Inland water ecosystems can undergo abrupt change transforming rapidly to a different state. Theory indicates that a detectable loss of resilience may occur in advance of some state changes. Using time series data, rolling window measures of increased variance and autocorrelation approaching one are key indicators of pending changes. We tested this theory with two types of whole lake manipulations - apex predator additions and nutrient fertilization. Adding largemouth bass (Micropterus salmoides) gradually to a lake destabilized the food web. Warnings of a regime shift to dominance by piscivorous bass were evident in time series of phytoplankton, zooplankton, and small fishes more than a year before the transition was complete. No warnings were observed in an unmanipulated reference lake. Whole lake inorganic nitrogen and phosphorus additions were used to study early warnings of phytoplankton blooms. We developed an automated alarm system based on continuously monitored phytoplankton pigments and dissolved oxygen percent saturation. We observed alarms in two fertilized lakes prior to blooms. More alarms were observed prior to large blooms while weak blooms produced few alarms. We tested if we could reverse the effects of nutrients in response to warnings. Transition to sustained blooms was reversed by halting nutrient additions in response to alarms. Overall, the lake manipulations demonstrated the potential of early warning statistics to detect loss of resilience. Prior knowledge and models as well as the availability of reference lakes greatly aided application of the approach.