We have conducted a survey of RNA editing in human brain by comparing sequences of clones from a human brain cDNA library to the reference human genome sequence and to genomic DNA from the same individual. In the RNA sample from which the library was constructed, approximately 1:2000 nucleotides were edited out of >3 Mb surveyed. All edits were adenosine to inosine (A-->I) and were predominantly in intronic and in intergenic RNAs. No edits were found in translated exons and few in untranslated exons. Most edits were in high-copy-number repeats, usually Alus. Analysis of the genome in the vicinity of edited sequences strongly supports the idea that formation of intramolecular double-stranded RNA with an inverted copy underlies most A-->I editing. The likelihood of editing is increased by the presence of two inverted copies of a sequence within the same intron, proximity of the two sequences to each other (preferably within 2 kb), and by a high density of inverted copies in the vicinity. Editing exhibits sequence preferences and is less likely at an adenosine 3' to a guanosine and more likely at an adenosine 5' to a guanosine. Simulation by BLAST alignment of the double-stranded RNA molecules that underlie known edits indicates that there is a greater likelihood of A-->I editing at A:C mismatches than editing at other mismatches or at A:U matches. However, because A:U matches in double-stranded RNA are more common than all mismatches, overall the likely effect of editing is to increase the number of mismatches in double-stranded RNA.