Not another rant on ethical training - a solution to AI music's attribution problem
Not another rant on ethical training - a solution to AI music's attribution problem
Ultimately this is about where and how the money flows
I wrote a medium spicy piece about generative AI x music last week. Tl;dr is that I think the democratization of music creation is a net good for humanity. Startups like Suno and Udio are leveraging state of the art models to generate music, enabling users to create songs with simple text prompts. More people than ever before will soon be able to create really great music, and that is undeniably awesome.
BUT, I do think it's important to acknowledge that we're standing on the shoulders of giants. Often, these AI models are trained on vast datasets of copyrighted music, raising important questions about data attribution, and therefore, compensation.
The Problem
When an AI model generates a new piece of music, it is essentially remixing and recombining elements from the millions of songs in its training data. But there is currently no clear way to trace which specific songs influenced the generated music. This is problematic because it means there is no mechanism for the original artists and rights holders to be credited or compensated when the AI-empowered music creators go on to commercialize their work.
Some argue this is a fair use of the training data, since the AI is not "directly" copying songs. After all, doesn't every human musician use all of their accumulated influences (whether consciously or subconsciously) to create "new" music?
Others take the opposing view, stating that nonconsensual use for training violates copyright law, as the AI-enabled musicians are deriving commercial benefit from copyrighted works without permission or payment.
Irrespective of which side of the argument you're on, I think both sides would agree that the ability to attribute AI-generated music to it's upstream creators is worthwhile - whether it's for auditing models and sending takedown requests (doomers) or for unlocking new licensing models for consensual training (maxis).
Okay, if both sides want it, why doesn't it exist?
Historically, it has been computationally hard to solve the attribution problem at scale (not to mention all the baggage the traditional music industry brings with it's longstanding metadata / rights management issues).
People way smarter than I am have written at length about the latter, so let's talk about the former—we might finally have a solution to solve data attribution at scale.
In December 2023, Junwei Deng and Jiaqi Ma from the University of Illinois wrote a paper titled Computational Copyright: Towards A Royalty Model for AI Music Generation Platforms.
They applied an existing ML approach called data attribution to music. The idea is to systematically measure how much each song in the training data contributed to a particular output. They used techniques like influence functions to quantify how the generated output changes when a particular training song is removed.
For example, if removing song X from the training data substantially decreases the probability of the model generating song Y, then song X can be said to be highly influential on song Y. By applying this type of analysis across the full training dataset, it's possible to create an "attribution score" for each training song in relation to each generated song.
These attribution scores could then be used as the basis for a more granular and transparent payout model. The owners of the songs with the highest attribution scores would receive the largest payouts. These payouts could be made on a per-stream or per-use basis, depending on the consumer application layer.
What would this unlock?
I think putting this into production could be really interesting. Solving the attribution problem for music could unlock better creative and financial outcomes for everyone involved. Some examples just to cement how important this could be:
Artists could choose to provide their work for training, and get paid a "royalty" every time someone uses an AI tool that is trained based on their work—opening up new revenue streams for them and new business models for startups providing high-quality training data to AI companies.
AI-generated music creators will strongly accelerate an existing trend - the decentralization of rights. Solving attribution will let startups benefit through a decentralized, demand-driven licensing approach, removing traditional blockers and enabling faster, more cost-effective innovation. Founders can go from 0-1 without having to worry about being sued by the labels.
There would be greater transparency around what data is used to train AI models. Artists and labels could audit the datasets and attribution scores.
Smaller artists could potentially benefit, as the attribution scores would be based on actual influence, not just popularity. If an indie artist's song is particularly influential on an AI model, they would be compensated accordingly.
AI-based sampling and remixing could become more fluid and open, as the attribution scores would provide a clear basis for credit and compensation.
The rise of generative music is inevitable. As technical barriers to create music are obliterated, we're going to have an unprecedented amount of people with the ability to create music. This could be music's Canva moment. Solving the data attribution problem could help kickoff a virtuous cycle: More Creation -> Accurate Attribution -> Fair Compensation -> Even more Creation. Psyched to see what's next.
Yash