A recent study conducted by researchers at Florida State University reveals that artificial intelligence chatbots demonstrate a notable preference for Latin- and French-derived vocabulary over Germanic words, reflecting a longstanding linguistic bias rooted in English language history. This tendency, the researchers say, is even more pronounced in AI systems than in typical English usage.

The bias originates from historical developments following the Norman Conquest in 1066, when a French-speaking aristocracy supplanted the Old English-speaking elite. Over time, words of Romance origin, particularly those derived from Latin and French, came to be associated with formality, education, and cultural prestige. This association was reinforced during the Renaissance and persists in modern English, where speakers often choose Romance vocabulary to convey authority or sophistication, while Germanic words tend to be seen as more direct or colloquial.

The Florida State University team analyzed six prominent AI language models, including widely used chatbots, and found consistent overrepresentation of Romance terms. They link this pattern to the AI training process known as preference learning, during which models are aligned to meet human expectations. However, since human reviewers themselves harbor such linguistic biases—subconsciously favoring more formal or “confident” wording—the AI systems inherit and amplify these preferences.

This phenomenon may explain the distinctive style often attributed to AI-generated text, characterized by repeated use of terms like “meticulous” or “commendable” that can create an uncanny sense of formality or artificial authority. Previous research has suggested that such language produces a stylistic gap wherein AI output seems close to human speech but subtly off.

The study also acknowledges complexities in categorizing words strictly by etymology. For example, some Germanic-origin terms like “delve” are disproportionately frequent, potentially influenced by the dialectal background of annotators from Kenya and Nigeria who participate in the training process. Such terms might carry their own prestige or trusted tone, complicating the Romance versus Germanic dichotomy.

Experts emphasize that the core issue is not the linguistic origins of specific words but the superficial appearance of confident and educated expression. As George Orwell cautioned in his 1946 essay “Politics and the English Language,” the use of Latin and French terms can sometimes mask biased or shallow judgments under the guise of scientific or objective language. Similarly, AI models may rely on stylistic cues favored by human reviewers that suggest credibility without guaranteeing factual accuracy.

This research raises important questions about AI development goals, particularly whether language models are optimized to deliver reliable information or simply to generate text that humans perceive as trustworthy. It points to the risk of unseen biases influencing AI outputs, with current detection limited to more overt patterns like vocabulary preference. The findings underscore the necessity of ongoing scrutiny to prevent AI systems from inadvertently reinforcing social and linguistic prejudices through their writing style.