The paper (http://resources.mpi-inf.mpg.de/DeepScreenSpace/) just says that
"The ambient occlusion contribution of a surfel to a pixel [..] as well as the influence radius [..] are computed using the point-to-disk form factor [Wallace et al. 1989]"
Wallace et al. 1989, A ray tracing algorithm for progressive radiosity
Does that mean that, in order to test for occlusion, rays are shot from a surfel to other surfels within its influence radius? That's just a speculation. Unfortunately the paper does not really explain how they implemented AO.