Sony's new system can identify original tracks inside AI-generated music

midian182

Posts: 11,734   +177
Staff member
What just happened? One of the many controversies surrounding AI-generated music is its use of copyrighted material without the consent of rights holders. Sony, however, has reportedly developed a solution: a technology capable of detecting copyrighted music within AI-generated tracks.

According to multiple reports, Sony Group's new system can identify the original works embedded in AI-generated songs and estimate how much each source contributed to the final output.

This attribution capability could allow rights holders to claim compensation when their material influences synthetic tracks. That would make it a potential breakthrough in one of the most contentious areas of generative AI.

The technology appears to build on neural fingerprinting and training-data attribution techniques designed to trace how generative models learn from existing recordings.

Sony researchers have previously explored methods to identify which audio files most influenced a generated piece, making attribution traceable even when the output isn't a direct copy.

Industry collaborations have also explored similar approaches: Sony Music and Universal Music Group partnered with research lab SoundPatrol to deploy neural fingerprinting tools capable of detecting the influence of original human-created music within AI content.

If the new system performs as intended, it could help address a growing legal and ethical crisis around AI music. Record labels and artists have spent the past two years battling AI track generators accused of training on copyrighted material without permission, while platforms have struggled to detect violations at scale. Sony alone has pushed to remove tens of thousands of AI-generated tracks mimicking major artists, highlighting the scope of the problem.

Also read: Spotify removes AI-generated song falsely attributed to country singer who died in 1989

Beyond enforcement, attribution technology could reshape the AI music industry. By measuring how much a generated song relies on specific source material, it may enable licensing frameworks or revenue-sharing models rather than outright bans. That shift would align with the industry's broader push toward "ethical AI" systems trained on licensed content and designed to compensate creators.

The challenge now is adoption. Detection tools must work across streaming platforms, content libraries, and AI generation services, all while remaining accurate even as models improve.

Permalink to story:

 
It's only fair for those with the original ideas to get compensated.

During this Trump administration everyone manipulating A.I will try to get away with anything as there's absolutely no oversight and no A.I regulation of any type. It's only natural for people affected to find a way to put some kind of order to the situation since this current administration couldn't care less.
 
Just curious, but I wonder what happens if that tool is ran on human produced music? Would be worth testing at least.
Bet the tool they're using is also AI. Which will lead to another arms race as AI generated music is built with tracks to fool ai detection ai.
It's only fair for those with the original ideas to get compensated.

During this Trump administration everyone manipulating A.I will try to get away with anything as there's absolutely no oversight and no A.I regulation of any type. It's only natural for people affected to find a way to put some kind of order to the situation since this current administration couldn't care less.
Much of this data was taken in during the gold rush in 2023-2024. Biden was president in 2024. But that doesn't fit into the "orange man bad" narrative.
 
These companies can't sue one another for their theft so they go after the common man. These copy right tears are nothing but a scam as they're all stealing anything and everything that they can get their grubby paws on.
 
I spent over an hour trying to get Gemini to show me the best Japanese music and songs, focusing on composition, sound design, performance, and other sound qualities, not just popularity. I've heard over 100 tracks, and on a scale of 0 to 100, they score at most a 30, with most scoring under 20. This music video, however, represents the peak of Japanese achievement in visuals and sound.

The visuals are top-notch, but the sound is very weak. I don't want to sound harsh towards Japan(which has some of the most intelligent people in the world), but sound quality seems weak across the globe, with only a few Europeans exceptions. It's as if Europe is the Hollywood of sound design. Songs generated by AI are definitely orders of magnitude better than the best Japanese songs I've encountered. They are so weak in terms of sound. And ultimately, this is likely the main reason they are scared of the competition and want to suppress it.

Peak Sound (and Visuals) of Japan ->
and ¯\_(ツ)_/¯
 
People will machine learning models to generate slop, other people will pay machine learning models to delete the slop, then finally people will pay machine learning models to agentically find and download whatever is left. Now I see what all the datacenter growth is about... stupidity.
 
It's only fair for those with the original ideas to get compensated.

During this Trump administration everyone manipulating A.I will try to get away with anything as there's absolutely no oversight and no A.I regulation of any type. It's only natural for people affected to find a way to put some kind of order to the situation since this current administration couldn't care less.
Bet the tool they're using is also AI. Which will lead to another arms race as AI generated music is built with tracks to fool ai detection ai.
Much of this data was taken in during the gold rush in 2023-2024. Biden was president in 2024. But that doesn't fit into the "orange man bad" narrative.

Yes, a lot of people are unhappy with both the current and former administrations. We get it. Please stay on topic.

On topic: I find this really interesting if this attribution system really works at scale. Instead of endless, and probably fruitless, takedowns and lawsuits (because courts are slow as molasses), we’d get measurable influence, structured licensing, and revenue splits.

It would be great to see tech that could help separate real creative value from low-effort AI slop and keep the industry from hollowing out.

Whoever controls attribution controls the economics of music. I think that could reshape the entire industry more than regulation ever would.
 
It is not possible to do what this tool claims.

The training phase uses thousands if not millions of sound recordings. It's not 5 tracks in, 1 track out.

At most it will detect patterns in the songs and say "oh the beat used in this song is very similar to a beat used by Snoop Dogg". But that's not enough to claim copyright infringement.
 
Good stuff. I hope to see AI used to identify every byte of date that any AI used to create a thing.
It would not of course scann everything people post. But those wealthy companies using AI will have to share with small creators who could easily use the hunter AI to identify their work.
Furthermore, this system would benefit the creator greatly as well. Each creator and artist would be more than happy to pay a percentage of the money they receive from settlements.
If small artists cannot make money due to AI, we will kill art as it is. There will be very little enthusiasm to create. And it is important that top AI companies share with everybody whose work is used for training
 
My guess is the tool is inaccurate but they will use it anyways because of the implication that almost no creator will challenge a mega corpo in court

Prepare for news of legitimate artists getting flagged. This is going to be the AI essay detectors all over again.
 
Back