I don't think there is an easy answer yet; it feels as if the heuristics are still very much emerging as the field matures. If you're using BirdNET data downstream in an occupancy model, I have found inspiration in this recent publication from Cole et al. 2022. It illustrates the reality of tradeoffs in true positives and false positives under varying thresholds and provides some precedent for how to deal with these. Recent work by Rhinehart et al. 2022 may also be relevant with respect to incorporating uncertain automated detection data into an occupancy framework.
One suggestion I've heard is to require a minimum number of detections within a timestep of interest, to minimize the chance that non-target noise is triggering false alarms. This still doesn't answer the stickiest part of the question -- how to choose what number of detections above what confidence level over what timestep?
Since it may be inadvisable to trust results without some degree of expert review, one method for conducting verifications quickly is to conduct "top-down listening" (where clips are sorted from highest to lowest confidence to quickly assess species presence). I'm dreaming about an eventual community science approach to solicit expedient BirdNET validations from local site experts, which would require an accessible platform and clear training heuristics guiding these field experts in how to verify (for example, how to verify songs vs. calls? And does it count as a correct detection if only a very small portion of the signal of interest occurs within the 3-second window?).