In May 2023, Google unveiled a new, AI-powered tool called ‘Help Me Write’. Currently available in beta for Gmail and Google Docs, the tool automatically drafts email and form letters based on a user’s simple instructions.
It promises to be a great boon for productivity, not to mention for those of us who loathe writing email. But the introduction of AI-generated communication may also be the death knell for endangered languages.
Thanks to the advanced capabilities of newer large language models, AI-authored text is about to be incorporated in most mass-market writing programs, from texting to email to general document generation.
Some of these platforms have incorporated AI in the past, for example to help make suggestions for how to finish a sentence. But AI is about to have a much larger hand in writing than ever before.
Currently, bot-written text can come across as a little hokey or generic-sounding (although ChatGPT does a pretty good job composing sarcastic text messages to my friends.) But it’s already sophisticated enough for plenty of standard work email.
It’s also improving, and with enough training over time, an AI tool may even be able to learn individual preferences and write in a more personal style.
The prospects for business are obvious, and the bots may help with more than just email laziness. AI-written text could be an equaliser, improving accessibility for people who have trouble writing appropriate prose themselves, whether that’s for disability, educational, or other reasons.
And while we may wind up in a world where AI-tools simply compose emails to each other, I personally cannot wait to leave most of my crushing and tedious inbox to the machines.
But bot-composed text can also cause trouble. Language learning models have already created controversy by generating unexpected content, from inappropriate advice to harmful language, making clear the importance of careful editing and oversight.
Plus, the use of AI can be insensitive or rude in certain contexts. Earlier this year, Vanderbilt University administrators forgot to remove a ‘written by AI’ note from a condolence email they sent out after a school shooting, disgusting their student population.
A lesser talked-about concern is what AI-written text will do to language. I spent half of my life immersed in Swiss German, which is an umbrella term for a family of dialects spoken in the German part of Switzerland and some alpine towns in Italy.
Swiss German dialects are verbal languages with no universal spelling, but that hasn’t stopped people from writing in them. And because the spelling is purely phonetic, each person’s words tend to reflect their specific regional accent, as well as their personal quirks.
The introduction of spellcheck and autocorrect changed part of the communication amongst my Swiss friend group. Suddenly, nearly every Swiss German word in our emails and text messages was squiggly-lined or incorrectly altered.
Many of us ended up disabling the correction tools in annoyance. But some people switched to official German, a formal language that was supported by the tools.
As AI-writing becomes common, there’s no question that most of us will begin using formal German. Anything else would be too impractical.
AI-writing tools will be available in hundreds of languages. Google also has an AI-project called the ‘1,000 Languages Initiative’, which they claim will support the thousand most popular languages on Earth, including rare ones spoken by less than a million people.
But it’s not clear whether those efforts will translate to verbal languages, like the unique Swiss dialects. More generally, every edge case will probably conform to whatever formal words the AI has been trained on, causing us to lose some of the richness and diversity of human language.
And that’s not all. AI is becoming incredibly adept at translating from one language to another, which may soon reduce the need for human translators in publishing.
While this is an exciting development for some publishers, some human translators have played an important role in oppressed countries by inserting political or other kinds of subversion into texts.
That’s just one example of something that could get lost in AI-translation, and there’s a lot more that we don’t even know about.
When it comes to language, AI can only be trained on an existing corpus of vocabulary, meaning it can only look backwards.
An AI tool may be able to combine words in creative ways, or learn a style from past writing, but it won’t naturally invent new slang, and it will always perpetuate the past into the future. We have no idea what this means for language development.
Amidst larger concerns about the risks of AI, there are many smaller ripples like these; side effects that are difficult to anticipate.
So, as we adopt AI-powered writing tools, it’s worth thinking about how they will impact both the conservation and the progress of language … and whether we care.
Read more about artificial intelligence:
