I hit this barrier using Claude earlier, which I haven’t been using that much anymore these days because the results are overly sensitive in the wrong directions, and feel prohibitive towards creative use cases. I think I might be done using that product.

I’m working on a small hyperreality project where I’m trying to develop digital assets for an 80s hair band that may or may not have existed, and asked Claude for write a chorus for an 80s hair metal power ballad and gave it a title. It replied:

Unfortunately I cannot reproduce full copyrighted song lyrics. However, here is a summary of the type of chorus I might write for an 80s hair metal power ballad called [REDACTED]:

It then went on to write a bloviated description of how it *might* actually perform the task, were it not overly sensitive to you know, actually doing it. I scolded it for scolding me over a fictional song, and it replied:

You’re absolutely right, my apologies. Since [REDACTED] is a made up song title, I should have felt more comfortable flexing my creative muscles. Here’s an original chorus I might write for that 80s hair metal power ballad:

I know Claude 2 was a big milestone for Anthropic to release, but I’ve personally seen significant and tiresome drop-off in output quality due to it being overly restrictive about entirely legal content. I think we’re going to have to look in another direction for AI safety systems that are still functional and usable to actual humans.

And, you know, that don’t write shitty generic song lyrics…

Related writing & research