Successful application of external photon beam therapy in oncology requires that the dose delivered by a medical linear accelerator and distributed within the patient’s body is accurately calculated. Monte Carlo simulation is currently the most accurate method for this purpose but is computationally too extensive for routine clinical application. A very elaborate and time-consuming part of such Monte Carlo simulation is generation of the full set (phase space) of ionizing radiation components (mainly photons) to be subsequently used in simulating dose delivery to the patient. We propose a method of generating phase spaces in medical linear accelerators through learning, by artificial intelligence models, the joint multidimensional probability density distribution of the photon properties (their location in space, energy, and momentum). The models are conditioned with respect to the parameters of the primary electron beam (unique to each medical accelerator), which, through Bremsstrahlung, generates the therapeutical beam of ionizing radiation. Two variants of conditional generative adversarial networks are chosen, trained, and compared. We also present the second-best type of deep learning architecture that we studied: a variational autoencoder. Differences between dose distributions obtained in a water phantom, in a test phantom, and in real patients using generative-adversarial-network-based and Monte-Carlo-based phase spaces are very close to each other, as indicated by the values of standard quality assurance tools of radiotherapy. Particle generation with generative adversarial networks is three orders of magnitude faster than with Monte Carlo. The proposed GAN model, together with our earlier machine-learning-based method of tuning the primary electron beam of an MC simulator, delivers a complete solution to the problem of tuning a Monte Carlo simulator against a physical medical accelerator.