AI models overwhelmingly choose Estonia’s “kuid” for “but”

Artificial intelligence language models overwhelmingly preferred the Estonian conjunction “kuid” when asked to choose the equivalent of “but”, ERR language columnist Lydia Risberg writes.

Risberg said she noticed around a decade ago that “kuid” was the first choice in many translated and original texts, although “aga” felt more natural to her. According to a language corpus, “aga” is the most common of the three contrasting conjunctions, with 15 billion matches. “Kuid” has nearly 6 billion matches, while “ent” has just over 700,000.

After developing an aversion to “kuid” and later overcoming it, Risberg said she began noticing the word repeatedly in the output of large language models. She then conducted a small experiment with the ChatGPT-5.5-mini, Claude Sonnet 5 and Gemini 3.6 Flash chat applications.

The models were asked to act as Estonian speakers and select a suitable conjunction for 10 sentences in which the conjunction had been removed. Risberg repeated the request three times to see whether the models changed their answers. She did not provide a list of possible choices in advance.

“Kuid” was the first choice in 75 percent of ChatGPT’s responses, 84 percent of Claude’s and 100 percent of Gemini’s. “Aga” was the first choice for ChatGPT in 22 percent of cases, for Claude in 16 percent and for Gemini in none. ChatGPT chose “ent” first in 3 percent of cases, while Claude and Gemini never selected it first.

When asked about the differences between the words, all three models said that the conjunctions are synonyms and often interchangeable. They described “aga” as the most common and neutral option, and the one most suitable for spoken language.

Risberg said the language corpus showed a similar pattern: “aga” is common across many types of text, while “kuid” and “ent” are more typical of journalism and literature. “Ent” is particularly literary and may sound emphatic in spoken language.

She concluded that there was a slight gap between the models’ stated knowledge and their actual choices. Although individual words may irritate people, Risberg said she now believes “kuid” should be allowed to continue existing alongside “aga” and “ent”, which have coexisted in Estonian for centuries.