I suspect that is using one of the off the shelf NLP chat bots and speech recognition, perhaps from a platform like Hugging Face, since you are allowed to chat about anything you want and have a dozen or so languages enabled. While really cool, as others have pointed out in complex responses received, I suspect it is a long way for being useful to a language learner because the user has no way of knowing if what the chatbot is saying makes any sense. For example I used keyboard entry to speak in Japanese with it. The bot said hello and asked my name. I replied with my name in katakana and the desu (I am fragment) in hiragana. The bot responded in a way that just smashed those together as my name grammatically which is incorrect.
Perhaps the language models can get so good some day one could just prompt a model to "teach me Japanese through conversation from the beginning", but for now I think it is difficult enough to make an engine for a single language pair. If one is far enough along in a language to be conversational, it is low-cost, or even free, to find an actual language exchange conversation partner.