Date / time: 10 December, 2:00 pm - 4:00 pm
The paper examines how the material used to train large language models is drawn overwhelmingly from a handful of dominant languages and sources, and what this means for the smaller linguistic communities that appear in these systems only through thin, translated, or manipulated material. I develop the argument through the case of the Balkans, where I read the sources directly, and through my recent work on the deliberate insertion of political content into the corpora these models learn from. The framing is oligopolistic/imperial: a small number of firms, concentrated in a few countries, now decide what the rest of the world’s languages are taken to mean.