Why AI Attribution Matters for the Music Business — And How It Can Become a Reality (Guest Column)
As the industry enters a new era, creating workable attribution systems will benefit everyone in the equation — from artists to AI music companies.
Black News 
AI attribution is the key to unlocking the limitless opportunities available in a music world that seeks to embrace generative AI technology. It can calm artists’ fears around compensation and unauthorized use. It can reduce litigation risk and increase profits for platforms that provide generative AI music products. And it can give rights owners more certainty on license scope, plus more nuanced and enhanced revenue shares. In short, if designed and deployed the right way, AI attribution can be a boon to the entire generative AI music ecosystem.
What is AI Attribution in the Music Industry?
In the context of generative AI and music, attribution refers to the process of attempting to trace which training inputs contributed to a given AI-generated output (and in some cases, how much each input contributed).
Related
For artists, copyright owners and platforms alike, attribution offers what the industry has demanded since the early days of generative AI: transparency. By shining a light into what was once a black box, attribution gives copyright owners visibility into whether and how their works are being used by AI systems.
It also creates the potential for more customized compensation opportunities. Where licensing arrangements exist today, they generally take the form of upfront fees or revenue-sharing models that are not tied to the actual contribution of any specific work. Attribution, in theory, changes that calculus by enabling compensation to be linked to traceable impact.
The potential upside is industry-wide. For AI developers, attribution could make licensing discussions less contentious. It is much easier to negotiate with artists, labels and publishers when there is a credible way to offer visibility into how works are used and how value is tracked. More transparent systems could also lend greater comfort to investors mindful of the legal exposure of generative AI platforms in today’s climate. Plus, this creates an opportunity to have more targeted data about what music fans and music creators find most useful in generative AI products.
Importantly, attribution offers more legitimacy. Platforms can point to sourcing and compensation mechanisms that are more trustworthy and easier to build on. Music fans and creators, and copyright owners, are more likely to partner with or use generative AI music platforms if they believe the underlying system can explain where the value came from and whether the relevant rights were compensated.
The Attribution Landscape Today
In recent years, a range of technical methods have been proposed to investigate whether particular training materials may have influenced a given AI-generated output. Some methods compare generated outputs to candidate training materials in a separate database to identify similarity or proximity. Other methods examine the AI model itself, using signals from the model’s parameters to estimate whether, and to what extent, particular training materials contributed to a given output. Watermarking training content is another method, where the presence of a watermark in the output has been suggested to indicate which specific training materials contributed to it.
At present, however, none of these methods yield answers about influence with certainty. Some rely on probabilities, while others may find correlations, which does not necessarily mean causation. Often the results depend on underlying assumptions and, in some cases, on access to technical information about the AI model that may not be available in practice. Some options rely on significant computing power, which can be expensive.
Because current attribution methods carry various limitations, it may take time before any single approach is widely adopted across the industry, or before effective hybrid approaches that combine the best of these methods emerge. That doesn’t mean that these technologies shouldn’t be used. As long as everyone understands the limitations, the benefits of these technologies merit development because the better they get, the better the commercial opportunities will be in the music space. And the technology won’t improve at the pace users will expect without actually using them, perhaps in limited data environments (“sandboxes”) to mitigate risk. In the meantime, as these technologies evolve, several questions matter from a commercial contracting perspective.
Where the Rubber Meets the Road: Questions Worth Asking in Your AI Music Deals
Capability – Before the attribution data is used to support payment or other rights-sensitive decisions, the parties should first ask: Is the attribution system capable of identifying the kind of influence you intend to compensate? This is important to ask because, in practice, an attribution tool’s capabilities (and its limitations) become part of the parties’ commercial bargain. Consider a hip-hop producer who licenses his catalog to an AI platform under an agreement intended to compensate him when the model draws on his unique production style. The model later generates a track in a new genre that incorporates his distinctive production signatures, but the output sounds nothing like any particular recording in the producer’s catalog. An attribution tool that looks for close resemblance to recordings in a reference database may return no close match to the producer’s catalog (or low confidence), potentially leaving his contribution uncompensated despite the parties’ intent to the contrary.
Auditability – If attribution data will inform payments to copyright owners, there should be visibility into how the system works and how the results are produced. What documentation exists to describe the attribution methodology? Are independent audits permitted or even possible with respect to these technologies?
Liability – Current attribution technologies are still developing and, like any technology, are not immune from error. If the attribution system over-credits, under-credits, or fails to trace the influence or credit at all, what are the consequences? Who should bear the cost of investigation and resolving these claims? Is there a process for dispute?
Data ownership – the data generated about how often each work in a training set influenced an output, in what context, to what degree, and other considerations could have secondary value for everyone in this ecosystem. The data can reveal which training content is commercially relevant or stylistically influential, which may be useful for A&R professionals in picking which artists and songwriters to sign; copyright owners and generative AI music platform developers in what content should be most valued in licensing; and to anyone interested in learning more about what music fans and creators find most relevant in their worlds. Who owns this attribution data and any derived analytics? Can this attribution data be leveraged for other purposes?
While AI attribution technology may feel like a nuts-and-bolts topic, it is worthwhile for the various music and AI stakeholders, creators and technologists alike to engage with its development and work cooperatively to maximize its efficacy. Implementation of reliable attribution technology could be a boon to the entire music and AI ecosystem, helping the commercialization market mature and bringing both more certainty and higher revenues to its constituents.
Adrian Perry is a partner at global law firm Covington & Burling, co-chair of its Entertainment and Media Industry group, and a driving force behind the firm’s artificial intelligence transactional and advisory work.
Nicole Canales is an associate at Covington & Burling who advises on transactional matters across the firm’s technology and music industry practices.
