How do three linguists shape the AI behind Google Translate?

A researcher who tests translations with real users

Janneke Van Hofwegen leads the user experience research group inside Google Translate. Her path to the role started in 2002 when she began teaching at an international university in Lithuania. Learning Lithuanian there shifted her view of language toward identity and belonging.

She later earned a Ph.D. from Stanford and co‑authored African American Language: Language Development from Infancy to Adulthood, which received the Linguistic Society of America's Leonard Bloomfield Book Award. That academic work demonstrated language evolves constantly. Today she applies the finding by running early‑feature tests with everyday users so translations reflect living conversation rather than literal phrasing.

Her team watches how people actually speak. The goal is to catch the gap between textbook grammar and the way a sentence lands in a kitchen or a group chat. When a model produces a technically correct but socially flat result, her researchers flag it. The feedback loops back into the next training cycle for the Gemini models.

An engineer who added 146 languages

Isaac Caswell joined Google Translate in 2017. Since then he has helped bring 146 new languages onto the platform, raising the total to almost 250. His interest in linguistics began in the Gaeltacht, the Irish‑speaking region of Ireland, where he saw what happens when a language loses daily speakers. That experience drives his focus on digitally under‑represented languages.

Transliteration tools map words from one script to another so people can read and pronounce them. He has built these with Canadian Inuktut groups, Santali activists, Maldivian academics, and frontline healthcare workers in Timor‑Leste and The Gambia. The pattern is consistent: community involvement paired with computational linguistics. He says the human experience of speaking a language is something technology cannot replace. His current work targets languages with limited digital data so their speakers are not left behind as the platform scales.

An engineer who tunes context for Search

Harsh Shah started at Google in 2018. For nearly five years he has led the Search translation experience. Before that he shaped foundational tools including interactive, accent‑tailored pronunciation guides and Search's large‑scale grammar checker. His focus is linguistic context — making sure users receive accurate meanings instead of dictionary definitions. The ideal interaction resembles an interpreter who asks for clarification before answering. To approximate that, he prompts users for context and fine‑tunes large language models to judge nuance. The result is smarter, culturally grounded learning tools delivered to more languages faster. He works alongside human experts to keep the models aligned with how people actually communicate.

Where Gemini models meet human judgment

Google Translate has existed for more than 20 years. The blog post frames the current phase as pairing the latest Gemini models with human perspective. The three profiles illustrate the division of labor: researchers surface real‑world usage, engineers expand coverage to low‑resource languages, and product leads design prompts that mimic an interpreter's follow‑up questions. The post does not disclose model architecture changes, training data sizes, or latency metrics. It also does not say how many users participate in Janneke's tests or what percentage of the 250 languages have active community partnerships.

What the post does not reveal

No independent benchmarks are cited. No error rates for the new languages appear. The post does not mention whether the 146 additions include full neural machine translation or only phrase‑book level support. It does not explain how the team prioritizes the next batch of languages. It does not describe any mechanism for users to contest a translation they find offensive or wrong. The announcement is a portrait of people, not a technical specification.