We examine the recent COMPASS data on the spin structure function g1 singlet. We show that it is rather difficult to use the data in the present form in order to draw conclusions on the initial parton densities. However, our tentative estimate is that the data better agree with positive rather than negative initial gluon densities.