BCP 47: Identifying Languages and Their Variants

Language is not the same as country

For this project, the broad language identifiers are en, nl, de, es and pl. A more specific tag such as nl-BE identifies a regional variant when the content actually targets that variant. A country choice and a language choice should not be treated as synonyms.

The right level of specificity depends on what the content promises. Our recommendation is to identify the audience first and select the tag that accurately describes the result. Adding a region does not make a generic translation regionally adapted.

An illustrative publishing decision

Suppose a site offers one English article intended for a broad readership, then later commissions a version specifically adapted for a regional audience. The second version should have a documented purpose and its own editorial treatment, not merely a renamed language tag.

Our suggested exercise is to inventory which versions actually exist, how readers select them and what happens if an exact preference is unavailable. Check that the displayed version, content metadata and routing behavior agree.

This is an original publishing example. It does not replace the RFCs' detailed syntax or matching algorithms.

Matching and fallback need a policy

A valid tag does not itself define the site's complete fallback behavior. A product needs a clear rule for unavailable content, and users should be able to change the selection. Our recommendation is to avoid silently presenting one regional adaptation as another.

Language identification also does not prove that the text is written in the claimed language. Content review and technical metadata checks remain separate tasks.

Relationship to LISA and AI

BCP 47 is an IETF reference, not a LISA standard. It supports the broader interoperability story by giving systems a shared way to identify language information.

For AI-assisted work, our recommendation is to preserve explicit language and audience instructions throughout generation and review. A model's detected language can be useful evidence, but it should not silently override the project specification. Mixed-language content and quoted material deserve particular attention.