Which kind of experts in which loops? Redefining the relationship between translators, data, and models

MT and generative AI have led and will lead to new translation practices. On the one hand, the language industry requires experts-in-the-loop who coordinate AI services and turn mechanically produced text into high-quality output (Txabarriaga 2024). On the other hand, many human translators work as data labellers via platforms for extremely low rates in uberized conditions (Fırat, 2021). These circumstances ignore that the long-term risks of decreasing high-quality human data in Large Language Models have been highlighted (Simard, 2014; Briesch et al., 2023; Martínez et al., 2023): Less than ten generations of recycled synthetic training data lead to complete degradation and/or a dramatic loss of diversity in the output. New (business) models are certainly called for! We investigate to what extent quality of the output increases when a GPT-like model is trained with a dataset consisting of multiple translations per source text – compared to a model trained with only one translation per source text. We expect to see a dramatic improvement not only in the quality of translations of the same kind of texts present in the training data, but also regarding emergent aspects. If this leverage effect arises, i.e., that an extremely high increase in knowledge results in the model with relatively little, high-quality data, this logic could also be applied to other areas (programming, graphic design, etc.) where there is more than one correct solution to a problem. Human, manual work would increase in value and long-term dangers of training with (purely) synthetic data would decrease. If translators can monetise alternative translations of existing translations, the resulting business model raises a number of legal, economic, but also practical issues, and allows for a potential redistribution of the economic value associated with translators’ work.