Although efficient methods exist to assemble synthetic oligonucleotides into genes and genomes, these suffer from the presence of 1-3 random errors/kb of DNA. Here, we introduce a new method termed consensus shuffling and demonstrate its use to significantly reduce random errors in synthetic DNA. In this method, errors are revealed as mismatches by re-hybridization of the population. The DNA is fragmented, and mismatched fragments are removed upon binding to an immobilized mismatch binding protein (MutS). PCR assembly of the remaining fragments yields a new population of full-length sequences enriched for the consensus sequence of the input population. We show that two iterations of consensus shuffling improved a population of synthetic green fluorescent protein (GFPuv) clones from approximately 60 to >90% fluorescent, and decreased errors 3.5- to 4.3-fold to final values of approximately 1 error per 3500 bp. In addition, two iterations of consensus shuffling corrected a population of GFPuv clones where all members were non-functional, to a population where 82% of clones were fluorescent. Consensus shuffling should facilitate the rapid and accurate synthesis of long DNA sequences.