I suppose it's a bimodal distribution: very high effort and very low effort will both have near-perfect grammar, albeit for different reasons. Since the low effort applications dominate in volume, the discriminator learns perfect grammar to be a negative signal.

Law of unintended consequences I guess.