That's not true and its exactly the point - real life data is noisy, difficult to parse, and ultimately often ambiguous. Text is the distilled information. It's significantly easier to distilled text into a likely response than it is to distill into lessons a continuous data stream whose contents are partially controlled by your own actions.