Comment by metabagel
> 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.
Seems so, yes, but tests showed that the models were better at predicting the future (or any time past their cutoff date) when they were less aligned and still used percentages.