Peer review is often defended as the mechanism that makes published science reliable, a description its own performance does not support. Studies in which deliberate errors are inserted into manuscripts find that reviewers detect a minority of them. Reviewers are unpaid, working in time taken from their own research, and are not given the data or code that would let them check what they are assessing. What review does reasonably well is quite different: it improves the presentation of an argument, catches the claim that outruns its evidence, and filters out work that no competent person would publish. Treating it as a guarantee of correctness has placed a weight on the institution that it was never built to carry.