> For example, the unaligned model would give uncertain answers in terms of percentages, and the aligned model would use less informative words like "likely" or "unlikely" instead.

Percentages seem too granular and precise to properly express uncertainty.