Most modern translation tools are language models and this was true even before the LLM chatbot explosion. The difference is they were trained on smaller (and less dubiously sourced) datasets and the output that was trained for was translations directly rather than conversations.