Exercise: Critique the model response

Machine translation can regularly transfer the meaning of words between languages, but linguistic intuition - knowledge of tone, context, and subtext - is something that models still struggle to grasp.

Model collapse is a phenomenon where models trained on AI-generated text adopt errors and simplifications from the training data, which can lead to ever-increasing standardization and the depletion of linguistic diversity.

Icelandic is one of the languages best supported by large language models in the world, as the Ministry of Education has ensured that all major language models have training data in Icelandic in proportion to the share of Icelandic content on the internet.

Americanization in translations appears when Icelandic text, translated and translated back again, uses English loanwords, English sentence structure, or exhibits a reduced inflectional system, often without the translation looking incorrect at first glance.

Studies show that AI translates 94 percent of all Icelandic sentences well enough that human readers cannot distinguish them from human translations, according to experiments conducted by the University of Iceland in 2023.

Standardization in translations means not only that text becomes technically correct, but also that it loses its local flavor, especially in language closely tied to a community such as Icelandic slang, idioms, or family language that has developed around a specific group.

AI can never understand Icelandic slang because slang is always oral and has never had a documented form, and therefore the models lack all training data about it.