'Those who use artificial intelligence are fences'
At the Frankfurt Book Fair, writers and translators try to fight against the biggest theft in history: that of the texts on which Llm, the large-scale language models, are trained
by Lara Ricci
"I call it 'recel': when you benefit from a theft knowing that your benefit originates from a crime": using artificial intelligence is reception for Jörn Cambreleng, director of Atlas, an association for the promotion of literary translation. Llm, the large-scale language models have been trained on texts, images, forms produced by human ingenuity, covered by copyright.
«La capacità dei sistemi di intelligenza artificiale è resa possibile dal maggior furto di dati e della storia» aveva detto l’anno scorso Karin Schmidt-Friderichs, presidentessa della German publishers & booksellers association, durante la conferenza stampa di apertura della fiera del libro di Francoforte, la più grande al mondo. Un anno dopo, alla Buchmesse la battaglia è ancora aperta, ed è delle più impari: contro i giganti della tecnologia si devono battere i traduttori, gli scrittori, i giornalisti, i fotografi, gli artisti, professioni perlopiù terribilmente sottopagate e senza garanzie, sottolinea Francesca Novajra, presidentessa del Ceatl, l’European council of literary translator’s association (incarico non stipendiato), cui fanno parte Aiti e Strade per l’Italia. La risposta dell’Unione europea è per loro del tutto insoddisfacente (e quella dell’Italia ancora di più), per esempio l’Ai Act prevede che le società di intelligenza artificiale debbano dichiarare solo il 10%
Novajra explains how transparency is needed: 'transparency to know which texts have been produced with artificial intelligence, transparency about the texts on which the Ai has been trained, and transparency about the output, to calculate the licences'. The battle against AI is not only an economic one, because of the impact it has on the publishing industry and on jobs related to writing, it is a political, social battle - AI influences cognitive processes and power relations - and also a cultural one, because it also determines the content and the way cultural products are generated and distributed: it is an ethical battle.
There are the cognitive damages of the abdication of thinking, of writing, that begin to emerge, 'Ai is dangerous for critical thinking, and it is so in a very subtle way,' Novajra observes, 'besides introducing sexist, racist, homophobic, etc. errors and prejudices into texts, as Tiziana Catarci, professor of Computer Engineering at the Sapienza University of Rome, explains well, since Llm learns from the statistical analysis of the data it is given. If these data reflect mental schemes, cultural stereotypes or even prejudices of those who produced them, the machine takes them as models and re-proposes them'.
Also interesting was the small experiment by Waltraud Kolb, a translation researcher at the University of Vienna, who demonstrated the 'anchoring bias': she asked ten literary translators to translate a Hemingway short story (A Very Short Story), assigning five of them to translate from English into German from the original alone, and five others to translate from the original and a machine-generated 'pre-translation'. A seemingly very simple sentence: 'Luz sat on the bed', which, depending on how one reads it, can be understood as completed or unfinished (Luz sat or sat), was translated divergently among the translators in the first group, while the other five all chose the machine-generated translation.



