追查人工智能制作的盗版歌曲，16年前YouTube推出的一款产品可能有大用场！

In the world of technology, sixteen years is an eon. That many years ago, Apple launched its first iPhone, and IBM created Watson. YouTube, which had just been acquired by Google, rolled out a groundbreaking tool that could identify copyrighted music within the videos that users uploaded to its site.

Now, in a remarkable indication of how much the world has changed since that time, YouTube has a new mission for its trusty copyright detection tool: to identify an expected deluge of songs composed by artificial intelligence.

On August 21, YouTube CEO Neal Mohan said that Content ID, the name of the tool, could play an important role detecting potential copyright infringement in A.I.-generated music.

“AI is ushering in a new age of creative expression, but it must include appropriate protections,” Mohan said in a blog post detailing YouTube’s plan for addressing A.I.-generated music on its site.

Mohan said the company will embrace A.I. wholeheartedly but responsibly. It will collaborate with artists and record labels to explore new ways to use A.I. in music, while also prioritizing protecting the creative works of artists, which includes continuing to develop its Content ID system.

But with so few guidelines and established best practices for the new era of generative A.I., YouTube will be in uncharted waters. As it puts its plan into practice, YouTube’s approach to policing A.I.-generated music on its platform, as well as its success and struggles in the effort, is likely to have an impact that goes well beyond its own website, according to experts.

It’s not easy to determine the song that trained the A.I.

YouTube initially developed Content ID as a result of a lawsuit from Viacom which alleged that the site was not doing enough to remove copyrighted material, such as a hit song playing in the background of a user’s video or a full album uploaded to the site. Content ID scans YouTube’s vast corpus of user uploaded videos, checking to see if a particular song is present. Copyright owners who use the system can then decide whether they want to have an infringing video taken down, or to leave it be and earn a cut of their advertising revenue the video racks up on YouTube.

Content ID has generated billions of dollars for the industry over the years, Mohan said in the post. “A new era of generated content is here, and it gives us an opportunity to reimagine and evolve again,” he said.

But detecting copyright in A.I.-generated material isn’t as easy as finding a video with a Beyonce song playing in the background. The current iteration of generative A.I. produces content based on pieces of things that already exist, so a song could be informed by dozens of voice cadences, notes sequences, and stylistic choices from artists in their own work. Style and voice can’t be copyrighted.

Rights laws come into play when the original recording trains an A.I. model or if an A.I.-generated song profits from the name and likeness of an artist. These issues spark a larger conversation about the rights of everyone involved in the production of music, experts say. Does the person who mixed the background music that informed A.I. have a claim? What about the artist whose voice was removed from a song? There are no clear guidelines on when A.I. infringes on copyright and who is owed if it does, and these issues are a long way from being solved in the courts. As a music industry leader with some of the most advanced audio recognition systems, what YouTube does next will set a precedent for who is owed and when, experts say.

It will be difficult for Content ID to claim exactly what media trained A.I.-generated music, said Jane Davidson, a technology and entertainment lawyer who previously worked as a talent manager for YouTube musicians. YouTube’s system will have to learn to recognize specific voices, notes or sequences in music that are similar enough to copyrighted material in YouTube’s catalog and flag them for rights owners to decide if they want to file a claim. A generated video’s metadata might also clue into the content that trained the A.I., Davidson told Fortune.

All of the major music companies have technology similar to Content ID, and they are all likely training their systems to recognize copyrighted material in generated music, experts agree. But when it comes to the social media companies, like Meta and TikTok, they will be following YouTube’s lead, said Samuel Tall, executive at Studio71 who has been using YouTube’s system and those of its competitors for a decade.

“Content ID came out of the gate really advanced and has been improved upon over time,” he said. The technology from other social media companies doesn’t compare, but Content ID still isn’t perfect, he said. “It’s really, really good, but it struggles with nuance, and there’s a certain amount of leakiness to it.” For example, it might not be able to differentiate a woman-led rock band covering a Paramore song from Paramore itself, he said. And it really needs to master nuance to detect copyright within A.I.-generated material. That said, other platforms don’t do a better job, he said. YouTube declined to comment.

Artists and other music industry workers are concerned A.I. will compromise their originality, Tall told Fortune. For musicians, that means their style and artistry. For producers, DJs and audio mixers, that means the skills they’ve built around creating music, which they bill for, he said.

Some artists have already had front row seats to A.I. stealing their thunder. In April, an anonymous TikTok user posted a song that used A.I. to mimic the voices of Drake and the Weeknd. The catchy song was available on TikTok, Spotify, Apple Music and YouTube, and it went viral before being taken down for copyright claims. It is unclear if Content ID flagged the song or if the artists’ teams reached out first.