Context: Dark cloud chemical models usually predict large amounts of O2, often above observational limits. Aims: We investigate the reason for this discrepancy from a theoretical point of view, inspired by the studies of Jenkins and Whittet on oxygen depletion. Methods: We use the gas-grain code Nautilus with an up-to-date gas-phase network to study the sensitivity of the molecular oxygen abundance to the oxygen elemental abundance. We use the rate coefficient for the reaction O + OH at 10 K recommended by the KIDA (KInetic Database for Astrochemistry) experts. Results: The updates of rate coefficients and branching ratios of the reactions of our gas-phase chemical network, especially N + CN and H3+ + O, have changed the model sensitivity to the oxygen elemental abundance. In addition, the gas-phase abundances calculated with our gas-grain model are less sensitive to the elemental C/O ratio than those computed with a pure gas-phase model. The grain surface chemistry plays the role of a buffer absorbing most of the extra carbon. Finally, to reproduce the low abundance of molecular oxygen observed in dark clouds at all times, we need an oxygen elemental abundance smaller than 1.6E-4. Conclusion: The chemistry of molecular oxygen in dense clouds is quite sensitive to model parameters that are not necessarily well constrained. That O2 abundance may be sensitive to nitrogen chemistry is an indication of the complexity of interstellar chemistry.Read less <