Apple Explores AI's "English Brain" and Proposes a Global Communication Solution
Apple and researchers at top European universities have been investigating why it happens and are proposing a master solution to make AI a global communicator.
The "English Brain" of New AI
As luck would have it, most of the big AI language models that we work with these days are actually built around English. Even the ones that are designed to be multilingual have a strong English accent, as one might say. Apple's research identifies what this means: when AI is generating text, in a language like French or Chinese, it may use a phraseology or sentence structure that sounds a little stilted or unnatural to a native speaker of the language. It's as if the AI is still "thinking" in English patterns, even as it's trying not to.
How Do You Measure "Naturalness"
To try to get a better hold of this, the researchers came up with two new ways of measuring how well a given AI does in different languages:
- Lexical Naturalness: Does the AI choose words like a native speaker, or does it choose maybe slightly unnatural, less common words.
- Syntactic Naturalness: Do the sentences appear to have been put together properly for the grammar of that language, or is it sort of a straight, awkward translation.
They did this by comparing AI output with original articles created by native speakers in Chinese, French, and English on Wikipedia.
The Results. AI Still Needs Language Lessons
The results pretty much validated the "English brain" theory. Amazingly, even an AI model developed in China, Qwen, didn't sound as natural as one would hope, even when speaking in Chinese. While Meta's Llama 3.1 model was generally the best at sounding natural, it was still not truly a human level. This is more than sounding a bit unnatural; Carnegie Mellon University discovered in 2023 that AI security filters are simpler to deceive when they are shown non-English data.
Apple's Smart Solution: Teaching AI What Not to Say
So what is Apple's fix for this issue. It's genius. They've had a program learn to recognize these unnatural-sounding outputs and substitute them with something more natural.
The brilliance is how they get examples of "unnatural" language. Instead of tediously collecting them manually, they use a trick of "back-translation." Imagine taking a perfect Chinese text, translating it into English, and then translating the resulting English text back into Chinese. Somewhere along the course of this double-translating, subtle, slightly unnatural patterns (typically called "translationese") come into play. Apple uses these slightly crooked examples to teach the AI what "unnatural" sounds and looks like.
A More Global Voice for AI
With this new method, Apple has been successful in significantly improving the way natural AI sounds, word choice as well as grammar, without disrupting its overall performance on other tests. This research can be a giant leap towards solving one of the largest obstacles for AI systems to communicate with people all over the world. It's about equipping AI with a substantially multilingual voice, and not just an English voice with the inclusion of added translations.
