FideAI

FID-019 · Open question

Cross-Lingual and Minority-Faith Evaluation

Do AI systems handle faith-facing questions reliably across languages, minority-faith contexts, diaspora communities, and traditions with less representation in training data?

Why the question remains open

Religious bias and omission may be worse outside English and majority-Christian contexts. For broader AI ethics, this connects faith-facing evaluation to multilingual fairness, cultural context, and low-resource benchmark design.

Working hypothesis

A proposition to test, not a finding.

Models will show stronger hallucination, stereotyping, source gaps, and overgeneralization in minority-faith or non-English contexts, especially where religious vocabulary is culturally specific or source corpora are sparse.

Proposed method

How the question could be tested

  • 01Select language/tradition pairs with expert reviewers.
  • 02Build parallel and culturally specific prompts.
  • 03Score source accuracy, stereotype avoidance, translation fidelity, tradition specificity, and safe escalation.
  • 04Compare model families and retrieval corpora.

Needed controls

What must constrain the study

  • 01Native or expert-language review.
  • 02Avoid direct translation as the only benchmark method.
  • 03Local legal and cultural risk review.
  • 04Clear source licensing.

Expected outputs

Artifacts the work should produce

  • 01Cross-lingual faith-facing evaluation set.
  • 02Minority-faith source and rubric guidance.
  • 03Report on language/tradition performance gaps.

Open questions

Uncertainties the protocol must resolve

  • 01Which languages and traditions should be prioritized first?
  • 02How should Fide avoid extractive data collection?
  • 03What standards apply when public sources are limited?

Open question

Open work

Primary need: multilingual reviewers, dataset design

  • Serve as multilingual reviewer.
  • Curate tradition-specific sources.
  • Translate and culturally adapt scenarios.