Observation bias pervades data collected during aerial surveys of large animals, and although some sources can be mitigated with informed planning, others must be addressed using valid sampling techniques that carefully model detection probability. Nonetheless, aerial surveys are frequently employed to count large mammals without applying such methods to account for heterogeneity in visibility of animal groups on the landscape. This often leaves managers and interest groups at odds over decisions that are not adequately informed. I analyzed detection of feral horse (Equus caballus) groups by dual independent observers from 24 fixed-wing and 16 helicopter flights using mixed-effect logistic regression models to investigate potential sources of observation bias. I accounted for observer skill, population location, and aircraft type in the model structure and analyzed the effects of group size, sun effect (position related to observer), vegetation type, topography, cloud cover, percent snow cover, and observer fatigue on detection of horse groups. The most important model-averaged effects for both fixed-wing and helicopter surveys included group size (fixed-wing: odds ratio = 0.891, 95% CI = 0.850–0.935; helicopter: odds ratio = 0.640, 95% CI = 0.587–0.698) and sun effect (fixed-wing: odds ratio = 0.632, 95% CI = 0.350–1.141; helicopter: odds ratio = 0.194, 95% CI = 0.080–0.470). Observer fatigue was also an important effect in the best model for helicopter surveys, with detection probability declining after 3 hr of survey time (odds ratio = 0.278, 95% CI = 0.144–0.537). Biases arising from sun effect and observer fatigue can be mitigated by pre-flight survey design. Other sources of bias, such as those arising from group size, topography, and vegetation can only be addressed by employing valid sampling techniques such as double sampling, mark–resight (batch-marked animals), mark–recapture (uniquely marked and identifiable animals), sightability bias correction models, and line transect distance sampling; however, some of these techniques may still only partially correct for negative observation biases.