Bots aren’t great listeners

Friday, May 22, 2020

The Washington post has a great piece ("Copyright bots and classical musicians are fighting online. The bots are winning.") that illustrates a substantial flaw of copyright filters:

These oft-overzealous algorithms are particularly fine-tuned for the job of sniffing out the sonic idiosyncrasies of pop music, having been trained on massive troves of “reference” audio files submitted by record companies and performing rights societies. But classical musicians are discovering en masse that the perceptivity of automated copyright systems falls critically short when it comes to classical music, which presents unique challenges both in terms of content and context. After all, classical music exists as a vast, endlessly revisited and repeated repertoire of public-domain works distinguishable only through nuanced variations in performance. Put simply, bots aren’t great listeners.

It is well known that copyright filters cannot recognise the context of a particular use of a copyrighted work, which pretty much disqualifies them when it comes to determining if a particular use is lawful or not. The WaPo article shows that the problem runs much deeper than that. State of the art content recognition technologies (the WaPo article mainly discusses YouTube’s ContentID and Facebooks Rights Manager) are also incapable of reliably differentiating between different recordings of a work.

This is even more evidence for the fact that automated content recognition technology is simply not up to the job that proponents of Article 17 of the DSM directive have envisaged it to play. At the very minimum this means that content matches involving classical music must always be subject to human review before they can result in blocking or takedown actions.