ABSTRACT We examine the nature of kpc-scale clumps seen in high-redshift galaxies using a suite of cosmological simulations of galaxy formation. We identify rest-frame UV clumps in mock HST images smoothed to 500 pc resolution, and compare them with the intrinsic 3D clumps of young stars identified in the simulations with 100 pc resolution. According to this comparison for the progenitors of Milky Way-sized galaxies probed by our simulations, we expect that the stellar masses of the observed clumps are overestimated by as much as an order of magnitude, and that the sizes of these clumps are also overestimated by factor of several, due to a combination of spatial resolution and projection. The masses of young stars contributing most of the UV emission can also be overestimated by factor of a few. We find that most clumps of young stars present in a simulation at one time dissolve on a timescale shorter than ∼150 Myr. Some clumps with dense cores can last longer but eventually disperse. Most of the clumps are not bound structures, with virial parameter αvir > 1. We find similar results for clumps identified in mock maps of H α emission measure. We examine the predictions for effective clump sizes from the linear theory of gravitational perturbations and conclude that they are inconsistent with being formed by global disc instabilities. Instead, the observed clumps represent random projections of multiple compact star-forming regions.