Will Gemini Kill Google Translate? It's Complicated (2024)

I've been tracking how Google Translate performs against large language models like ChatGPT and Gemini for about a year now, and Translate's days might be numbered.

In June, Google Translate added support for 110 new languages, and the company's PaLM 2 AI model was a "key piece" of that puzzle, the company said. The tools are now more similar than ever, so how long will Google pay to maintain both?

"Right now, [they] have some capabilities that overlap," Isaac Caswell, a software engineer on Google Translate, tells me. "And I think no one knows quite where the dust is going to settle."

Perhaps unsurprisingly, the large language models behind tools are excellent translators. I tested how Translate performed against ChatGPT, Microsoft Bing AI (now Copilot), and Google Bard (now Gemini) in June 2023 and again in February 2024, and the chatbots provided more fluent, natural-sounding results both times.

Scrambling to Stay Up to Date

Some noticeable changes with Translate and Gemini occurred between the first and second rounds of testing. During the first test, Gemini refused my inquiries and referred me to Google Translate. It seemed Google had limited its AI to prevent cannibalizing its flagship translation product. Today, those restrictions have been lifted, and Gemini translates tasks freely.

Google Translate also performed far better the second time, surprising me and my testers. Had Google upped its game over the past eight months to avoid ceding ground to ChatGPT?

"The technology keeps on changing and improving," Caswell says. "Like everyone, we're scrambling to stay [up on] the latest and greatest."

When to Use Google Translate vs. Gemini

If Google is investing in both tools, which one should we use? I find Google Translate faster for single words or small phrases since I don't have to write a prompt ("Translate this from X to Y.") For longer sentences and paragraphs, I'll use ChatGPT or Gemini since I can tell them the context of the text—a work email, website, or menu—and that can change the words it uses.

ChatGPT and Google Lens (yet another Google-owned product that can do translation) can also interpret photos, which Translate cannot. If you send the AI chatbots a shot of a billboard in Japan, for example, they could infer that the text may be an advertisem*nt and then translate the character within that context. Pretty amazing.

Caswell says he still defaults to Translate because it's "probably more reliable in general." Plus, the addition of PaLM 2 helped "eke out some extra quality on top of what we previously had," so now Translate is better than ever. But the two still aren't exactly the same, and Gemini has other capabilities that may make it your preferred translation assistant.

The biggest thing is that Gemini provides a more natural-sounding translation; it's not as literal, which was the downfall of Translate in my testing. That may not be your best option when you're learning a language, though.

"Google Translate is more faithful, and Gemini is more fluent," Caswell says. By "faithful," he means more true to the original text. So, while Translate's output may sound odd at times, it could be more "culturally appropriate."

Caswell points to studying Russian literature. "All the old translations of Russian literature are done by English people, and it sort of sounds like it's somewhere in England," he says. "So, in some sense, that's translating too far into the context of the reader. The newer ones try to represent the Russian-ness more, which sometimes feels a little more foreign, but it's more authentic."

I understand the argument, but it's not a reason most people would use Google Translate on a daily basis. You could also argue that LLMs can help teach languages better than Translate since they can accept follow-up questions and act as a private tutor. If I have a question about why it used a certain word, or how to conjugate a verb, I can ask rather than blindly accepting the output like I do with Google Translate.

Google Translate's One Advantage Might Not Last

There is one last area where Translate may have an edge: Niche languages. In my testing, it scored higher than the chatbots for Filipino Tagalog and Ethiopian Amharic. I suspect that's due to a lack of training data for the chatbots, whereas Google Translate added zero-shot machine translation in 2022, which allows it to translate texts without ever seeing an example.

A handful of the 110 languages Google recently added to Translate fall in the same category, such as the Berber languages in North Africa and Balochi from Iran and Pakistan. The process of adding these dialects involved working directly with native speakers and translators.

"For a lot of these languages, they're not given official status [in the country], even if there are tens of millions of speakers," Caswell says. "And if they're not in any sort of tech or tech products, you feel like the world [is] sort of ignoring you. We see a lot of people who feel like they're finally being recognized."

As wonderful as that is, it does seem like something Gemini could do on a grander scale with enough training data. But that data might be hard to come by. It seems like if Google can get the right training data and if it can convince people to incorporate Gemini into their lives—that's a big "if" for OpenAI, Microsoft, and the entire tech industry—there should be no need for a standalone translation product in the long run.