Magic linguistic ethics

When to Use

• Any new dataset being considered for training/eval (early — before download).
• Endangered or Indigenous language data of any kind.
• Religious / sacred text use (Bible-NLP, Quranic, Vedic, Indigenous oral histories).
• License audit before release (open / community-gated / restricted decision).
• Attribution and provenance tracking design.
• Drafting a model card's "Ethics" / "Limitations" / "Intended Use" sections.
• Routing decisions involving community-controlled archives (DELAMAN, ELAR, AILLA, PARADISEC).

When NOT to use: the dataset is your own English-only synthetic data with no community attribution issues; or the operation is a pure technical refactor with no data implications. Even then, ask once — under-using ethics is the modal failure mode.

Why this skill is A-tier (and not optional)

A "good" engineer can build a tokenizer, mine bitext, fine-tune a model. None of that protects against:

• Training on a dataset whose community didn't consent to model use, even if the license technically permits it.
• Releasing a model that generates Sacred Indigenous content without permission.
• Stripping attribution lineage during a dataset merge so credit becomes irrecoverable.
• Using "open" Bible-NLP for a commercial chatbot whose generations the source community would not endorse.

These are the high-cost mistakes. They damage communities, harm professional reputation, and increasingly carry regulatory consequences (EU AI Act, Indigenous data sovereignty laws, GDPR-like cultural data protections). This skill is mandatory routing — not optional polish.

CARE complements FAIR (does not replace it)

FAIR (Findable, Accessible, Interoperable, Reusable) is the technical-data discipline. CARE adds the social discipline:

FAIR	CARE
Findable — metadata + identifiers	Collective benefit — does this serve the source community?
Accessible — fetchable	Authority to control — community decides terms
Interoperable — standard formats	Responsibility — for harms downstream
Reusable — clear license	Ethics — through engagement, not just consent

A dataset can be FAIR (standardized, downloadable, openly licensed) and still violate CARE (community wasn't asked, terms favor the user not the source).

The rule: every dataset routing through magic-linguistic-corpus or magic-linguistic-bitext MUST pass through this skill at least once for a CARE check, even if FAIR is satisfied.

FPIC — Free, Prior, Informed Consent

FPIC is the standard for Indigenous and traditional knowledge use. Often misunderstood as "got a signature → done". Real FPIC requires all four:

| F: Free | Without coercion or pressure | | P: Prior | Before the data is used / collected (not retroactive) | | I: Informed | Community understood what models trained on this data could do | | C: Consent | Affirmative; can be withdrawn |

FPIC is process, not document. A signed license without ongoing engagement is not FPIC.

For projects involving Indigenous data, route to a partner organization (e.g., Te Hiku Media for te reo Māori; First Languages Australia; ELAR for many endangered languages) BEFORE first contact with raw data.

License Audit Workflow

MANDATORY READ references/license_audit.md before any dataset-mix release.

Quick decision tree:

Does the dataset have an explicit license?
├── NO → don't use; or contact source for clarification
└── YES
    ├── CC0 / Public Domain → check for community norms anyway (CARE step)
    ├── CC-BY → attribution required; track in registry
    ├── CC-BY-SA → ShareAlike obligation propagates to model output license
    ├── CC-BY-NC → COMMERCIAL use blocked; common dataset-mix gotcha
    ├── CC-BY-ND → no derivatives; usually blocks training
    ├── Custom community license (Māori Data License, etc.) → respect terms; usually requires community-gated release
    ├── ODbL / similar database licenses → propagate ShareAlike to derived data
    └── No-NC, derivative-allowed, attribution-required → standard "open" dataset

Compatibility check: when mixing N datasets for training:

• Any CC-BY-NC in the mix → entire model is non-commercial-use only.
• Any CC-BY-SA in the mix → entire model output must propagate the SA term.
• ND blocks derivative use; if mixed, you've already violated terms.

Sacred-Text / Culturally-Sensitive Material Gating

Framework + 5 canonical examples (Q6 resolution from ralplan iter-2). NOT a hardcoded blocklist — a decision framework anchored by examples.

MANDATORY READ references/sacred_text_gating.md before training on or releasing models that touch any of:

Example	What's restricted	Why
Quranic text	Generation/transformation; respectful contextual handling	Religious community standards on representation
Indigenous oral histories	Public release; transformation into model output	Custodian permission required; many are not for outsider use
Sami yoik recordings	Use in non-Sami contexts; commercial training	Cultural ownership; Sami Council guidelines
Aboriginal Australian songlines	Recording, distribution, model use	Custodian-permission required; Indigenous Cultural and Intellectual Property (ICIP) protocols
Bible-NLP / liturgical text	Commercial training; distortion of canonical text	Many translations under permissive license but communities prefer non-commercial use

The framework: source community involvement → use intent → distribution scope → technical safeguards. Apply it to any text you're unsure about.

Attribution Registry

Every dataset entry traces to:

• Source URL or citation.
• License (with version + date).
• Date acquired.
• Permission record (if FPIC required).
• Community contact (if applicable).
• Lineage (parent dataset(s) it was derived from).

When merging datasets, lineage MUST be preserved — never strip attribution to "clean up" a merged file.

MANDATORY READ references/attribution_registry.md for the schema.

Workflow

Step 1 — Scope ethics (early, before data acquisition)

• Read workspace_state.md for target language + vitality (from magic-linguistic-scope).
• Vitality 6b+ (threatened or worse): heavy ethics — community partner identified.
• Vitality 0-6a + non-Indigenous: standard ethics — FPIC + license check.
• Note in workspace_state.md under "Ethics Status".

Step 2 — Per-dataset ethics check (during Acquire)

Before recommending any dataset to magic-linguistic-corpus or magic-linguistic-bitext:

1. License: known + compatible with target use.
2. CARE: does it serve the source community?
3. FPIC: required? If yes, obtained?
4. Sacred-text framework: any concerns?
5. Lineage: traceable?

If any answer is NO or UNCLEAR, BLOCK the dataset until resolved. Document in workspace_state.md.

Step 3 — Mix audit (before training)

MANDATORY READ references/care_fpic_checklist.md.

When N datasets are mixed:

• License compatibility matrix (above).
• Attribution registry combined.
• Most-restrictive-wins for downstream model license.

Step 4 — Release gate (final)

Three release modes:

Mode	Requirements
Open	All-open licenses + attribution complete + no community restrictions + standard model card
Community-gated	Community sign-off documented; access criteria + revocation path; model card cites community partner
Restricted	Use-policy + access controls (auth, audit log); legal review

Step 5 — Model card

Every release includes a model card with:

• Datasets + licenses + lineage.
• Ethics statement (CARE alignment).
• Intended uses + restrictions.
• Limitations and known biases.
• Contact for community concerns / opt-out.

Anti-patterns (NEVER do)

• NEVER recommend a dataset without checking community norms beyond license. Bible-NLP is open but the community use norms apply (commercial chatbot generation can violate these).
• NEVER skip FPIC because the user is in a hurry. A 30-second flag now prevents a multi-week rollback later.
• NEVER strip attribution lineage when merging datasets. Once the lineage is gone, you cannot rebuild it; you cannot honor the propagated obligations.
• NEVER assume "open-licensed" = "ethically usable" for endangered/Indigenous data. License gives legal permission; CARE gives ethical permission.
• NEVER treat FPIC as a one-time signature. Consent can be withdrawn; engagement is ongoing.
• NEVER mix CC-BY-NC datasets into a commercial training run. The license forbids it; the resulting model is technically infringing.
• NEVER release sacred-text-trained generative models without community sign-off. The downstream generations are the harm vector, not the training data itself.
• NEVER propose "we'll add the model card later". Model cards are part of the release decision; without one, release is incomplete.

Edge Cases

• Unknown license (data with no LICENSE file): treat as restricted. Contact source. Default deny.
• Aggregator dataset that wraps many sources (CulturaX, MADLAD-400, OPUS): per-source license check is needed; the aggregator's license isn't sufficient.
• Multi-source paper-released datasets with no license statement: cite + use academic-only norm; do NOT mix into commercial training.
• Personal communication (email/Discord) saying "use my dataset": get a written license; informal permission is not FPIC.
• Dataset whose attribution-required source is unknown (orphan data): treat as restricted; do not propagate.
• Rights conflict between original speaker and field documenter: speaker rights take precedence; documenter has secondary rights only.
• Languages with no formally recognized speaker community (recently-revitalized, language with diaspora): consult ELDP / DELAMAN partners for guidance.

Output Format

## Ethics Assessment: <Dataset / Mix>

**Source(s):** ...
**License(s):** ...
**License compatibility (mix):** OK | CONFLICT — details
**CARE check:** PASS | NEEDS-WORK — gap: ...
**FPIC required?** NO | YES — status: OBTAINED | OUTSTANDING | NOT-APPLICABLE
**Sacred-text framework concerns:** NONE | SEE: ...
**Attribution registry status:** COMPLETE | INCOMPLETE — missing: ...
**Recommended release mode:** OPEN | COMMUNITY-GATED | RESTRICTED
**Outstanding actions:** ...