Training multiple models might lead to different emerging flavors of synthetic interlingua(s). This might translate to humans being familiar with a host of them, with time investment driven by the multi-polar dynamics of the associated AGIs. There's also a question of bridging questions. Humans comunicating with augmented humans through a bridger language, before augmented humans comunicate with AGI through a different one. Beyond a cliche SF trope, it's not entirely clear how those intermediate entities would be like, how fast would humanness fade across the spectrum, etc.
In such a bridger chain of languages, there would be a problem of information being lost in translation, at the intermediate points of "bilingual" speakers along the chain. However, in the vanilla formulation, such phenomena might be less likely, because the very premise is that both sides use one shared language as a common ground for their interaction.