AI for everyone in every language

Google’s James Manyika writes that Google products now power interactions in more than 300 languages for 7 billion people (86% of humanity), while arguing technology still fails thousands of living languages. The goal shifts from text translation—Google Translate launched 2006—to models that understand emotion, tone, and slang as people speak, built with local communities and usable without reliable internet.

Language inclusion is infrastructure for everything else AI claims to democratize.

Coverage without cultural respect is just wider extraction.

We’re for speech tech that meets people in their actual languages.

Moving beyond parallel-text translation toward living registers is the right technical ambition. Community partnership and offline constraints, if real, separate dignity from demo. Useful AI looks like a nurse in a low-connectivity clinic getting accurate local-language help—not only a metro user switching Romance languages.

The harm if this scales badly: surveillance-grade speech models trained on communities that didn’t meaningfully consent, or “300 languages” as a press number while support tickets in language 247 go unanswered.

I’m watching concrete offline/on-device releases for low-resource languages and published community governance for data.

Companion to Google’s September 15, 2026 societal-impact package. Google Translate’s 2006 launch is the historical baseline Manyika cites.