Using GPT-3 to Generate Fake Bob Dylan Songs
I supplied the AI system with the following prompt:
"You are a respected folk singer, famous for writing moving and meaningful songs that seem to capture the spirit of the times and the pulse of society. You have been invited to sing a new song for a group of important thinkers. Here are lyrics you have sung in the past:"
I then supplied the lyrics to three Bob Dylan songs ("Tangled up in Blue", "Like a Rolling Stone", and "Just Like a Woman"; here is the exact text I gave it: https://pastebin.com/YxaPF189). This text was used in order to "condition" the model— to give it a sense of context for what I wanted. The idea was to get GPT-3 to write me new song lyrics based on a new song title given by me in quotes.
In some cases, I supplied the song title several times and combined the output together. Other times I took the partial lyrics already generated for the same song title and included that as part of the conditioning text. But aside from that, this is pretty much direct output from the model without any intervention from me conditioned only on the Dylan lyrics. I tried a variety of different titles to see what would happen. Included here is the vast majority of the output I got, so there is very little cherry-picking on my end to find the best stuff.
To me, it definitely sounds reminiscent of Dylan in certain ways. The titles have a big impact on the lyrics, though— when I supplied the title "Putrid Smells and Acrid Smoke," the results sounded more like a death metal song than folk music. Other parts are truly bizarre and amusing, like its exhortations to get drunk and commit credit card fraud. There is certainly a lot of stuff about the moon being out, as well as many "Oh, baby" exclamations.
Anyway, here are the lyrics the AI generated: https://pastebin.com/qyBuTMB6