When an assistant seems worse, it may have changed, or people may simply expect more after using newer, better systems. Ordinary ratings mix those explanations.
The researchers examined votes on pairs of unchanged language models, software that generates text. Voters compared answers without knowing which model produced them, then chose the better answer or said both were bad.
For unchanged pairs, “both are bad” votes rose by 1.71 percentage points a month, while preference for the stronger model stayed steady. Ratings can fall because standards rise, not because a product worsens, so evaluations need stable reference points to tell the difference.