Every multilingual data vendor advertises the same thing: native speakers, dozens of languages. Behind most of those claims is a dropdown menu where a worker rated their own fluency, and maybe a short test graded by another worker who did the same.
What self-rating gets you
Inflated coverage in the long tail, translated-English data wearing a local costume, and dialect blindness — a 'native Arabic speaker' label that says nothing about whether the person writes Gulf, Levantine, or Modern Standard. The model trained on it speaks committee.
Fluency you can audit
Employment history encodes language the way nothing else does. Where someone worked, in what role, serving which customers, in which country's payroll — that record establishes not just fluency but register: the difference between conversational Portuguese and the Portuguese of a Brazilian claims department.
- Confirmed language fluency is one of the ground-truth fields we hold from a decade as the employer.
- Our coverage is deepest exactly where labeling workloads run heaviest: Europe, LATAM, China, the GCC including Saudi Arabia, and West, East, and Central Africa.
- Regional depth means occupational range per language — Arabic-speaking petrochemical engineers, not just Arabic speakers.
You can fake a dropdown. You can't fake four years of payroll in Guadalajara.
If your model's next frontier is speaking to the world, its training data should come from people who verifiably live and work in it.
We were the employer