You’re reading Superintelligent, a newsletter about AI. Click here to get the full edition in your inbox.
Adam Aleksic is the author of “Algospeak: How Social Media Is Transforming the Future of Language.”
In his 1946 essay “Politics and the English Language,” George Orwell argued that bad writers “are nearly always haunted by the notion that Latin or Greek words are grander than Saxon ones.” Today, artificial intelligence chatbots have fallen victim to the same blunder.
The idea that Latin-derived terms like “exhibit,” “triumph” and “consume” sound fancier than Germanic words like “show,” “win” and “eat” traces back to at least the Norman Conquest, when the previous Old English-speaking elites were suddenly replaced by a new French-speaking aristocracy. As Romance vocabulary became associated with the educated upper class, it increasingly took on a feeling of cultural prestige — a feeling that compounded as English borrowed additional waves of Latin loanwords during and after the Renaissance.
Even in the 21st century, this bias continues to permeate our social interactions. We’ll use more Latin terms when we want to speak formally or authoritatively; we’ll use Germanic words to sound crass or casual. I’m writing this article using Romance words like “permeate” and “authoritatively” because they make it sound like I know what I’m talking about.
A new study from a group of researchers at Florida State University suggests that AI chatbots have also inherited this proclivity. After testing six AI models, the researchers found consistent favoritism for words coming from Latin and French over those with Germanic etymologies — even more than you would typically encounter in the English language.
This bias appears rooted in the preference-learning stage, when the models are trained to align with human expectations about language. This process poses an inescapable problem: that you need real people to make sure the machine is aligned, but the human workers are ironically biased as well. As annotators click through sample texts, for example, they are probably subconsciously disposed to approve those that sound confident and incisive.
This new finding could help explain why large language models like ChatGPT and Claude seem to have a distinctive writing style. Previous research has found that AI chatbots tend to overuse words like “meticulous” and “commendable,” creating a kind of linguistic uncanny valley that sounds similar to how you speak, but ever so slightly off. Perhaps the ghosts of Latin and French haunted these words during preference learning, leading human workers to reward more prestigious-sounding sentences.
Of course, the Germanic versus Romance distinction is a simplification of a messier etymological reality. The notoriously overrepresented word “delve” is actually Old English in origin, and its presence may have been influenced by the local dialects of human annotators in Kenya and Nigeria. It’s also quite possible that the verb carries its own separate prestige, sounding more trustworthy than a commonplace synonym like “dig in.”
Romance words were never really the underlying issue here. Jorge Luis Borges praised English as beautiful precisely because of its ability to switch between registers. A good writer should know when to alternate between a “fine, dark Saxon word” like “kingly” or “ghost” and a “light Latin word” like “regal” or “spirit.”
Instead, the real danger is the semblance of a confident, educated tone, which could obscure underlying problems with AI models. In Orwell’s words again, French and Latin are “used to dress up simple statements and give an air of scientific impartiality to biassed judgements.” The same could be true for other common AI constructions, like the em dash or the negative parallelism “it’s not just X — it’s Y,” which sounds impressively analytical even when it’s not conveying anything of note.
How can we be sure that we will not also be hoodwinked by prestige language, convinced that an AI model is saying something profound simply because it’s using French words like “profound”?
More broadly, the chatbot preference for confident vocabulary raises questions about what exactly we’re measuring during preference learning. The goal of a large language model should be to optimize for accurate, reliable information, but AI companies may be optimizing for human trust signals instead, creating the appearance of accuracy rather than the real thing. What other biases are slipping through the seams to cloud our senses?
AI models could be dressing up their arguments with all kinds of linguistic trickery, but we only know about the Romance vocabulary preference because it’s a more easily observable phenomenon. It’s one thing to look up word origins and compare them across texts, but all sorts of language can shape our headspaces in less obvious ways. Unless we are careful, we will just end up echoing our own pretensions back to ourselves.
The post AI chatbots have a Romance language problem appeared first on Washington Post.
