The output of a model is only ever a close approximation of the training data, so asking these systems questions about events (in any language), that happened AFTER training, will mean no meaningful reply could be given. Searle perfectly describes the kind of systems we have today, but still ignores how answers are generated by people. We are not doing syntax and symbol manipulation according to rules; we are "thinking" in terms of concepts, even fragments of concepts, not directly expressed in the text. It's only after we have thought about it, and have some idea to express, that we "output" according to grammatical norms.