Recent data on large-$p_T$ $\overline{p}p\to\gamma X$ and $pp\to\gamma X$ at $\overline{p}p$ collider, ISR and fixed target energies are analyzed using complete next-to-leading $(O(\alpha_s^2))$ corrections. We employ physical scales $\mu=M=p_T$ and use three parton distribution sets differing essentially in the softness of the gluon. With $\mu=M=p_T$ we find that, on the whole, the data favor a distribution of intermediate softness, in contrast to optimization procedures which lead to a very soft gluon distribution. We discuss theoretical uncertainties related with the photon bremsstrahlung contribution and with the variation of the scales ($p_T/2\leqslant\mu=M\leqslant 2p_T$). We conclude that inherent ambiguities in the theory lead to significant uncertainty in the gluon distribution.