How often does automatic annotation give results that are consistent with manual annotation?

FAQ tags: 

n general, electronic annotations are rarely incorrect, as they are annotations to very high-level GO terms. For example, the GOA group at EBI reports:

Usually manual annotation simply provides deeper-level terms in GO. In 93% of cases GOA's electronic annotation is in the same GO lineage as the manual annotation. Some users have used our manual annotation to assess the quality of their automatic GO annotation techniques. They have found a few manual annotation errors by Proteome Inc. but no errors (so far) of manual annotation by Swiss-Prot staff have been reported to GOA. A few InterPro2GO errors have been reported but not very many. So, in general, our electronic techniques are very accurate, and are sometimes based on manual annotation. For example, Swiss-Prot keywords are usually manually annotated to Swiss-Prot entries; by using a mapping of Swiss-Prot keywords to GO, GOA inherits the high quality of Swiss-Prot manual annotation.

Text from Clark, et al., 2005

"The quality of electronic annotation has recently been assessed in some detail (Camon et al., 2005). This research found that in the worst case scenario, the generation of electronic annotations using the interpro2go, spkw2go, and ec2go mapping files precisely predicted the correct GO term 60% to 70% of the time, with the remainder of the predictions being to insufficiently specific GO terms. The high precision was found to be due to the basing of electronic annotations on manually curated mapping files. Curators noted that it was more important for database curation to be accurate than to have complete coverage, and the figures above demonstrate that this is the tendency with electronic annotation."