Opinion | AI could hasten the extinction of languages — and ways of thinking

Far fewer. There is reason to believe that these tools could bring about a mass extinction of languages. And that, more worryingly, would expunge a diversity of ways of thinking and creating.

How this might happen holds clues for how it could be stopped.

First, some basics: Large language models (LLMs) are trained to make predictions about the best fit for what to generate next in a conversation or piece of writing. They learn that the most likely word after “bacon and” is “eggs” and that the probability of “eggs” being correct is higher if other preceding words included “breakfast” or “coffee.”

Except we are already far past bacon and eggs, and much closer to a Joël Robuchon menu combining all the offerings from his restaurants with 32 Michelin stars total. And each new iteration of LLMs surpasses the last more and more quickly.

AI language models are trained on enormous amounts of data — from the books, journals, newspapers and online content available. The more data, the better. But what’s available to train a model varies widely across the thousands of languages used today.

The most powerful models will be those trained on about 20 “high-resource languages” such as English, Mandarin, Russian, German and Japanese. In turn, AI will churn out massive amounts of new text mostly in those languages. Like invasive species, such dominant models could drive out languages for which fewer resources exist for training.

Some of these languages — such as Hawaiian, Quechua and Potawatomi — are already critically endangered because of globalization, migration and cultural homogenization. Currently, about nine per year snuff out. LLMs could increase that extinction rate dramatically.

Crucially, this is about much more than language. If a majority of languages die in the space of a few generations, that will also bring about a collapse of ways of thinking and being. Because the interaction between language and the mind is a two-way street.

Language shapes the brain. It is one of the most powerful ways to organize, process and structure information. The languages we use influence how we perceive the world, what we remember, the decisions we make, the emotions we feel and the insights we have.

In experiments in my lab at Northwestern University and others, we see that people who speak different languages make different eye movements and have different brain activity. Different things in the environment capture their attention; their memories and interpretation of the world and of reality vary. Those who speak many languages have somewhat different neural networks activated by each.

Reality as each of us perceives it is a subjective experience. It results from how our brains combine the input from our senses with knowledge and experience. Language experience gives us a prism through which to see the universe.

The demise of languages winnows the number of prisms through which to refract the world.

Here’s an irony. When I was on sabbatical at Stanford University, most Silicon Valley AI scientists I met spoke, studied, grew up with or were exposed to more than one language. Many spoke two or more languages. The very people potentially contributing to the demise of the multilingual mind are those harnessing its prowess to build these extraordinary artificial intelligence software programs.

If our reality becomes filtered through a much more limited linguistic set, shaped largely by the symbolic systems of math, logic and artificial languages, could our thinking change?

One insight comes from a version of the classic dilemma used to study morality. A trolley is speeding toward five workmen who cannot see it; you are standing on a footbridge above a train track next to a person. If you push them onto the tracks below, they will die. But they will also stop the trolley, saving the five workmen. What to do?

When responding in a native language, 20 percent of people opt to sacrifice one person to save five. Responding in a second language, 33 percent do so. This 13 percent shift toward utilitarian decision-making is called the foreign language effect.

But just when new brain imaging and computational science are beginning to give us insight into why the multilingual mind works the way it does, we are setting a course that might eliminate one of its core features.

Two things. First, AI research, development and use should be regulated in the public interest. At a minimum, this should be at a level similar to other private-sector industries. Better still would be a level akin to that of the defense sector.

Second, to buy time to come up with more solutions, it is imperative to keep as many natural languages actively engaging human minds for as long as possible.

Now is not the moment to reduce the supply of new ideas. Our many languages are one of the most powerful sources of diversity of thought. In the human experience, multilingualism is not noise, it is the signal.