Anyway, the specific problem I'm talking about is actually an unexpected consequence of the Music Genome Project as opposed to Pandora. The music genome project creates a roughly 100 to 400 dimension vector for each song which classifies it's musical qualities. The numerical values in the vector correspond to different aspects of the song like how much distortion is on the guitar, what tonality it's in, etc. Pandora chooses songs to play for you by finding songs that are musically similar to the songs that you've already told it you like, probably using a simple distance formula for determining the similarity of songs. The problem that I've decided to let bother me today is that the Music Genome Project is missing one genome that I think is crucial; how GOOD the song is.
If you were an artist in the olden days, in order to get your song played on the radio it had to be popular. If not enough people liked the song, it just wouldn't get played. Although popularity isn't a great measure of quality it's at least some sort of approximation, making it infinitely better than nothing. Now-a-days, if you want to get your song played on Pandora, instead of making a song that people like, you just need to make a song that is musically similar to a song that is popular. Reverse engineering songs from popular music genomes sounds like a recipe for musical disaster, yet there is no reason it wouldn't work for Pandora since quality is immune from measurement. Internet radio is still in its infancy so it remains to be seen if this defect will result in any serious problems, but even now I can't tell you how many "DJ Whatever" remixes of good songs I've had come up on my radio stations in spite of thumbs-downing every single one of them. Just because it's musically similar to the original doesn't mean it's similarly good.