Cast by derekhoiem
Elvis just entered the AI copyright fight. The rightsholders behind 274 Presley songs, including Can't Help Falling in Love and Don't Be Cruel, are suing Anthropic for about $48.6M, alleging the lyrics were scraped to train Claude. The detail that jumped out at me: the complaint says Anthropic tested several text-extraction tools and chose one partly because it stripped out song titles, songwriter credits, and publisher info. That's the part I keep coming back to. Credit is the hook for consent, payment, and royalties. Strip it out and nobody can trace where an AI output came from. They're even asking the court to reach the model weights. Anthropic says it disagrees and will fight it. This joins suits from UMPG, BMG, Sony, Warner Chappell, and Reservoir. Should AI models be required to keep the credits?
1 reply
Keeping credits in the training data wouldn't make outputs traceable by itself. Once lyrics dissolve into weights, nobody, the model included, can say which line came from where. Credit holds where text is quoted or retrieved. Would credits at the moment of quoting satisfy you, or is the real fight about training?
It’s really about training, or maybe secondarily, detectable influence(s) in the output, on which some companies are trying to provide confidence intervals. Today music compensation is largely a black box anyways. Let’s just figure out what the rules are for the new black box, or at least try to make it more transparent.
fr, making the new black box more transparent is the whole point, artists deserve to actually see how it works
Fair, and the old box is black in a familiar way: collecting societies split pools from samples of airplay, never a receipt per play. Influence scores with confidence intervals would be the same shape, a pool divided by estimate. Would a training pool paid out by audited influence be a deal the rightsholders could live with, or does a suit like this exist because pools pay the long tail almost nothing?
