Google Translate just added a feature that solves the tool's biggest problem

As I was scrolling through Instagram Reels the other day, I came across a video of a woman discussing her engagement and follower reach.

Having no knowledge of the woman or her content, I was about to scroll past it when she said something that made me sit up and pay attention.

She spoke of how she had noticed a spike in comments written in Spanish, which struck her as odd because she spoke only Hindi in her videos.

After a brief investigation, she realized that Instagram was using Meta AI to translate her videos into other languages.

Not only was the audio translated, but Meta AI was editing the video to sync her lips with the translated audio.

Paying attention properly, I began to notice words that didn't fit and phrases I had never heard of before.

Now that I'm aware of it, I can pick out these awkwardly translated videos without needing the "Translated with Meta AI" tag at the bottom of the video.

While this has been a shock, what interested me was the poor translation quality.

It's a sloppy use of AI that undermines how effective AI can be for supporting translation tools, as demonstrated by Google Translate in its latest update.

Related

7 simple ways to use Google Translate

Translate text and speech on the fly with Google Translate

Posts By  Jon Gilbert

Google Translate was primed to take advantage of AI

The latest update is a huge step forward for the service

Google Translate switched from a statistical machine translation service to a neural machine translation engine.

This meant that rather than translating words one by one, Google Translate would analyze whole sentences at once.

It still wasn't perfect, but Google has refined the service over the years.

A 2021 study carried out by the UCLA Medical Center found that when asked to translate English to Spanish, it managed a 94% accuracy rate. However, for rarer languages like Armenian, the accuracy rate dropped to 55%.

Google Translate is brilliant for simple translations; you can rely on it for foreign travel where you don't need exact translations.

However, its inability to understand idioms, local expressions, or slang often means nuance is lost in translation.

To fix this problem, Google updated Google Translate with Gemini capabilities to help translate the meaning of idioms.

In its announcement post, Google uses the English idiom "stealing my thunder" as an example. This is a grammatically simple sentence that can be accurately translated word-for-word into another language.

But when we say "They stole my thunder," we don't mean it literally. Thus, a non-English speaker with no knowledge of this idiom would not understand a direct translation.

Now, the Spanish translation of "stole my thunder" reads as "stole the spotlight," which is, of course, what the idiom means.

This update marks a significant step forward for translation tools. But AI can't solve everything.

Better translation tools than Google Translate exist, but perhaps not any more

Google's biggest weakness may no longer be relevant

Google Translate has been the most accessible and widespread translation tool for years. However, its inability to accurately translate context is well-documented.

Its strengths lie in its quick, offline translation capabilities that are good enough for everyday situations.

Tools like DeepL or PONS offer fewer languages with limited offline capability and a paywall for full functionality compared to Google Translate.

However, their accuracy surpassed Google Translate, so they were marketed to business, not consumer markets.

Until the advent of AI tools, this was the state of affairs.

But now that Google Translate's biggest problem has been addressed, we can look to the most accessible translation tool for accuracy again.

What makes Google Translate fantastic is that it doesn't wholly rely on AI.

AI-powered translation tools are deeply flawed

AI is not built to accurately translate text

Looking at the success of Google Translate's Gemini-powered translations, it's easy to see AI as the future of translation tools. However, like any AI tool, the reality doesn't match up.

The auto-translated Instagram Reels that have appeared on my feed offer broadly accurate translations. I get a good idea of what the audio is about, but it's clear that AI is behind the translation.

AP Recommends

Tech insights about everything mobile directly from the Android Police team.
Subscribe

AP Recommends

Tech insights about everything mobile directly from the Android Police team.

Subscribe
By subscribing, you agree to receive newsletter and marketing emails, and accept Valnet’s Terms of Use and Privacy Policy. You can unsubscribe anytime.

Just as AI-generated text has clear signposts, AI translations are marked by unusual wording, incomplete phrases, and poor grammar.

You can ask any AI chatbot to translate for you, but this is where the problems arise.

AI doesn't understand the data; it's making an educated guess about what should come next. It cannot understand context, nuance, and meaning.

Google's new Gemini-powered translation feature is still in beta.

Understanding context and nuance is still challenging, so for a feature like this to work, Google can't just add Gemini support and call it a day.

AI has to be used to support existing translation tools, not replace them

In 2020, Google replaced its translation engine with a deep learning network built on the same architecture that powers LLMs. So while an LLM does not power Google Translate, it shares a lot in common.

This is the correct approach, as proven by Meta AI's terrible Instagram Reel translations; LLMs are not as reliable as purpose-built translation tools.

Google Translate's latest update has yet to be extensively tested, but it marks a significant step forward in the service's accuracy.

If you want to try it, it's currently available in the US and India with English translations in nearly 20 languages, including Spanish, Hindi, and Chinese.