A hearing device (HD) incorporates a method for adaptive cancellation of acoustic feedback (AF). The method comprises generating an estimated feedback signal (EF) and subtracting the estimated feedback signal (EF) from the microphone signal (MS) before feeding it to a signal processor (SP) providing the primary hearing device function. The estimated feedback signal (EF) is generated in an adaptive filter (FE1), which is controlled using a least-mean-square algorithm, which operates on an error signal (E) and a reference signal (R). The algorithm may behave erroneously if the feedback path changes while a signal with low HF content, such as speech, is received. In this case, the hearing device (HD) will not be able to quickly adapt the HF characteristic of the adaptive filter (FE1) to the changed conditions. The adaptive filter (FE1) may thus have an incorrect HF gain when a subsequent signal with high HF content is received. This may lead to whistling or, alternatively, to an unwanted suppression of the subsequent signal. The problem is solved by modifying a filter function (H) applied to the error signal (E) and to the reference signal (R) in dependence on estimated relative amounts of high- and low-frequency signal content in the microphone signal (MS).