real, disclosed reliability gaps and when the per‑topic policy is otherwise matched. The same asymmetry hurts fairness and widens reliance gaps when it is invisible, large, or decoupled from actual performance.