It makes more sense when viewed as a fancy autocomplete, not an intelligence. There's no intelligence behind it that is reading your statement and understanding your meaning. It's responding with text that mathematically likely matches some sort of reply that would fit your statement.

Your statement included Y and the algorithm landed on result that includes Y. There's no intelligence that could understand that you meant no Y.

That bullshit about the model getting fine tuned just means they are data mining you. It doesn't make the more LLM intelligent. All it does is add your data to their dataset of which the LLM can draw from for possible future replies. The fundamental limitations of the technology still exists.