1. The dynamics of the ventilatory response to carbon dioxide inhalation were studied in ten healthy young men using four different inspired fractions of carbon dioxide (FI, CO2) in air (0.015, 0.030, 0.045 and 0.060) successively increasing and decreasing stepwise. 2. Seven such different progressions were performed for each subject and each of seven different durations of the steps (t) ranging between 0.1 (i.e. one ventilatory cycle) and 10 min ('steady-state' conditions). The overall duration of one test (T) was taken as the sum of the seven successive FI, CO2 steps (t) plus one step, t, of air breathing. Thus, the values of T ranged between 0.8 (i.e. eight ventilatory cycles) and 80 min. Three subjects were tested twice. 3. We measured, as a function of T, the magnitude of the loops formed by the curves PA, CO2-VE and the value of the highest ventilatory response (VE max) to each progression. For all ten subjects, both functions had two maxima, one for T values of 2.6 or 8.0 min and one for T values of 24 or 40 min, and one minimum at T equal to 12 min. 4. The same measurements were made on tidal volume-response curves (PA, CO2-VT) and ventilatory frequency-response curves (PA, CO2-f) and yielded the same results except for the ventilatory frequency-response curves, for which we only found a statistically insignificant single maximum for T values of 24 or 40 min. 5. The locations of the maxima in loop magnitude and VE max were similar in duplicate tests in three subjects, whereas the quantitative values of these variables showed wide differences. 6. We compared our results with what is expected from the current linear dynamic model of ventilatory control submitted to the same forcing function: the first maximum in the loop magnitude is predicted by the model, but the second is not. The model shows no peak in the evolution of VE max. 7. We conclude that controlled system dynamics, which are the only ones included in dynamic models of ventilatory control, cannot by themselves account for our observations, and that one should take into consideration the dynamics of the controlling neuronal network.