Auditory filters broaden with increasing level. Using a recently developed method of fitting filter shapes to notched-noise masking data that explicitly models the nonlinear changes in filter shape across level, results at 2 kHz from 9 listeners over a wide range of levels and notch widths are reported. Families of roex(p,w,t) filter shapes lead to models which account well for the observed data. The primary effect of level is a broadening in the tails of the filter as level increases. In all cases, models with filter parameters depending on probe level fit the data much better than masker-dependent models. Thus auditory filter shapes appear to be controlled by their output, not by their input. Notched-noise tests, if performed at a single level, should use a fixed probe level. Filter shapes derived in this way, and normalized to have equal tail gain, are highly reminiscent of measurements made directly on the basilar membrane, including the degree of compression evidenced in the input-output function.