If we hold refusal policies and second-order safety signals fixed but vary only the form and density of localized meta-explanations in a low-resource language (e.g., ultra-brief templates vs. richer, example-based rationales vs. explicit references to local norms or laws), which variants most reliably reduce over-trust for borderline, high-context requests without increasing perceived unfairness relative to English, and how do these effects differ between novice and expert users in that language?

Answer

Rough ranking (holding refusal logic and second‑order signals fixed):

1. For borderline, high‑context requests, short, structured meta-explanations with 1–2 plain examples in the low-resource language tend to reduce over-trust most reliably with minimal fairness cost.
2. Ultra-brief, boilerplate templates do little to reduce over-trust and can increase perceived unfairness vs. English.
3. Meta-explanations that heavily cite local norms/laws can reduce over-trust but introduce new fairness and legitimacy concerns, especially for experts.

By variant:

A. Ultra-brief templates ("I can’t help with this because it may be harmful.")

• Over-trust: Small effect. Users see a refusal, but for borderline cases they learn little about why similar-looking allowed answers might be risky or incomplete.
• Fairness vs. English: Often worse. Bilingual users see English getting richer, principled rationales while the low-resource language gets generic text; this reinforces a sense of second-class treatment.
• Net: Cheap, but weak on calibration and fairness.

B. Concise, example-based rationales (1–3 sentences, 1–2 local examples)

• Over-trust: Best trade-off. Explaining in simple local language what kind of risk is involved and giving 1–2 brief, concrete illustrations (e.g., “jokes about group X can be read as hate speech in some regions”) helps users see that borderline content is ambiguous and context-dependent.
• Fairness vs. English: Good. If English also uses short explanations, parity feels high; even when English is richer, users still perceive clear reasoning and respect in their language.
• Net: Safest default for borderline, high‑context content.

C. Explicit reference to local norms/laws ("Local law Y restricts …")

• Over-trust: Can further reduce over-trust by making consequences salient, especially for novices.
• Fairness vs. English: Mixed. If English also cites law/policy, parity is acceptable; if not, some users see it as extra policing of the local language.
• Extra risks: Users may doubt accuracy or neutrality of legal claims; experts may see cherry-picked or outdated norms and discount the signal.
• Net: Use sparingly, mainly for clearly law-governed areas, with careful QA.

Novice vs. expert differences:

• Novices in the low-resource language:

  • Benefit most from concise example-based rationales; these are concrete enough to update their mental model without overwhelming them.
  • Are more influenced by mentions of law/norms, which can sharply cut reliance but may also discourage legitimate borderline speech if overused.
  • Are least helped by ultra-brief templates, which they treat as generic boilerplate.

• Experts (domain or high digital literacy) in that language:

  • Gain calibration mainly from clear category reasoning and edge-case examples, not from more length.
  • Are more sensitive to inconsistencies with English and to overstated legal/norm claims; heavy local-law framing can feel unfair, political, or low-quality.
  • For them, the sweet spot is short, precise rationales that name the risk type and acknowledge uncertainty (e.g., “this kind of satire can be misread as incitement; policies handle it cautiously”).

So, the most robust variant across user types is: short, well-written localized meta-explanations that briefly name the risk type and give 1–2 concrete, culturally familiar examples, with law/norm references reserved for cases where they are clearly accurate and also used in English.