1. Temporal changes in semi-arid ecosystems can include transitions between alternative stable states, involving thresholds and multiple domains of attraction, but can also include relatively continuous, symmetric and reversible shifts within a single stable state. Conceptual state-and-transition models (STMs) describe both types of ecosystem dynamics by including state transitions (plant community changes difficult-to-reverse without substantial input or effort) and phase shifts (easily reversible community changes) as consequences of management practices and environmental variability. Grazing management is purported to be the primary driver of state transitions in current STMs for North American grasslands, but there is limited empirical evidence from these grasslands showing that grazing can cause difficult-to-reverse transitions between alternate stable states. 2 .In a northern mixed-grass prairie in Wyoming, USA, we examined plant community responses to (i) long-term (33-year) grazing intensity treatments (none, light, moderate and heavy stocking rates) and (ii) 8 years of light or no grazing in pastures that were grazed heavily for the previous 25 years. 3. Long-term grazing treatments were associated with distinct, but not stable, plant communities. From year 22 to 33, heavier stocking rates decreased cover of dominant C3 grasses and increased cover of the dominant C4 grass Bouteloua gracilis. 4. Reversing stocking rates from heavy to light or no grazing resulted in reversal of changes induced by prior heavy stocking for dominant C3 grasses, but not for B. gracilis. For both groups, rates of change following grazing treatment reversals were consistent with rates of change during the initial years of the experiment (1982–1990). 5. Synthesis and applications. In a semi-arid rangeland with a long evolutionary history of grazing, different long-term grazing intensity treatments caused slow, continuous, and directional changes with important management implications, but did not appear to induce alternative stable states. For this and similar ecosystems, quantifying the time-scales and compositional gradients associated with key phase shifts may be more important than identifying thresholds between alternative stable states.