“Driven by the de facto dominance of English, AI multilingual alignment shows a form of cross-linguistic deference, a systematic tendency to resolve cross-linguistic disagreement by treating the English judgment as authoritative, rather than acknowledging the natural cross-linguistic inconsistency of human natural language.”
That’s from the “Cagliari Statement on Linguistic and Conceptual Justice in AI“, which objects to correcting such inconsistencies” in ways that “suppress legitimate semantic plurality and progressively establish English (spoken natively by only about 6% of the world’s population) as the de facto standard for what meanings should remain equivalent across languages.”
Rather, according to the statement, “language models require language-specific alignment, grounded in rich data on how ordinary speakers actually use their words in each language.”
The statement, which is currently open for signatures, includes distinct recommendations for AI developers, researchers in linguistics and philosophy, institutions and funders, and standard-setting bodies, among others.
It was originally drafted by Çağla Çimendereli (Old Dominion), Filippo Contesi (Cagliari), and Masaharu Mizumoto (Japan Advanced Institute of Science and Technology) and worked on during the Linguistic Justice Society Satellite Workshop at the 6th European Experimental Philosophy Conference.
The statement provides some background on the issue:
While the capabilities of current frontier AI models have shown remarkable progress in areas such as mathematical ability, their semantic competence in natural languages—including, in particular, lower-resource languages—fails to accurately reflect basic human semantic intuitions (i.e., how ordinary speakers actually use their words in diverse contexts). This gap is what this statement addresses.
Multilingual AI systems now answer questions about foreign facts, reports, opinions, etc. in local languages without explicitly indicating that translation mediates the response. Moreover, because these systems acquire meaning from patterns of usage, they inherit the semantic norms of the larger corpora they are trained on—which are predominantly English. We therefore hold that decisions about how such systems behave across languages are not merely technical, but normative.
The full statement is here.
(via Çağla Çimendereli)
This is all news to me.
So, is it true that most LLM’s are trained on a corpus that is English-only, and then generate non-English output only at the end of the “thinking” process? I.e. that their language of thought is English?
If so…yikes.
Is this true even of the LLM’s that China is spinning up?
Some future Jerry Fodor is going to have a field day arguing that the LoT is really English.
Not English-only but English is superdominant and so most LLMs default to English. China’s LLMs might be different…