Tree rings are thought to be a powerful tool to reconstruct historical growth changes and have been widely used to assess tree responses to global warming. Demographic inferences suggest, however, that typical sampling procedures induce spurious trends in growth reconstructions. Here we use the world's largest single tree-ring dataset (283,536 trees from 136,621 sites) from Quebec, Canada, to assess to what extent growth reconstructions based on these - and thus any similar - data might be affected by this problem. Indeed, straightforward growth rate reconstructions based on these data suggest a six-fold increase in radial growth of black spruce (Picea mariana) from ~0.5 mm yr-1 in 1800 to ~2.5 mm yr-1 in 1990. While the strong correlation (R2 = 0.98) between this increase and that of atmospheric CO2 could suggest a causal relationship, we here unambiguously demonstrate that this growth trend is an artefact of sampling biases caused by the absence of old, fast-growing trees (cf. "slow-grower survivorship bias") and of young, slow-growing trees (cf. "big-tree selection bias") in the dataset. At the moment, we cannot envision how to remedy the issue of incomplete representation of cohorts in existing large-scale tree-ring datasets. Thus, innovation will be needed before such datasets can be used for growth rate reconstructions.