On May 31, 2023, the US Copyright Office hosted a listening session on artificial intelligence (AI) and copyright. Below are notes from KEI’s oral statement.
Copyright Office Listening Session on AI and Copyright
KEI Statement
May 31, 2023
My name is James Love. I’m speaking today on behalf of Knowledge Ecology International, a non profit organization that works on intellectual property rights. The AI conversation today often features polarized discussions about the opportunities and benefits of new AI generative models, on the one hand, and grave concerns about the potential for harm to society, on the other.
Without dismissing the potential for harm, I’ll focus on what I see as some practical issues, focusing on the topic of this session, music and AI.
As everyone will note, there are already amazing AI-related music services. Someone like me, who can’t play an instrument or sing, can create interesting music with AI tools. And people with actual talent in this area can use the tools to make good music.
Some artists want the Copyright Office to require rules for the 3 C’s, Consent, Credit and Compensation, for training data.
Consent
I personally hope that any requirements on consent involve opt-out rather than opt-in, but note that even that may prove difficult to enforce.
One reason why opt-in is problematic concerns monopolies. Some entities will have the time, money and legal and management resources to acquire large training datasets, while smaller entities won’t. There is a risk of monopolies or dominant platform scenarios emerging, similar to what we see now in Internet search or streaming platforms.
Sorting out who actually owns the rights in works in an opt-in system is complicated and resource intensive, and prone to mistakes. If this is done outside of Section 230, it can be very costly and burdensome to create large datasets.
There are substantial risks of monopolization in AI training data.
For music, one might also look at some of the entities that may already control sufficient rights in music to benefit from new licensing revenues when consent is required, and also to facilitate unwanted market concentration, amplifying the challenges that exist today with a high concentration in the areas of labels and platforms.
In general, AI services are better if they have more data. Of course, if people just want to kill off or slow down the development of AI, then copyright might seem like a tool to use, but this is probably a mistake, and more direct forms of regulations may be more appropriate to address non-copyright concerns.
While using copyright consent rules can slow down and temporarily degrade services, over time, the AI programs and well financed entities can eventually overcome this, for music and many other types of art. The leverage one expects from consent may not be as robust as some expect or hope.
It’s also useful to note that artists themselves don’t ask for consent when training for their own art. They often soak up as much as they can from others, and in some countries, like the United States, you can record music under a mechanical, or compulsory, license, without consent from the author. (17 U.S. Code § 115 – Scope of exclusive rights in nondramatic musical works: Compulsory license for making and distributing phonorecords).
Credit (Attribution)
Credit and attribution is important, for all innovative and creative efforts, and it’s often controversial. People argue over who should win Nobel prizes for the same discoveries, and who influenced or contributed to popular songs. One thing that can help is better metadata for recorded music. The current inconsistent standards and practices and problems with accuracy and completeness are well known problems in the industry. Society, globally and not just nationally, needs to create better incentives to improve metadata collection, curation, verification, sharing and use.
But here, AI can play a positive role in providing better metadata and better credits.
Compensation (remuneration)
Compensation is important, but is also going to be challenging. Having good metadata is important, and here too, AI may be helpful.
Litigation over copyright infringement is expensive and time consuming. AI programs can provide a relatively low cost and fast way to evaluate and resolve disputes over remuneration for works, whether the works used AI or not.
Transparency
If you want AI services to do a better job on metadata, credit and compensation, the training data needs to be robust, and policy makers, artists and others need to be able to audit and test the services, to ensure they are functioning in ways that are considered fair and useful, and with such diversity of interested parties and jurisdictions involved, some thought needs to be given to the best governance structure for such services, since they may replace in some cases and to some extent, judges and juries.
Fair use, licensing
To the extent that credit or and compensation can be resolved better, faster and cheaper, it may change our notions of what constitutes a fair use of works, or the role of licenses. (1) Significant exceptions may be in order to get the data to train services dealing with metadata, credit and compensation issues. (2) Artists may claim that even small claims are appropriate, given the feasibility of lower transaction and evaluation costs. (3) Society may want to alter the distribution of remuneration in ways that are inconsistent with contracts, and (4) there may be a considerable expansion of content generated and used by non commercial actors.
How will AI-generated music impact listeners?
To the extent that AI-generated music will be cheap or free to use, people will be using it in a variety of ways, including making their own music or making videos with AI generated music that don’t get take-down requests over a copyrighted song being used.
Listeners will be confronted with many fakes involving artists they love. Some listeners place considerable value on authenticity. It’s not as if the industry is not without its synthetic aspects already, with lip syncing a multitude of enhancements available to producers in both studio and live performances.
Consumer protection and copyright laws may address many of these issues, but probably not all.
Take home messages
Take home message 1:
Overly burdensome consent procedures for training data will mostly delay and degrade services, and create problematic market structures.
Take home message 2:
If big is better for training data, there may be a role for mandating sharing access to large datasets. The essential facilities doctrine is relevant here, but existing antitrust laws are difficult to use, and if policy makers know what they want, rules that make this work better can be implemented.
Take home message 3:
AI can be used to establish credit and sort out compensation, but having good metadata can be critical, when training the AI.
Take home message 4:
We need to look deeper into the role of transparency and auditing for some AI services, particularly those that become dominant platforms or are used in some official or quasi official way to resolve disputes over credit and remuneraiton.
Take home message 5:
The measures to address better metadata, credit and remuneration will have a global dimension. The Copyright Office should ask that the WIPO SCCR use the current agenda item on Copyright in the Digital Environment to discuss these issues.