With artificial intelligence (AI) labs’ demand for new data sources showing no signs of slowing, companies that sell and route data — ranging from Mercor and other data vendors to services such as OpenRouter — have quickly become some of the most important and valuable companies in the AI industry.
Yet this data pipeline faces growing risks, fuelled by declining trust in AI companies. The fight around data centres might be grabbing headlines but another core driver of this mistrust lies in AI’s original sin: scraping data to train their models. The debate over whether this data was fair use is making its way through the courts, but in the meantime, there is a growing movement of creators interested in blocking AI scrapers from accessing content.
To address these concerns and ensure a healthy market equilibrium, AI companies and policy makers can take inspiration from performing rights organizations (PROs) in the music industry. As the number and size of data intermediaries continue to grow, there will be significant impacts on data and labour markets. Research on data value estimation and economics has begun to highlight the potential benefits of collective rights management, such as lowering transaction costs and creating clear market signals around which types of data are valuable. We believe the history, market structure and regulation of PROs offer useful lessons for designing effective data intermediaries — intermediaries that can deliver the high-quality data AI companies need while returning economic value to creators in a way that sustains incentives for creation.
PROs ensure that rights holders (such as music artists) are paid each time their work is played publicly (by radio stations, for example, or in restaurants or stadiums). They are collectives that bundle rights from different artists together into a single catalogue that users can access for a fee. AI data intermediaries, meanwhile, collect, structure and resell data to AI labs for training large language models (LLMs). While many organizations currently act as a kind of data intermediary, there are three main categories for which the PRO comparison is especially relevant: organizations that intermediate retrievable content (such as Tollbit, Cloudflare and Prorata), those that intermediate human labour (Mercor and Afterquery, for example) and those that route user requests to AI products (such as OpenRouter).
The Original Data Dispute
The early history of PROs resonates with today’s concerns about AI data use. In 1847 in Paris, as the story goes, the composer Ernest Bourget refused to pay his bill at Les Ambassadeurs Café-Concert, arguing that he owed nothing to the café because it had used his music without compensating him. The café was banned from playing Bourget’s works and the composer was awarded damages. Then, in 1914 in New York, American composer Victor Herbert sued a restaurant after hearing his compositions being played by the restaurant band. In 1917, the US Supreme Court ruled that composers must be paid whenever their music is performed publicly for profit.
Despite these one-off cases, wide-scale enforcement remained economically challenging. A composer could not realistically travel from café to café to identify performances of their work and negotiate bilateral licences. Equally, no venue operator could plausibly track down every composer whose work might be played that evening.
PROs solved this through aggregation on both sides of the market. The American Society of Composers, Authors and Publishers (ASCAP), the first US-based PRO, was formed in 1914. A single blanket licence replaced thousands of individual negotiations, and a single collecting organization replaced thousands of individual enforcement actions.
AI data intermediaries have developed similarly. While larger content providers such as NewsCorp and Reuters have negotiated bilateral data deals with AI labs, Tollbit, Prorata and other companies were founded to help small and medium data creators protect their content from these labs and decide how much to charge for access.
Information about the structure of recent AI licensing deals is scarce. In light of this limited information, looking to historical precedent may be especially useful. The PRO model normally involves a single blanket licence, meaning that users cannot pay for just one song or artist within the PRO’s catalogue. Instead, the PRO charges a percentage of the user’s revenues. Broadcast Music, Inc. (BMI), one of the largest PROs in the US, charges radio stations 2.20 percent of gross revenues coming from songs in its catalogue.
The PRO distributes revenue among rights holders based on how often their respective works in the catalogue are used. For example, while Taylor Swift and an obscure classical music composer might belong to the same PRO, their share of revenues depends on how many times users of the catalogue played their works.
In the context of AI training data, a local newspaper and The New York Times could belong to the same PRO-like organization. They would both benefit from increased revenue from AI labs, with the exact amount of revenue depending on how often NYT content was used as a source in LLM answers compared to the local newspaper.
In exchange for negotiating with thousands of users and distributing revenue equitably to rights holders in the collective, PROs take a percentage of revenue collected from users to cover their operating costs. In 2019, Australia’s PRO, the Australasian Performing Right Association, published the expense ratios of 57 PROs worldwide. The mean expense ratio was 20.14 percent, with a median of 17.9 percent.
Lessons Worth Borrowing
Most countries have singular non-profit PROs that act as regulated monopolies. To prevent abuse of their market positions, these countries have introduced laws that could also be applicable to regulating AI data intermediaries. The most significant is the United States’ consent decree laws governing its two biggest PROs, ASCAP and BMI. The decrees apply regardless of ownership: BMI converted to a for-profit company in 2022 and remains bound by them, while newer entrants such as Global Music Rights (GMR) are not. The decree has four main elements, all of which may be emulatable in the context of AI data intermediaries.
- Rate courts: If a licensee thinks ASCAP or BMI’s rates are unreasonable, they can challenge the fees in a specific rate court in New York. The rate court then decides the appropriate rate. Although rarely used, the rate court arguably depresses the amount ASCAP and BMI can charge. In the not-too-distant future, it is plausible that data intermediaries will become large enough to wield significant pricing power, especially over users with fewer resources than the major AI labs. In that scenario, something akin to rate courts could act as a mechanism to regulate prices charged for data.
- Open membership: ASCAP and BMI must accept nearly any songwriter, which benefits smaller artists. Because a blanket licence grants access to the entire catalogue at a single price, licensees pay for the works of unknown writers alongside those of superstars — the catalogue’s stars are what make the licence worth buying. PROs outside the decrees work differently. They are invite-only and their members are must-have superstars such as Drake and Bruce Springsteen. GMR is not bound by the consent decrees that govern ASCAP and BMI, allowing it to set rates outside the rate court and split the money among far fewer members.
- Non-discrimination: ASCAP and BMI have to charge the same amount to similar users. A similarly sized radio station in New York pays the same rate as one in Los Angeles.
- Non-exclusivity: Members only give ASCAP/BMI a non-exclusive right to license their works. Songwriters and publishers retain the right to bypass ASCAP/BMI and negotiate direct licences with users.
Bourget’s dispute with a café nearly 180 years ago spawned a system that today licenses billions of performances a year worldwide. PROs solved the problem of prohibitive transaction costs between many rights-holders and many users — the same model can solve similar barriers in AI licensing agreements today. AI labs cannot negotiate millions of individual agreements with data providers, while many data providers lack the market power or technical expertise to negotiate agreements with AI labs.
A PRO-like model for AI licensing agreements would prevent years of litigation, blocked crawlers and bilateral deals. Policy makers will need to decide whether to allow for for-profit PROs such as Cloudflare, which seems to be the best candidate, or to establish regulated non-profit monopoly PROs who negotiate deals with AI labs on behalf of thousands of data creators in each country. We do not need to recreate the wheel — in the quest for better AI licensing agreements, mimicking aspects of the music industry might hit the right note for labs, data creators and the government.
