Abstract We present the results of a new study of the relationship between infrared excess (IRX ≡ LIR/LUV), ultraviolet (UV) spectral slope (β) and stellar mass at redshifts 2 < z < 3, based on a deep Atacama Large Millimeter Array (ALMA) 1.3-mm continuum mosaic of the Hubble Ultra Deep Field. Excluding the most heavily obscured sources, we use a stacking analysis to show that z ≃ 2.5 star-forming galaxies in the mass range $9.25\le \log (M_{\ast }/\rm M_{{\odot }}) \le 10.75$ are fully consistent with the IRX–β relation expected for a relatively grey attenuation curve, similar to the commonly adopted Calzetti law. Based on a large, mass-complete sample of 2 ≤ z ≤ 3 star-forming galaxies drawn from multiple surveys, we proceed to derive a new empirical relationship between β and stellar mass, making it possible to predict UV attenuation (A1600) and IRX as a function of stellar mass, for any assumed attenuation law. Once again, we find that z ≃ 2.5 star-forming galaxies follow A1600–M* and IRX–M* relations consistent with a relatively grey attenuation law, and find no compelling evidence that star-forming galaxies at this epoch follow a reddening law as steep as the Small Magellanic Cloud (SMC) extinction curve. In fact, we use a simple simulation to demonstrate that previous determinations of the IRX–β relation may have been biased towards low values of IRX at red values of β, mimicking the signature expected for an SMC-like dust law. We show that this provides a plausible mechanism for reconciling apparently contradictory results in the literature and that, based on typical measurement uncertainties, stellar mass provides a cleaner prediction of UV attenuation than β. Although the situation at lower stellar masses remains uncertain, we conclude that for 2 < z < 3 star-forming galaxies with $\log (M_{\ast }/\rm M_{{\odot }}) \ge 9.75$, both the IRX–β and IRX–M* relations are well described by a Calzetti-like attenuation law.