Most people who use Google Translate have never thought about what it took to build it. Macduff Hughes spent 12 years thinking about almost nothing else.
In Episode 223 of the Localization Fireside Chat, I sat down with Macduff Hughes, former Director of Engineering at Google Translate, to talk about one of the most consequential journeys in the history of language technology. Macduff joined the Google Translate team in 2012, guided the product through the landmark 2016 shift from statistical to neural machine translation, and then led it into the LLM era before retiring in 2024 after 35 years in the industry. Before Google, he spent eight years managing the Adobe Acrobat team — helping establish PDF as trusted global infrastructure. He is, by any measure, someone who has been in the room when it mattered.
What struck me most in this conversation was not the technology. It was the human story underneath it.
How Statistical Machine Translation Actually Worked
When Macduff joined the team in 2012, statistical machine translation was the state of the art. The system worked by crawling the web for billions of pairs of sentences that appeared to be translations of each other, building a massive probability database of which words and phrases in one language tended to appear alongside which words in another, and then stitching those probabilities together using a language model trained on huge volumes of text.
It worked well enough to be genuinely useful. But it had a ceiling. Macduff described a category of errors he called negation errors — cases where the model would essentially flip a coin on whether to include a “not” in the translation, with no sense that dropping or adding a negation completely reverses the meaning of a sentence. The statistical model could see that 51% probability pointed one way and 49% pointed the other. It had no way of knowing those two outcomes meant completely different things.
“Some people on the team knew even in 2012,” he told me. “It’s gonna have to be neural networks.”
The 2016 Neural Transition
By 2016, the case had become undeniable. What followed was one of the most exciting and consequential years of Macduff’s career. The Google Translate product team and the Google Brain research team worked in close collaboration — the researchers bringing neural architecture expertise, the product team bringing the practical knowledge of what hundreds of millions of users actually needed. Macduff was named as a co-author on the landmark GNMT research paper alongside Jeff Dean and 29 other contributors. The result was an error reduction of up to 85% on some language pairs, launched overnight.
But the transition was not just a technology swap. Macduff was candid about the human side. Some engineers who had spent years building and optimizing the statistical system simply did not find neural networks interesting. “They just said, I don’t find this interesting. That’s not the kind of problem that’s in my sweet spot.” Some moved on.
Others walked on air. Macduff described native speakers of various languages running the neural system on their own language for the first time. “Chinese now sounds like Chinese instead of a computer’s attempt to imitate Chinese.” The quality jump was visceral and personal for the people who built it.
500 Million Users and the Language Access Question
At its peak, Google Translate was serving hundreds of millions — sometimes approaching a billion — users across all platforms. But Macduff always tried to keep one number in mind: roughly 80 to 90 percent of the internet is in just 10 languages.
For native English speakers, that statistic barely registers. The internet feels like it was built for you. For a monolingual speaker of Vietnamese, or Welsh, or any of the thousands of languages with limited digital presence, the experience is completely different. You are standing outside a vast ocean of information that you simply cannot access.
“The first big promise of machine translation,” Macduff said, “was opening that enormous ocean of information in the big languages and making it available to everyone.”
He described receiving messages roughly once a year from users who had found a partner, resolved a medical emergency, or navigated a crisis in a foreign country because Google Translate had been just good enough to bridge the gap. Not perfect. Good enough. That distinction mattered enormously to him.
The team also had to deal with the politics of scale. Governments lobbied to have their languages added. Welsh, Macduff noted, was a language he had started learning as a hobby before he even joined the team — drawn in partly by Welsh ancestry, partly by a trip to Wales where he heard the language alive and in use. That personal thread ran quietly through his entire tenure.
Bias in the Corpus
One of the harder problems the team grappled with was gender bias embedded in training data. Because the statistical and early neural models were trained on web text, they inherited the internet’s own inequalities. The most visible example: in languages where personal pronouns are gender-neutral, translating a sentence like “they are a doctor” into English would produce “he is a doctor” — because the corpus reflected a world where the majority of written references to doctors were male.
Macduff was clear-eyed about what this meant. It was not just a translation error. It was a product used by hundreds of millions of people every day subtly reinforcing the idea that doctors are male. “If you’ve got a service that hundreds of millions, maybe billions of people are using every day, it’s your responsibility to think about what larger social impact you have.”
The team built a system to detect cases where gender was unspecified in the source but required in the target, and offer users multiple options. It helped. But Macduff acknowledged the larger problem remains unsolved — and has only grown more complex as general-purpose LLMs have taken over.
The LLM Era and What It Means for the Industry
By the time large language models arrived, Macduff was still in the chair. His view of what it meant for translation products was clear: dedicated machine translation systems were going to become downstream of LLMs, not independent of them. The question for a product like Google Translate was no longer just how to produce the best translation, but how to use the conversational and reasoning capabilities of LLMs to let users understand and interrogate the translation — to ask about formality in Japanese, about dialect in Arabic, about gender in any language that requires it.
He also had a frank message for the localization industry. Straight text translation, in his view, is asymptotically approaching solved. The bigger frontier is speech and video. And the deeper disruption is a question most LSPs have not fully confronted: if a user in Thailand can go to any frontier AI, type a question in Thai, and get a step-by-step answer drawn from an English user manual — without anyone ever commissioning a translation — what exactly is the translation industry selling?
“Spend some time getting your head out of the translation task,” he said, “and looking into what problem you are solving for businesses and users and how you can help.”
That is not a dismissal of human expertise. It is an argument for repositioning it.
35 Years, Two Paradigm Shifts, No Regrets
Macduff retired in 2024. He told me at the end of our conversation that the thing that surprises him most — even now — is how far conversational translation still has to go. The vision of setting your phone on a table and having a fluid cross-language conversation in real time has been “around the corner” for years. He appreciates now, having worked with professional human interpreters, just how genuinely difficult that task is even for people.
He left the industry with something rare: a completed chapter and the perspective that comes with it. He came on this podcast because he still enjoys keeping in touch with the people doing this work. That generosity came through in every answer.
Watch the full episode on YouTube, listen on podcast, or read more at the links below.
YouTube: https://youtu.be/dwYgYj_Cmvg | Podcast: https://localization-fireside-chat.simplecast.com/episodes/the-google-translate-journey-from-statistical-to-neural-to-llm-macduff-hughes | Connect with Macduff: https://www.linkedin.com/in/macduff-hughes-98665a3 | LFC on LinkedIn: https://www.linkedin.com/company/localization-fireside-chat | Robin on LinkedIn: https://www.linkedin.com/in/robinayoub | Blog: https://www.robinayoub.blog | N49Networks: https://www.n49networks.com | Book a call: https://calendly.com/robin-ayoub/30min
Robin Ayoub 
Leave a comment