The most recent SH0ES measurement of the Hubble constant employs corrections of type Ia supernova magnitudes due to extinction in their host galaxies. These corrections are estimated using a probabilistic model which is trained on Hubble flow (z > 0.03) supernovae and extrapolated to the calibration galaxies (those with observed Cepheids), despite the fact that the latter are selected based on criteria favouring disky and dust-rich systems. We show that this standard approach underestimates the brightness of reddened supernovae in the high stellar-mass (M⋆ > 1010 M⊙) calibration galaxies. This can be traced back to the fact that for these galaxies, a low total-to-selective extinction coefficient (RB ∼ 3) is assumed, while for the low stellar-mass analogues a more standard RB ∼ 4 is adopted. We propose a minimalistic modification of the extinction model in the calibration galaxies in order to alleviate this systematic effect. The modification is twofold and it involves: (i) the same, Milky Way-like distribution of RB (with mean RB of 4.3 – consistent with the extinction curve used for colour corrections of the Cepheids – and scatter 0.4) and (ii) a modified shape of the E(B − V) reddening distribution while keeping the same effective slope of the supernova peak magnitude-colour relation and the same mean E(B − V) reddening as measured for supernovae in the Hubble flow. We show that this new approach yields a significantly better fit (ΔBIC = −11) to the calibration data and results in a lower value of H0. Our result is H0 = 70.5 ± 1 km s−1 Mpc−1 implying a reduction of the Hubble constant tension from 5.2σ to 2.8σ.