Human perceptual decisions and working memory reports are systematically biased toward recent perceptual history, a phenomenon known as serial dependence. A prominent theoretical account proposes that serial dependence serves an adaptive function by exploiting temporal continuity in the environment to reduce noise and improve decision accuracy. Here, we tested this hypothesis in a large-scale mega-analysis compiling trial-level data across a decade of continuous-report serial dependence studies spanning multiple visual feature dimensions. Contrary to the prevailing normative assumption that serial dependence improves performance, we found that attractive serial biases consistently increase overall estimation error and deteriorate perceptual decision-making accuracy. These results challenge standard optimal-observer accounts of serial dependence and call for a re-evaluation of the functional role of history biases in human perception and working memory.