La traduction automatique neuronale (TAN) et ses discours. Réflexions sur les implications discursives et sociétales des usages actuels de la TAN

While we hear more and more about the miracle of so-called 'neural' machine translation (NMT), all too little attention is paid to discourse on NMT. In this paper I analyse this discourse focusing on the hype around artificial intelligence in general and machine translation in particular, including the now often mentioned generative AI (based on generative pre-trained models, GPT). how it is actually used and how users perceive it. As a counterpoint to these analyses, I then focus on actual uses on NMT. Who are the users of machine translation systems today, and what are their practices? What do they know about machine translation and how do they perceive it? Based on data available in the literature and from recent surveys, I seek to define the current uses of NMT, in order to show the advantages that certain well-supervised uses can present, as well as the risks incurred by unsophisticated users. These risks are linked with well described effects of influence, persuasion and sometimes disinformation, in order to show that the risks incurred are still considerable but that what users need most is long-term projections, ethical discussions and AI commons.