Biodiversity offsets are seen as a policy mechanism to balance development and conservation goals. Many offset schemes employ habitat restoration in one area to recreate biodiversity value that is destroyed elsewhere, assuming that recovery is timely and predictable. Recent research has challenged these assumptions on the grounds that restoration implies long time delays and a low certainty of success. To investigate these assertions, and to assess the strength of empirical support for offset policy, we used a meta-analytic approach to analyze data from 108 comparative studies of secondary growth (SG) and old-growth (OG) habitat (a total of 1228 SG sites and 716 OG reference sites). We extracted species checklists and calculated standardized response ratios for species richness, Fisher's alpha, Sorenson similarity, and Morisita-Horn similarity. We modeled diversity change with habitat age using generalized linear models and multi-model averaging, correcting for a number of potential explanatory variables. We tested whether (1) diversity of passively and actively restored habitat converges to OG values over time, (2) active restoration significantly accelerates this process, and (3) current offset policies are appropriate to the predicted uncertainties and time lags associated with restoration. The results indicate that in the best case, species richness converges to OG reference values within a century, species similarity (Sorenson) takes about twice as long, and assemblage composition (Morisita-Horn) up to an order of magnitude longer (hundreds to thousands of years). Active restoration significantly accelerates the process for all indices, but the inherently large time lags, uncertainty, and risk of restoration failure require offset ratios that far exceed what is currently applied in practice. Restoration offset policy therefore leads to a net loss of biodiversity, and represents an inappropriate use of the otherwise valuable tool of ecosystem restoration.