Overview
In this final edition of our four-part series, we talk about AI-related IP licensing issues that implicate antitrust concerns. We discuss antitrust implications of licensing pools that may emerge to address AI content scraping, as well as the latest developments in standard essential patents (SEPs) that will emerge in the context of AI.
Licensing Pools for AI Learning Content / Training Data
Generative AIs heavily rely on ingesting vast amounts of content, identifying patterns, and reflecting those patterns in their outputs. This process has triggered significant IP challenges, as content providers allege that AI companies are using their copyrighted works without permission or compensation. For example, the New York Times sued Microsoft and OpenAI, alleging OpenAI copied content from the New York Times' sources when building its large language models. The claim framed this as a "free ride" on the Times' substantial investment in journalism. Judge Sidney Stein accepted most of the Times' claims and rejected the majority of Microsoft's arguments on a motion to dismiss.
The practical challenge is clear: if courts are going to require AI companies to get licenses to use copyrighted material in their large language models, negotiating individually with each of the innumerable rights holders is not realistic.
Similar issues arose in the past between music companies and broadcasters. Broadcasters needed to use thousands of songs, but negotiating with individual rights holders was impractical because of high transaction costs. The solution was the establishment of the Performing Rights Organizations (PROs), including ASCAP and BMI, which offered blanket licenses.
But PROs, which pursue the interest of content providers, have caused conflicts with users. After a 1941 antitrust litigation against ASCAP and BMI alleging price-fixing among music companies and abuses of market power, the DOJ entered what are now the longest standing consent decrees, under which ASCAP and BMI today operate their blanket licenses without violating antitrust laws.
AI companies can look to the examples of the PROs as a path to obtaining the copyright licenses that some courts suggest they need. Licensing pools would reduce transaction costs and create predictable licensing structures. But such licensing pools will require close attention to antitrust compliance – to avoid abuses of market power and allegations of price-fixing.
In addition, new challenges will arise for licensing pools. Pools with extensive content will attract more AI companies, creating strong network effects. If exclusionary practices occur, as is often alleged in SEP disputes, competition could be significantly affected. Additionally, protecting non-participating content providers could be challenging. As compared to the broadcasting of music, AI training involves massively more volumes of data processed at unprecedented speed, complicating rights management. These dynamics may revive debates over fractional licensing or even compulsory licensing as seen in the US Department of Justice's (DOJ) most recent review of the ASCAP and BMI consent decrees and in Judge Mehta's order in U.S. v. Google require Google to license certain search data to competitors.
Additionally, to the extent that large frontier model companies have taken licenses to broad swaths of training content (whether in a pool or not), concerns may arise about a concept known as anticompetitive acquiescence. By taking licenses to content that may be protected under copyright's fair use doctrine, arguments could potentially arise that licensing precedents are being set that could create barriers to entry for smaller AI developers that might feel compelled to take licenses to content that might not have otherwise been required (depending on the outcome of numerous pending copyright suits). This is yet another reason to keep current on trends in the licensing of training data.
Standards Essential Patents and AI
Just as a licensing regime for AI training data must be reasonable from an antitrust standpoint, so too must patent licensing comport with the antitrust laws. SEPs critical to AI and IoT technical standards include 5G, Wi-Fi 6, audio and video codec standards, model and agent interoperability standards, and numerous other technical standards. Without these essential patents, AI deployers will not be able to use vital AI technologies. For this reason, the antitrust laws require that SEPs are licensed on fair, reasonable, and non-discriminatory (or FRAND) terms.
In Europe, a recent proposal to introduce laws to regulate SEPs by increasing transparency through a SEP registry and a competence center at the European Union Intellectual Property Office (EUIPO) has been shelved. That proposal had been an attempt to improve access to the more than 75,000 SEPs worldwide by implementers but ran into political difficulties in light of concerns that it would disproportionately impact US tech companies.
Instead, in a recent review of its Technology Transfer Guidelines, the European Commission proposed guidance on two elements relevant to AI:
- first, an antitrust safe harbor for technology pools (that is, where companies combine various technologies that are then licensed to all members of the pool and potentially third parties); and
- secondly, an antitrust safe harbor for groups seeking to jointly negotiate terms for technology implementers (called "licensing negotiation groups" or "LNGs").
The European Commission has acknowledged the pro-competitive effects of both pools and LNGs, but has also recognized the competition risks arising from coordination and information sharing. Pools can enable "one-stop-shop" licensing and allow for pro-competitive standards, while running the risk of establishing de facto industry standards that make it more difficult for alternative technologies to be introduced. Pools must therefore ensure a degree of transparency as to what precise rights are included, and avoid "double-dipping" (that is, charging licensees more than once for the same technology rights). LNGs can reduce the transaction costs of technology licensing and improve negotiations but can also dramatically improve the bargaining position of implementers and run risks of anticompetitive information sharing. LNGs must therefore be limited, open to new joiners, voluntary, and have in place safeguards to avoid anti-competitive information sharing.
These moves followed informal guidance letters by the Commission and the German Antitrust agency approving joint negotiation by the German automotive industry for SEPs for communications technology. However, pools and LNGs are not without their critics, since pools and LNGs tend to disproportionately benefit those seeking access to SEPs, rather than the holders, and therefore benefit European companies while being seen to target US companies. The DOJ recently criticized the approval of these LNGs as akin to endorsing the formation of a "buyers' cartel" that enabled price-fixing. A final decision on these rules is expected next year.
Given that the European proposal was shelved and the prospect of a "one-stop-shop" for AI-related SEPs is unlikely, navigating the SEP licensing thicket will continue to be a challenge for technology implementers. Holders of SEPs will continue to pursue litigation in the context of AI-related technologies around the world. For example, KeeeX recently filed litigation in the Unified Patent Court against OpenAI. In this litigation, KeeeX asserted a patent that is allegedly essential to the Coalition for Content Provenance and Authenticity's (C2PA's) content authentication standard for detecting deepfakes. Many companies are using the C2PA standard to comply with new laws, including the California AI Transparency Act (SB 942), which requires certain AI companies to create an AI detection tool that watermarks and allows individuals to determine provenance of a potential deepfake. KeeeX illustrates that litigation will emerge over new AI-related standards, and discussions will need to be had over how licensing will be pro-competitive and FRAND as the litigation landscape evolves, and as the thicket of relevant standards pertaining to AI licensing becomes more complex. This is particularly the case as patent plaintiffs and defendants battle over FRAND licensing terms and injunctions in global forums such as the UK, Brazil, and China. Many view the improper use of SEPs to obtain injunctions to be anticompetitive and inconsistent with the concept of FRAND. But as the European example underscores, no one can seem to agree on an over-arching proposal that will comprehensively address how to manage SEPs in a rapidly evolving AI landscape.
Conclusion
As we conclude this series, one thing is clear: the intersection of AI and antitrust, is one of the most dynamic and consequential areas in today's legal landscape. The topics addressed in this series, algorithmic coordination, unilateral conduct, monopoly-related concerns, and the interplay between AI, antitrust, and IP, only begin to illustrate the complexity of issues emerging as technology evolves. Our analysis has merely scratched the surface, and the pace of regulatory and judicial developments will demand close attention. We are actively monitoring these trends, and nearly every business should do the same, as the implications for compliance, risk management, and competitive strategy are substantial.
