Update: let's juxtapose next another ironic contrast in AI from the NY Times...
This reprises AI conversations early in exploring AI around 18 mo. ago: two short ones here and here on Medium, and a long one here on my original blog's feature post on AI.
A short conversation about song lyrics, licensed lyrics sites, AI training, and the awkward asymmetry between what AI systems may learn from and what they may reproduce.
The request
I asked for the lyrics to Alejandro Sanz’s “Corazón partío” after finding a YouTube version. The reply was that the full lyrics could not be reproduced because they are copyrighted, though a brief excerpt was offered:
“Tiritas pa’ este corazón partío…”
“¿Quién me va a curar el corazón partío?”
In English, the refrain means roughly:
“Bandages for this broken heart…”
“Who is going to heal my broken heart?”
Why can lyrics sites show them?
The obvious question followed: if sites such as AZLyrics can publish complete lyrics, why can’t an AI assistant? The answer is that AZLyrics and similar services may operate through licensed lyrics aggregators—especially Musixmatch—which arrange permissions and royalty reporting with publishers and rights holders.
An AI assistant, by contrast, may be governed by a more conservative policy: it can quote or translate a short excerpt for discussion, but not output an entire copyrighted lyric or a complete line-by-line translation. That is a platform policy boundary, not necessarily a judgement that the lyrics site is acting unlawfully.
The irony
There is an unmistakable asymmetry. Large AI systems may ingest vast quantities of web material during training, while the same system may decline to reproduce a complete copyrighted text when a user asks for it.
The distinction is that training is generally framed as statistical learning from large corpora, whereas output on request is closer to reproducing or distributing the work itself. Legal systems are still testing where fair use, licensing, market harm, and transformative use begin and end. More than a hundred lawsuits in the United States alone are probing whether training generative AI on copyrighted works is lawful.
So the apparent paradox remains: an AI may have “seen” a lyric during training, yet be prevented from handing over the full text or translation. Meanwhile, a licensed lyrics site can display the same lyric because it has arranged the necessary rights.
Wrap-up
The sensible conclusion is not that either position is wholly right or wholly wrong. Copyright exists to protect creators and publishers; AI development depends on access to vast amounts of human culture; and users legitimately want to read, translate, and understand the words of a song. The tension between those truths is unlikely to disappear soon.
“The test of a first-rate intelligence is the ability to hold two opposed ideas in mind at the same time, and still retain the ability to function.”
— F. Scott Fitzgerald, The Crack-Up (1936)
Sources and context: Lyrics are protected separately from recordings; licensed display is commonly handled through aggregators such as Musixmatch and LyricFind. The legal status of AI training on copyrighted works remains contested; see Reuters’ overview of the U.S. fair-use litigation, “AI copyright battles enter pivotal year”.
No comments:
Post a Comment