ICLE Issue Brief

A Framework for Understanding and Evaluating AI Commercialization Strategies

I. Introduction

The rapid advancement of artificial-intelligence (AI) technologies has generated tremendous interest in how to regulate this emerging field. Implicit in the question are various assumptions about how AI will be developed and deployed in the marketplace. As it turns out, this is not nearly as simple a forecast as it may at first appear. This is particularly evident when one examines the differences between open-source and proprietary approaches to commercialization,[1] and is particularly relevant in light of the growing power costs needed to run large-scale models.[2]

As AI continues to transform industries and reshape our social and economic landscape, understanding the various commercialization strategies available to AI developers is crucial for policymakers, business leaders, and researchers alike.

This paper examines the diverse commercialization models for AI development, exploring their strengths, weaknesses, and potential implications for innovation and competition. Our analysis is grounded in the recognition that AI is not a monolithic technology, but rather a diverse array of techniques and applications with vastly different market potentials and societal impacts.

We begin by defining AI in the context of commercialization, highlighting the complexity and variety of technologies that fall under this umbrella term. We then explore four essential commercialization strategies that are likely to shape how AI will be deployed: proprietary private, open-source private, hybrid private/public, and publicly directed models.

Throughout our analysis, we emphasize that no single commercialization approach is universally superior. Instead, the optimal strategy depends on the specific context, including the nature of the AI technology, target market, and broader ecosystem in which it operates. We also consider how these commercialization choices intersect with and are shaped by regulatory decisions and public policy.

II. Defining AI in the Context of Commercialization

The term “artificial intelligence” has become ubiquitous in popular culture since the November 2022 launch of OpenAI’s ChatGPT made individual “generative artificial intelligence” capabilities broadly available to the general public.^[3] Indeed, ChatGPT was in its first year the fastest-growing consumer application in history.^[4]

But what do we mean by AI? The deceptively simple term remains obscure. As Jason Potts has noted: “Generative AI is not a single technology, or even industry, and is composed of extremely complex and varied ownership and governance at each composite layer.”^[5] Instead, we must consider the diverse ecosystems of AI technologies—from core technologies through foundation models and specific applications—and how they interact with various sectors and user communities.

Some economists characterize AI software as a “general purpose technology,” with some form of the technology today widely embedded in firms, as well as their products and services.^[6] Yet until the advent of ChatGPT, many consumers were unaware of AI’s pervasiveness, as it was largely hidden in the back offices of business and government. It was only in the last few years that consumers began to interact with AI technologies regularly, such as when they accessed their bank accounts, paid bills, made purchases online, or interacted with online assistants. The introduction of highly capable large language models (LLMs) has rapidly changed public perception of applied AI’s still-incipient capabilities. Specifically, there is now widespread recognition of the potential impact these technologies could have on society.

McKinsey & Co. offers a functional definition: “AI is a machine’s ability to perform the cognitive functions we associate with human minds, such as perceiving, reasoning, learning, interacting with an environment, problem solving, and even exercising creativity.”^[7] AI itself, however, consists of multiple technologies and designs, including such methodologies and applications as machine learning, deep learning, natural language processing, computational statistics, facial recognition, and robotics (among many others).^[8] This definition is quite expansive, and is perhaps unhelpful in its tendency to give the impression that “AI” refers to a single, unified concept. It includes everything from foundational models capable of producing human-like text to algorithms that can diagnose illnesses more precisely than human physicians to “simple” algorithms that enable more personalized online experiences.

When considering commercialization models for AI, it’s crucial to recognize that we are not dealing with a monolithic technology, but rather with the intersection of a variety of techniques and technologies that often serve very different purposes. From machine-learning algorithms used in predictive analytics to natural-language processing in chatbots, from computer vision in autonomous vehicles to reinforcement learning in robotics, each AI application may require a unique commercialization approach. This diversity in AI technologies means that no single business model will be universally applicable. Instead, successful commercialization strategies must be tailored to the specific capabilities, applications, and target markets of each AI technology, while also considering the broader ecosystem in which these technologies operate and interact.

III. Commercialization Models for AI Development

There are a variety of commercialization paths for AI developers. For our purposes, we will discuss four essential commercialization strategies, and related public policies that we believe are most likely to shape how AI will be deployed: proprietary private, open-source private, hybrid private/public, and publicly directed.

A. Private Models: Tradeoffs Between Proprietary and Open Approaches

1. Private-proprietary commercialization

A fully private business model for AI production would operate on a presumption that the technology that drives the AI development is proprietary, and the firm is free to leverage this proprietary technology along a spectrum of deployments, from a fully integrated/closed ecosystem to one that is relatively open to external integrators. This approach relies on a mix of different forms of intellectual-property protections and contractual obligations. For example, copyrights, patents, and trade secrets may form the basis for the core business model, while contractual relationships with integrators and other developers could then be used to facilitate functional integration (to varying degrees) via licensing.

A fully closed business model for AI production would leverage the firm’s proprietary technology and intellectual property to create a vertically integrated ecosystem. Under this model, the firm maintains tight control over the development, deployment, and commercialization of its AI technology.[9] The primary objective would be to capture value by creating a unique, differentiated offering that is difficult for competitors to replicate.

The foundation of this model rests on various forms of intellectual-property protection. Copyrights protect the original expression of ideas, such as the source code and documentation associated with the AI system.[10] Patents provide a time-limited exclusive property right over novel, non-obvious, and useful inventions, including (in some cases) AI algorithms and architectures.[11] Trade secrets protect confidential information that derives economic value from not being generally known, such as the training data, hyperparameters, data-center design, and optimization techniques used in the AI system.[12] Building on this intellectual-property foundation, the firm can then create a fully integrated ecosystem around its AI technology. This ecosystem may include proprietary hardware (e.g., specialized AI accelerators); software (e.g., development tools, APIs, and applications); and services (e.g., cloud-based AI platforms).

A license involves a mutually negotiated, contractual agreement—often employing a nondisclosure agreement during licensing negotiations—whereby the licensor retains title to the underlying IP but assigns a right to exploit the IP to a third-party licensee.[13] The licensee can typically make, use, and sell a product using the licensed AI technologies—either through an exclusive license or nonexclusive license—for specified financial terms or by paying a stream of predetermined royalties.[14] A similar arrangement can be struck for access to proprietary software through more restricted means—for instance, for providing consumption of services via an API.

Recent inquiries from federal regulators have raised concerns about the potential implications of various proprietary business models in the AI industry.[15] There is, however, little for lawmakers and enforcers to assume ex ante when considering whether firms should pursue an open or closed proprietary strategy. As federal regulators grapple with the challenges of overseeing this rapidly evolving industry, it is crucial to recognize that the commercialization choices firms make are a reflection of the unique contingencies of their production processes. Firms often blend elements of both open and closed models to strike a balance that aligns with their goals, market dynamics, and consumer preferences.

Open approaches to AI development (e.g., offering interoperable services) offer numerous benefits, such as fostering innovation, enabling scientific research, and promoting competition in the marketplace. By making foundational AI models accessible and adaptable, open approaches can lower entry barriers and encourage a diverse ecosystem of applications tailored to various needs and challenges. This openness can drive technological advancements, price efficiencies, and ultimately benefit consumers and society at-large.

But closed approaches (also known as “walled gardens”) also have their merits. Closed proprietary development can provide enhanced security, privacy, and a more streamlined user experience. Closed models may also allow firms to maintain greater control over their intellectual property, ensuring a level of competitive advantage and providing incentives to invest in research and development (R&D). Under the right conditions, closed systems can likewise foster a healthy ecosystem of complementary products and services.

Indeed, ecosystem development is not really an either/or matte, as even closed/proprietary approaches to monetizing AI can have downstream “market making” effects on AI development:

Licenses are…a direct rent, and a way to gate or control competition through their issuance. Additionally, the terms of these licenses influence subsequent related markets, such as product embeddings, thereby also governing the development of business ecosystems, typically by restricting their growth or creating bottlenecks. A critical aspect of using licenses to manage rents and ecosystems relates to the discovery of value and the property rights that accrue from opportunities identified by third parties or users. The implication of a right, akin to an option, within the license aligns with real options theory. Effectively, the license determines where the financial value of these strategic real options is capitalized, whether in the licensor or the licensee.[16]

Moreover, whether any given ecosystem is “open” or “closed” is not a binary question:

[AI] technology is a cumulative result of decades of research and development by individuals and teams that have built extremely powerful and often valuable capabilities and products. Some of this work circulates in public (open-source code), but also in publications, open forums, institutions (universities), labour markets (hires) and financial markets (acquisitions). Still, many elements remain closed and protected with intellectual property and corporate secrecy (algorithms, private training data), and through tacit knowledge (especially in training). So, its development through innovation is both open and closed, simultaneously public and private. It is institutionally complex. [17]

The optimal choice between open and closed models depends on the relevant market participants’ specific needs and preferences. As Jonathan M. Barnett notes in the context of digital platforms, open systems may yield no net social gain over closed systems; can pose a net social loss under certain circumstances; and can impose a net social gain under yet other circumstances.[18] According to Barnett, there is a “fundamental welfare tradeoff between two-sided proprietary…platforms and two-sided platforms which allow ‘free entry’ on both sides of the market.”^[19] Consequently, “it is by no means obvious which type of platform will create higher product variety, consumer adoption and total social welfare.”^[20]

The interplay between consumer and developer preferences significantly influences the adoption of open versus closed AI models. Consumers often prioritize ease-of-use, security, and seamless integration, which can favor closed ecosystems. But some users may place greater value on the flexibility and customization options of open systems. For instance, while Linux is popular among developers for its openness, consumer-oriented operating systems like Windows and MacOS dominate the desktop market due to their user-friendly interfaces and extensive software ecosystems.^[21]

These dynamics give rise to a complex landscape in which the success of a model depends on its ability to balance the sometimes conflicting needs of both consumers and developers. Ultimately, the coexistence of open and closed models in the AI ecosystem fosters innovation and provides diverse options to meet varying user requirements. Furthermore, the distinctions between open and closed models can significantly influence competition between brands. If public policy is enacted that favors one business model over another, it would likely lead to reduced choice and lower overall benefits along this dimension.[22]

The diffusion of firms exploring different business models yields benefits in many ways. For example, when it comes to ethical AI development, various types of organizations should be encouraged to explore different approaches, and even to combine them when appropriate. Anthropic’s “Collective Constitutional AI” approach adopts a “semi-open” model that combines private elements with certain aspects of openness in order to encourage innovation, while maintaining some degree of control.[23] This model may strike a suitable balance by ensuring a degree of proprietary innovation and competitive advantage, while still benefiting from community feedback and collaboration.

Conversely, a completely open-source approach to development could yield different, and possibly superior, outcomes that address a wider range of needs through community-driven evolution and iteration. It is impossible to determine, in advance, whether an open or closed approach to AI development will inherently lead to better results in creating “ethical” AI. Both approaches have their merits, and the most effective solutions will likely incorporate elements of both.

By differentiating themselves through a focus on ease-of-use, quality, security, and user experience, closed systems contribute to a vibrant competitive landscape in which consumers have clear choices between differing “brands” of AI. In essence, codifying a regulatory preference for one proprietary business model over the other would oversimplify the intricate balance of tradeoffs inherent to AI development.

2. Private/open-source commercialization

The next flavor of private AI commercialization models is open source. This is distinct from the “open vs. closed” private models discussed in the previous subsection. In contrast to the restrictions inherent in the proprietary models discussed above, “open source” is a decentralized software-development model that encourages open collaboration among potential developers.[24] The term generally refers to source code made freely available to individuals and companies to modify and redistribute, and includes legal permission to use the source code, design documents, or product content.[25] Source code is typically released under the terms of a software license, which typically allows other users to download, modify, and subsequently publish their enhanced versions.[26]

Stable Diffusion, an open-source text-to-image generation model developed by Stability AI, is a prominent example of open-source generative AI.[27] The model was trained on a large dataset of images and their associated captions, allowing it to generate novel images from textual descriptions. The source code, model weights, and training data were made freely available, enabling developers to experiment with and build upon the technology.[28] Another prominent example is PyTorch, an open-source machine-learning framework led by Meta’s AI Research lab.[29] PyTorch provides a flexible and intuitive interface for defining and training neural networks. Other notable open-source AI projects include TensorFlow, Keras, Scikit-learn, and OpenCV.[30] These frameworks and libraries provide powerful tools to develop machine-learning and computer-vision applications, lowering the barrier to entry for AI development and enabling rapid prototyping and experimentation.[31]

While there are more than 100 Open Source Interconnection (OSI) licenses, for our purposes, we will categorize them into two typologies: permissive and copyleft.[32] An open source permissive license governs how others (i.e., those other than the copyright holder) can inspect, use, modify, distribute (and freely redistribute), or enhance software code.[33] The open-source software licensor has the option to change the terms and conditions to grant the licensee the permission and rights to use or repurpose the code for new applications or to include the code in other derivative work.[34]

Users of open-source licenses must accept the legal terms of a license, but the terms of an open-source license differ from those of a proprietary license.[35] For example, in the open source copyleft AI license, the licensor grants the licensee permission to use, modify, and share the source code, but offers stipulations against relicensing through specific terms and conditions in the agreement, including that any derivative work is required to be released under copyleft license terms and conditions identical to those of the original work.[36]

There are numerous examples of open-source licenses, including Standard License (“non-copyleft”); General Public License (“strong copyleft”); Lesser General Public License (“weak copyleft”); Eclipse License (“weak copyleft”); Mozilla Public License (“weak copyleft”); Apache License (“non-copyleft”); MIT License (“non-copyleft”); and Berkeley Source Distribution (BSD) License (“non-copyleft”).[37]

A common example of permissive licensing can be found in the Creative Commons (CC) suite of licenses. Creative Commons offers a range of standardized licenses that creators can apply to their work, each with different levels of permissiveness.[38] While not strictly open source in the software sense,[39] these licenses share similar principles of openness and accessibility. A typical Creative Commons license allows the license holder to grant broad permissions for others to share, remix, use commercially, or otherwise utilize the license holder’s work without seeking specific authorization for each use.[40] This flexibility makes Creative Commons licenses a popular choice for various types of content—from academic papers to multimedia works—fostering a culture of sharing and collaboration similar to that seen in open-source software communities.

The Creative Commons license works alongside the rules of copyright, allowing the owner to authorize a license base from more strict to more open usage of the owner’s work, and thus to choose the level of protection that best suits the owner’s needs.[41] Under the Creative Commons license, there are six primary variants: attribution, attribution share alike, attribution no derivatives, attribution non-commercial, attribution non-commercial share alike, and attribution non-commercial no derivatives.[42] There are, however, also variants of the Creative Commons license (such as “commercial use is forbidden” and “derivatives are forbidden”) that are not OSI-approved.[43]

Meta’s Llama 3 is a good example of an AI system released permissively under a quasi-open-source license.[44] Like traditional open-source software, it allows for free use, modification, and redistribution of the model and its code.[45] But it diverges from standard open-source practices by including specific restrictions—such as prohibiting the use of Llama materials to improve competing LLMs and requiring prominent attribution—that are not typically found in open-source licenses.[46]

More broadly, an open-source-specific AI license would likely incorporate elements from existing open-source software licenses, while addressing the unique characteristics of AI systems. Key components might include:

1. Granting users the freedom to study, use, modify, and share the AI system and its components, including source code, model weights, and training data;
2. Requiring any derivative works or modifications to be released under the same open-source license terms, ensuring the continued openness of the technology (copyleft provisions);
3. Permitting commercial use of the AI system, fostering its adoption and integration into real-world applications;
4. Requiring attribution to the original creators and contributors, acknowledging their work and promoting collaboration; and
5. Disclaiming warranties and limiting liability, protecting the licensors from potential legal issues arising from the use or misuse of the AI system.[47]

The development of AI systems under an open-source model offers both benefits and challenges. On the positive side, open-source AI can promote transparency, collaboration, and rapid innovation. By making the underlying technology accessible to a wide range of developers and researchers, open-source AI enables diverse perspectives and approaches, leading to more robust and adaptable systems. Open-source development can also foster trust, as the AI system’s inner workings can be scrutinized and verified by the community.[48]

But open-source AI also presents legal and economic challenges. The lack of proprietary control over the technology may discourage commercial investment, as companies may struggle to monetize their efforts.[49] A flip side of the benefits of the open nature of the software is that, in some cases, malicious actors can access and modify the technology for harmful purposes.[50] Additionally, the open nature of the development process may complicate issues of liability and accountability, as the responsibility for any the AI system’s negative consequences may be diffused among multiple contributors.

To balance these tradeoffs, a hybrid approach that combines open-source and proprietary elements may be optimal. For example, companies could release certain components of their AI systems as open source while maintaining control over others, thus enabling collaboration and transparency while preserving competitive advantages.[51]

B. Hybrid Models: The Role of Standards

The third model for commercialization comprises various forms of hybrid public/private standards. The fair, reasonable, and non-discriminatory (“FRAND”) licensing approach is one example common in some patent contexts. FRAND licensing involves a voluntary agreement between a holder of a standards-essential patent (SEP) and a manufacturer or implementer, with the agreement allowing the implementer to use the patentee’s technology in its manufactured products.[52] Established by standards-setting organizations (SSOs) to ensure technical compatibility and interoperability in their products,[53] these SEPs are “patents that are considered essential to implement a specific industry standard” and are “governed by the terms and conditions of the licensing agreement.”[54]

The fair and reasonable part of FRAND means that “the patent holder should charge a licensing fee that reflects the economic value of the patent,” while “taking into account factors such as the importance of the patent to the standard, the contributions of the patent to the product, and the prevailing licensing rates in the industry.”[55] The non-discriminatory part of FRAND means that “the patent holder should offer the same terms to all companies wanting to license the patent,” thus “preventing the patent holder from discriminating against certain companies.”[56]

As it has been traditionally understood, however, patenting the core technologies underlying AI is exceedingly difficult under prevailing U.S. law.[57] It is therefore unlikely that a true FRAND approach is possible in the United States, although that doesn’t rule out that other forms of government involvement could be used to establish a similar forced-sharing regime. For example, a weaker form of this model would be to create “standards” to guide the development and commercialization of AI technologies, either through a direct governmental standards-setting approach, or a hybrid public/private model that involves government-led and/or enforced standards development.[58]

Standards setting, either on a private basis or through government-led initiatives, can play a role in shaping development practices and ensuring safety and interoperability across technologies in a way that can potentially lead to more commercialization options as a result of elevated consumer trust.  As is always true, however, there are both benefits and costs when thinking about the way that public or public/private standards shape commercialization decisions.

These standards can not only help to organize the industry by establishing common frameworks and protocols but can also enhance public trust by providing assurances that AI development is under diligent oversight. As AI systems increasingly permeate such critical sectors as health care, transportation, and finance, the perceived need for assurance regarding their safety, reliability, and ethical implications grows. Commercialization of these technologies is necessarily shaped by these realities.

These standards can originate in a number of ways. Firms can develop standards as part of the process of instituting their own internal best practices.[59] Large professional organizations also contribute important work in this area. For example, the Institute of Electrical and Electronics Engineers (IEEE) has been working to develop an “Ethics Certification Program for Autonomous and Intelligent Systems,” the goal of which is to “create specifications for certification and marking processes that advance transparency, accountability and reduction in algorithmic bias in Autonomous and Intelligent Systems (AIS).”[60] The IEEE also has a number of similar efforts underway.[61]

In 2017, the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) established the ISO/IEC JTC 1/SC 42 Artificial Intelligence Committee,[62] tasked with developing universal AI standards that would apply across different industries and applications. Their current projects include creating a comprehensive interoperable framework for AI systems.[63] The Consumer Technology Association (CTA) develops trustworthiness standards to address the use of AI in health care,[64] assessing the impact of AI’s trustworthiness from multiple perspectives—including those of physicians, consumers, professional and family caregivers, public-health officials, medical societies, and regulators.[65]

Since 2019, the National Institute of Standards and Technology (NIST) has actively contributed to AI governance, recommending strategies for the federal adoption of AI to enhance both the U.S. economy and national security. A February 2019 White House executive order directed NIST to promote the use of voluntary standards in AI,[66] coordinate with standards development organizations via the National Science and Technology Council,[67] and engage with the private sector on emerging AI standards.[68]

More recently, NIST developed the AI Risk Management Framework (AI RMF), a set of guidelines to improve trustworthiness in AI design and use, emphasizing consensus-driven, outcome-focused, and nonprescriptive standards.[69] The recently proposed bipartisan U.S. Senate framework also contains many of these elements, recommending studies where laws are insufficient to capture AI systems’ potential harms but otherwise recommending federal coordination on standards with largely voluntary compliance.[70]

On the more extreme side, there are calls to create a central AI regulator, with the proposed regulator having authority to grant licenses before AI development can be undertaken, and enforce ongoing compliance.[71] This centralization aims to standardize oversight and ensure adherence to established standards across the AI industry.[72]

It’s crucial to recognize that nominally voluntary standards can effectively become mandatory in practice. This can occur, for example, when courts adopt these standards in negligence litigation as benchmarks to establish a duty of care.[73] Thus, even in the absence of any direct regulatory mandates or common agreement, standards often shape legal obligations and liabilities, influencing practices across industries by establishing a de-facto regulatory framework.

While we should not understate the benefit of some government participation in developing AI standards, the introduction of mandatory standards can carry substantial costs, depending on their scope and content. Applying the schema proposed by Knut Blind et al. to AI commercialization, we can consider how the impact of standards and regulations might vary depending on market uncertainty.[74] In markets with low uncertainty, formal standards may lead to higher innovation costs than regulations would.[75] This could be due to regulatory capture, where established firms influence standards to their advantage, potentially raising costs for competitors.[76] Conversely, in highly uncertain markets, regulations might impose higher innovation costs than standards. This may occur because regulators have less information about rapidly evolving technologies, leading to potential mismatches between regulations and technological realities.[77]

A good example of the double-edged sword of government involvement in standards setting can be seen in the campaign to require interoperability and compatibility among diverse technology providers.[78] Standardization can help to reduce fragmentation, thus facilitating smoother integration and interaction between different AI products and services. By fostering a common understanding and predictable environment, these standards might aid in accelerating technological advancements and innovation within a structured and well-defined boundary. But as noted above, a more open, interoperable business model is not a panacea, and it’s not at all clear that regulation-driven interoperability standards are the best way to reduce friction among various AI systems. There are tradeoffs in pursuing varying degrees of openness, whether imposed by fiat or market demand.

Moreover, government-led standards can restrict AI firms from fully capitalizing on unique, proprietary technologies by mandating uniformity that may dilute the distinctiveness of more innovative or advanced solutions. This can discourage investment in new technologies that don’t conform to existing standards, potentially stunting growth and innovation within the sector. Firms often leverage proprietary advantages for competitive gain and market differentiation; thus, stringent standards could limit their ability to offer something unique to the market, reducing the incentive for groundbreaking developments.

Further, adhering to mandatory AI standards can result in significant compliance costs, particularly affecting smaller firms and startups with limited resources.[79] These costs include not only the financial burden of implementing new systems and processes to meet the standards but also the ongoing expenses associated with maintaining compliance. This financial strain can divert funds from other critical areas like product development and research, potentially stifling innovation and limiting the company’s growth potential in a highly competitive market.

Standards may also significantly limit AI developers’ flexibility to engage in creative and technical decisionmaking. When standards dictate or prohibit specific aspects of AI system design and functionality, developers might find their hands tied, unable to explore alternative or potentially more innovative solutions that fall outside the set guidelines.[80] This could slow the pace of innovation in AI development, as companies might be forced to prioritize compliance over creativity and innovation.

C. Public Models of Commercialization

The fourth approach would be some form of compulsory-licensing or forced-sharing regime that creates a public commercialization model.[81]  The term “compulsory license” refers to a governmental entity that provides permission to enterprises that seek to use intellectual property without the owner’s consent, while offering financial compensation to the rightsholder.[82] Advocates for this licensing approach favor “policies that shape US markets while aligning innovation and competition policies towards creating a domestic AI innovation system that better serves the public interest.”[83]

The exact shape such a policy might take is highly contingent on the forms of intellectual property employed, and the conditions under which such use is deemed necessary. For example, the World Trade Organization’s Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS Agreement) permits member nations to institute compulsory licensing of patents under certain circumstances.[84] Compulsory licenses for patents are most often applied to pharmaceuticals and public-health-related inventions, but may apply to any patented invention.[85]

In the context of copyright, Congress has seen fit to institute compulsory licenses in certain instances, such as satellite and cable retransmission[86] and the use of musical compositions in recording “cover songs.”[87] The compulsory license is a concept that has been formally part of U.S. intellectual-property law since the enactment of the U.S. Copyright Act of 1909.[88] U.S. patent law has not paralleled copyright law, however, as compulsory licenses have been uncommon in the patent arena.[89]

Additionally, Congress is able to place compulsory-license stipulations on funding it provides to researchers and developers. The U.S. government—representing all U.S. taxpayers—has historically played a role in funding R&D to develop the nation’s computer-technology infrastructure. This extends to AI technologies and efforts to bring these AI technologies to the marketplace.[90] Moreover, public-interest advocates who push for broader “fully public” models often argue that their approach will “better align domestic investment and AI capability development with economic, societal and national security objectives.”[91]

This fully public commercialization policy might, for example, focus on a nationwide health emergency (such as developing a public-health vaccine) or some nonprofit purposes (such as meeting an environmental-policy goal). Indeed, the Biden administration indicated its willingness to use the Defense Production Act (DPA) to intervene in how AI is developed and produced.[92] Some advocates have also raised the threat of eminent domain-like actions against intellectual property.[93]

Relatedly, some recent public discourse has focused on the possibility that AI companies might obtain monopoly status, with attendant conversation about the need to act to deter this likelihood.[94] This is another flavor of treating core AI technologies as some sort of public good or essential facility that would be fully subject to governmental licensing. In many respects, such advocacy closely mirrors conversations from a decade ago about “data monopolies.”[95] But as Geoffrey Manne and Dirk Auer suggest, this characterization is overly simplistic and potentially misleading.[96] History has shown that, despite the leading tech platforms’ vast accumulations of data, their market dominance in the realm of generative AI has not been as unassailable as one might expect.[97]

The concept of market “moats” (a term used to describe barriers to entry that protect incumbents from competitive pressures) is central to this discussion. While the urge to dismantle these moats through aggressive antitrust actions is strong in some quarters, pursuing this urge would likely be a mistake. The rapid advancement and dynamic nature of AI technologies often render these moats less impenetrable than they might appear, and their existence is unlikely to be supported by rigorous empirical examinations.[98]

Moreover, using competition law to force open the business models of leading AI firms carries significant risks. Chief among these is that such policies could stifle innovation and impede the growth of the AI sector, rather than foster competitive markets. Of course, it’s not impossible that competition harms might emerge in the AI industry, as they can in any industry, but the assertions that these firms and their products should be treated as anything like an essential facility or public utility should be subject to heightened scrutiny in order to avoid high error costs and reduced innovation and consumer welfare.

A mandated forced-sharing regime would be a special (and extreme) case of the “open” business models discussed above. Thus, it’s worth considering some of the recent history of governmental interoperability requirements, and their effects on the commercialization paths that firms take.

Open-banking initiatives are a prime example. The regulatory push in the United States fundamentally aims to democratize data access within the financial sector by enabling customers to have control over their financial data, promoting a shift towards more customer-centric financial services.[99] This transition is, however, fraught with complexities that could undermine its intended benefits.[100]

One significant issue with the implementation of open banking relates to the technical and operational challenges involved in ensuring seamless and secure data sharing. The infrastructure required to support secure APIs that allow third parties to access bank data must be robust, resilient, and effective.[101] There is a risk that inconsistencies in the technological capabilities of different financial institutions and third-party providers could lead to vulnerabilities in the system that expose sensitive data or cause service disruptions. Further, in at least some implementations of open banking, obligations are placed on providers in ways that distort the competitive landscape.[102]

So-called “device neutrality” or “sideloading” represents another cautionary tale for forced openness. Article 6(1)(c) of the EU’s Digital Markets Act (DMA) addresses “sideloading”—the installation of third-party software through app stores other than the one provided by the manufacturer (e.g., Apple’s App Store for iOS devices).[103] The sideloading mandate aims to give users more choice but can only achieve this by removing the option of choosing a device with a “walled garden” approach to privacy and security, such as Apple’s iOS.[104]

By eliminating the choice of a walled-garden environment, a sideloading mandate essentially forces users to use alternative app stores preferred by app developers, who have incentives to create their own app stores or move to those with the least friction. This may also mean the app stores that invest least in privacy and security. Apple is well-known for its commitment to user privacy and security. Thus, regulations like the DMA and similar attempts in the United States[105] represent a reordering of digital-ecosystem providers that currently have the incentive to account for a broad array of consumer demands. In the place of such nuance, mandated openness overrides the preferences of a vast swathe of consumers.

As suggested above, implementing strong forced-sharing obligations could severely limit the commercialization paths available to AI developers. One of the primary challenges with compulsory licensing is the potential disruption to proprietary business models. AI firms often rely on exclusive control over their intellectual property to maintain a competitive edge and secure funding for further R&D. Mandating that these firms license their technology to third parties could dilute that competitive advantage, making it less attractive for investors to fund high-risk, high-reward AI projects. This could ultimately slow the pace of innovation, as firms might be less willing to invest in groundbreaking technologies if they are not able to fully capitalize on their proprietary developments.

Additionally, imposing forced sharing could lead to a homogenization of AI offerings, reducing the diversity and specialization that currently characterizes the AI market. AI firms tailor their technologies to specific market needs, while forced-sharing regimes could undermine this customization by requiring firms to share their innovations broadly—including potentially with competitors. This could lead to a scenario where AI technologies become more standardized and less differentiated, which might diminish the overall quality and effectiveness of AI solutions available to consumers.

Furthermore, the administrative and regulatory burden associated with enforcing compulsory licensing could create significant operational challenges for both AI firms and regulators. Ensuring compliance with licensing agreements or the requirements of a strong regulatory regime, monitoring the use of an incredibly wide scope of shared technology, and resolving disputes over access would require substantial resources and oversight. This could divert attention and resources away from innovation and toward regulatory compliance, further hindering the growth and development of the AI sector.

IV. Conclusion

The commercialization of AI technologies offers a complex landscape of opportunities and challenges. As we have explored, there is no one-size-fits-all approach to bring AI innovations to market. Instead, a variety of models—from fully proprietary to open source, and hybrid approaches in between—each offer distinct advantages and tradeoffs.

Proprietary models allow firms to maintain tight control over their intellectual property, potentially driving focused innovation and creating unique value propositions. These closed systems may, however, limit broader collaboration and slow the overall pace of advancement in the field. On the other hand, while open-source approaches could foster transparency, collaboration, and rapid innovation, they also present challenges in the difficulty inherent in finding sustainable business models and exercising appropriate quality control.

Hybrid models and industry standards attempt to strike a balance, promoting interoperability and shared progress while still allowing for proprietary advantages. These approaches, however, require careful management to avoid stifling innovation or inadvertently creating barriers to entry for smaller players.

Public models of commercialization, including compulsory licensing and government-directed development, offer potential benefits in terms of equitable access and alignment with public-interest goals. But they also risk dampening private-sector innovation and investment if not carefully implemented.

The diversity of AI technologies and their applications means that different commercialization strategies may be appropriate in different contexts. What works for a general-purpose LLM may not be suitable for a specialized medical diagnostic tool or a financial-trading algorithm.

As the AI landscape continues to evolve, it is crucial that policymakers and industry leaders remain flexible and adaptive in their views of commercialization (with perhaps the fully public models being the only ones presumed inappropriate, pending strong evidence to the contrary). Regulatory frameworks should aim to foster innovation and competition while addressing legitimate concerns about safety, privacy, and the ethical use of AI. They should avoid prematurely locking in any single commercialization model, instead allowing for experimentation and learning as the field matures.

Ultimately, the success of AI commercialization will depend on finding the right balance between openness and proprietary development, between rapid innovation and responsible deployment, and between market forces and public interest. By understanding the strengths and limitations of various commercialization models, stakeholders can make more informed decisions that will shape the future of AI and its impact on society.

[1] See infra at nn. 9-80, and accompanying text.

[2] See Lynne Kiesling, Data Center Electricity Use III: Make or Buy?, Knowledge Problem (Aug. 1, 2024), https://knowledgeproblem.substack.com/p/data-center-electricity-use-iii-make.

[3] Introducing ChatGPT, OpenAI (Nov. 30, 2022), https://openai.com/blog/chatgpt.

[4] Laurie A. Harris, Generative Artificial Intelligence: Overview, Issues, and Questions for Congress, Cong. Research Serv. (2023), at 1, available at https://crsreports.congress.gov/product/pdf/IF/IF12426.

[5] Jason Potts, Sources of Innovation in Generative AI, The Network L. Rev. (Feb. 12, 2024), https://www.networklawreview.org/jason-potts-generative-ai.

[6] Nicholas Crafts, Artificial Intelligence as a General-Purpose Technology: An Historical Perspective, 37 Oxford Rev. Econ. Pol’y 521 (2021), https://doi.org/10.1093 oxrep /grab012. It is not, however, absolutely clear how to regard “AI.” Indeed, as we note below, there is no single technology that can be called “AI.” What exists are a varied collection of techniques and technologies that add some form of intelligence to automated systems.

[7] What Is AI, McKinsey & Co. (April 24, 2023), https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-ai.

[8] Id.; Matt Vella, How AI Is Transforming Our World, in Artificial Intelligence: The Future of Humankind, Time (Nov. 1, 2017); Michael Atleson, Keep Your AI Claims in Check, Fed. Trade Comm’n (Feb. 27, 2023), https://www.ftc.gov/business-guidance/blog/2023/02/keep-your-ai-claims-check.

[9] For an example of this approach, see Apple Intelligence: AI for the Rest of Us, Apple, https://www.apple.com/apple-intelligence (last visited Aug. 2, 2024).

[10] See Pamela Samuelson, The Uneasy Case for Software Copyrights Revisited, 79 George Wash. L. Rev. 1746 (2011);Copyright could theoretically be invoked to protect models’ parameter-embedding weights as a means to control distribution. See Thibault Schrepel & Jason Potts, Measuring the Openness of AI Foundation Models: Competition and Policy Implications (Sciences Po Digital, Governance and Sovereignty Chair, Working Paper 11, 2024), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4827358.

[11] See generally Adam Mossoff, Why History Matters in the Patentable Subject Matter Debate, 64 Fla. L. Rev. 23 (2012).

[12] See David S. Levine & Ted Sichelman, Why Do Startups Use Trade Secrets?, 94 Notre Dame L. Rev. 751 (2019).

[13] See, e.g., Patent Licensing & Legal Options, Justia, https://www.justia.com/intellectual-property/patents/licensing (last visited Aug. 2, 2024); Nathan C. Lovette & Carlos P. Garritano, Patent Licensing: A Brief Summary, Tucker Ellis LLP (Aug. 16, 2022), https://www.tuckerellis.com/ip-tip-of-the-month-blog/patent-licensing-a-brief-summary; Proprietary Software License: Everything You Need to Know, Upcounsel, https://www.upcounsel.com/ proprietary-software-license (last visited Aug. 2, 2024).

[14] Patent Licensing and Its Types: Everything You Need to Know, GreyB, https://www.greyb.com/blog/patent-licensing-101 (last visited Aug. 2, 2024).

[15] See, e.g., Kristian Stout, ICLE Comments to NTIA on Dual-Use Foundation AI Models with Widely Available Model Weights, Int’l Ctr. L. & Econ. (2024), https://laweconcenter.org/resources/icle-comments-to-ntia-on-dual-use-foundation-ai-models-with-widely-available-model-weights.

[16] Schrepel & Potts, supra note 10, at 10-11.

[17] Potts, supra note 5, at 2.

[18] See Jonathan M. Barnett, The Host’s Dilemma: Strategic Forfeiture in Platform Markets for Informational Goods, 124 Harv. L. Rev. 1861, 1927 (2011).

[19] Id. at 2.

[20] Id. at 3.

[21] Linux adoption rates consistently trail MacOS and Windows. See Desktop Operating System Market Share Worldwide Feb 2023 – Feb 2024, StatCounter, https://gs.statcounter.com/os-market-share/desktop/worldwide (last visited Aug. 2, 2024); Nonetheless, many developers have a preferences for Linux systems. See Joey Sneddon, More Developers Use Linux than Mac, Report Shows, Omg Linux (Dec. 28, 2022), https://www.omglinux.com/devs-prefer-linux-to-mac-stackoverflow-survey.

[22] See Michael L. Katz & Carl Shapiro, Systems Competition and Network Effects, 8 J. Econ. Persp. 93, 110 (1994), (“[T]he primary cost of standardization is loss of variety: consumers have fewer differentiated products to pick from, especially if standardization prevents the development of promising but unique and incompatible new systems”).

[23] Collective Constitutional AI: Aligning a Language Model with Public Input, Anthropic (Oct. 17, 2023), https://www.anthropic.com/news/collective-constitutional-ai-aligning-a-language-model-with-public-input.

[24] Sheen S. Levine & Michael J. Prietula, Open Collaboration for Innovation: Principles and Performance, 25 Organization Science 1414 (2014), https://doi.org/10.1287/orsc.2013.0872.

[25] The Open Source Definition, Open Source Initiative https://opensource.org/osd (last visited Aug. 2, 2024).

[26] Id. at 18.

[27] Stable Diffusion Public Release, stability.ai, https://stability.ai/news/stable-diffusion-public-release (last visited Aug. 2, 2024).

[28] Id.

[29] Learn the Basics, PyTorch, https://pytorch.org/tutorials/beginner/basics/intro.html (last visited Aug. 2, 2024)

[30] See Tim Mucci, Five Open-Source AI Tools to Know, IBM (Dec. 15, 2023), https://www.ibm.com/blog/five-open-source-ai-tools-to-know.

[31] See Robert Gorwa & Michael Veale, Moderating Model Marketplaces: Platform Governance Puzzles for AI Intermediaries, 16(2) Law Innovation & Technology 9-10 (2024), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4716865 (noting how Hugging Face “is developing a business model where it can bundle additional ‘premium’ deployment features, lowering the  barrier to entry for less technical users or those simply seeking convenience. For instance, one product, ‘Inference Endpoints,’ advertises itself as a way to ‘deploy models in minutes’ on Hugging Face’s own infrastructure. Similarly, their ‘AutoTrain’ product allows one to ‘train, evaluate and deploy state-of-the-art Machine Learning models’ by simply uploading a dataset, without having to write a single line of code.”).

[32] Open Source License: Types and Comparisons, Snyk, https://snyk.io/learn/open-source-licenses (last visited Aug. 2, 2024); 5 Potential Risks of Open Source Software, Snyk, https://snyk.io/learn/risks-of-open-source-software (last visited Aug. 2, 2024).

[33] Jeanelle Horcasitas, Understanding Open-Source Software Licenses, DigitalOcean (Nov. 2, 2021), https://www.digitalocean.com/community/tutorials/understanding-open-source-software-licenses; What Is Open Source, OpenSource.com, https://opensource.com/resources/what-open-source (last visited Aug. 2, 2024).

[34] Id.

[35] Id.

[36] Id.

[37] Sylvan Leroux, Open Source Licenses Comparison [Guide], It’s Foss (Sep. 9, 2023), https://itsfoss.com/ open-source-licenses-explained; Gina Häußge, A Dev’s Guide to Open Source Software Licensing, GitHub https://github.com/readme/guides/open-source-licensing (last visited Aug. 2, 2024).

[38] Id.

[39] While some Creative Commons licenses (like CC0 and CC BY) can be considered “open” in a broad sense, they are not recognized as open-source licenses by the Open Source Initiative, the body that formally approves open-source software licenses. See About, Open Source Initiative, https://opensource.org/about (last visited Aug. 2, 2024).

[40] Michelle Kaminsky, What Is Creative Commons? 5 Frequently Asked Questions, LegalZoom (Mar. 27, 2023), https://www.legalzoom.com/articles/what-is-creative-commons-5-frequently-asked-questions.

[41] Id.; Kela Parker, Creative Commons License – Pros and Cons Explained, Studiobinder (May 28, 2023), https://www.studiobinder.com/blog/creative-commons-license.

[42] Id.

[43] See Häußge, supra note 37.

[44] Meta Llama 3 Community License Agreement, Meta (Apr. 18, 2024), https://llama.meta.com/llama3/license.

[45] Id.

[46] Id.

[47] See, e.g., Andrew M. St. Laurent, Understanding Open Source and Free Software Licensing (2004), available at https://people.debian.org/~dktrkranz/legal/Understanding%20Open%20Source%20and%20Free%20Software%20Licensing.pdf.

[48] Niklas Leicht, Given Enough Eyeballs, All Bugs are Shallow – A Literature Review for the Use of Crowdsourcing in Software Testing, Scholar Space (2018), http://hdl.handle.net/10125/50404.

[49] Josh Lerner & Jean Tirole, The Economics of Technology Sharing: Open Source and Beyond, 19(2) J. Econ. Perspectives 99, 105 (2005), (“Because firms do not capture all the benefits of the investments in the open source project, however, the free-rider problem often discussed in the economics of innovation should apply here as well”).

[50] Liam Tung, Open Source: Almost One in Five Bugs Are Planted for Malicious Purposes, ZDNet (Dec. 3, 2020), https://www.zdnet.com/article/open-source-software-how-many-bugs-are-hidden-there-on-purpose.

[51] In many cases, this will be a commercialization strategy that tries to quickly build a set of complimentary products supporting the firm’s primary revenue source. Joel West & Scott Gallagher, Challenges of Open Innovation: The Paradox of Firm Investment in Open-Source Software, 36(3) R&D Mgmt. 13-14 (2006). Apple, for example, released Swift and WebKit as open-source projects at least partly as a strategy to encourage development of software generally—and web apps, specifically—that would work well for MacOS- and iOS-based hardware products. See Open Source at Apple, Apple, https://opensource.apple.com (last visited Aug. 2, 2024).

[52] What Is FRAND Licensing?, Lerman Law Associates P.C. (Jun. 26, 2023), https://www.lermanlawpc.com/blog/ 2023/06/what-is-frand-licensing.

[53] Herbert Hovenkamp, FRAND and Antitrust, 106(1) Cornell L. Rev. (2020), https://www.cornelllawreview.org/2020/09/15/frand-and-antitrust.

[54] Garth Brian Hedenskog, What Is a Standard Essential Patent (SEP)? SHIP Global Intellectual Property (Oct. 29, 2019), https://shipglobalip.com/blog/what-is-a-standard-essential-patent-sep-.

[55] Id.

[56] Id.

[57] See Kevin Madigan & Adam Mossoff, Turning Gold to Lead: How Patent Eligibility Doctrine Is Undermining U.S. Leadership in Innovation, 24 George Mason L. Rev. 939, 955 (2017), https://ssrn.com/abstract=294343; see also Eric Schmidt et al., Final Report, National Security Commission on Artificial Intelligence, NSCAI (2021), at 201-202, available at https://www.dwt.com/-/media/files/blogs/artificial-intelligence-law-advisor/2021/03/nscai-final-report–2021.pdf.

[58] See U.S. Leadership in AI: A Plan for Federal Engagement in Developing Technical Standards and Related Tools, NIST (2019), available at https://www.nist.gov/system/files/documents/2019/08/10/ai_standards_fedengagement_plan_9aug2019.pdf; see also AI Risk Management Framework, NIST, https://www.nist.gov/itl/ai-risk-management-framework (last visited Aug. 2, 2024).

[59] Navdeep Gill, Abhishek Mathur, & Marcos v. Conde, A Brief Overview of AI Governance for Responsible Machine Learning Systems, arXiv (Nov. 21, 2022), https://arxiv.org/pdf/2211.13130.

[60] The Ethics Certification Program for Autonomous and Intelligent Systems (ECPAIS), IEEE SA, https://standards.ieee.org/industry-connections/ecpais (last visited Aug. 2, 2024)

[61] See also IEEE Standard Model Process for Addressing Ethical Concerns during System Design, IEEE SA, https://standards.ieee.org/ieee/7000/6781 (last visited Aug. 2, 2024); 7000-2021 – IEEE Standard Model Process for Addressing Ethical Concerns during System Design, IEEE Xplore, https://ieeexplore.ieee.org/document/9536679 (last visited Aug. 2, 2024), (Aiming to develop organizational standards for incorporating ethical considerations into system design); The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, IEEE SA, https://standards.ieee.org/industry-connections/ec/autonomous-systems (last visited Aug. 2, 2024), (General project for developing ethical standards for technology).

[62] Congratulations, AI Trailblazers! Iso/Iec Jtc 1/Sc 42 Receives Prestigious Iso Lawrence D. Eicher Award, ANSI (Sep. 22, 2023), https://www.ansi.org/standards-news/all-news/2023/09/9-22-23-congratulations-ai-trailblazers-isoiec-jtc-1sc-42-receives-prestigious-iso; ISO/IEC JTC 1/SC 42 Artificial intelligence, ISO, https://www.iso.org/committee/6794475.html (last visited Aug. 2, 2024).

[63] Id.

[64] The Use of Artificial Intelligence in Health Care: Trustworthiness (ANSI/CTA-2090), Consumer Tech. Assoc. (2021), https://shop.cta.tech/products/the-use-of-artificial-intelligence-in-healthcare-trustworthiness-cta-2090.

[65] Id.

[66] U.S. Leadership in AI: A Plan for Federal Engagement in Developing Technical Standards and Related Tools, NIST 18 (2019), available at https://www.nist.gov/system/files/documents/2019/08/10/ai_standards_fedengagement_plan_9aug2019.pdf.

[67] Id. at 22.

[68] Id. at 19.

[69] NIST, supra note 58.

[70] Majority Leader Schumer Floor Remarks on the Release of the Roadmap for AI Policy by the Senate Bipartisan Senate AI Working Group, Senate Democrats (May 15, 2024), https://www.democrats.senate.gov/newsroom/press-releases/majority-leader-schumer-floor-remarks-on-the-release-of-the-roadmap-for-ai-policy-by-the-senate-bipartisan-senate-ai-working-group.

[71] See Model Legislation: Responsible Advanced AI Act, Center for AI Policy (Apr. 9, 2024), https://www.aipolicy.us/work/model.

[72] Id.

[73] See, e.g., The Tj Hooper, 60 F.2d 737 (2d Cir. 1932) (industry standards can inform the bounds of negligence even though they do not necessarily define the limits of negligence; instead, what is “reasonably prudent under the circumstances” might go beyond current practices); Daubert v. Merrell Dow Pharmaceuticals Inc., 509 U.S. 579 (1993), (consideration of whether the claimed expert scientific methods adhere to standards recognized in the relevant scientific community.); Nikolov v. Associated Env’t Servs., 52 F. App’x 975, 976–77 (9th Cir. 2002).

[74] Knut Blind, Sören S. Petersen, & Cesare A.F. Riillo, The Impact of Standards and Regulation on Innovation in Uncertain Markets, 46(1) Research Pol’y 249 (2017), https://doi.org/10.1016/j.respol.2016.11.003.

[75] Id. at 251.

[76] Id.

[77] Id. at 253.

[78] See infra, nn. 99-105 and accompanying text.

[79] See Jennifer Huddleston, The Price of Privacy: The Impact of Strict Data Regulations on Innovation and More, American Action Forum (Jun. 3, 2021), https://www.americanactionforum.org/insight/the-price-of-privacy-the-impact-of-strict-data-regulations-on-innovation-and-more/#ixzz8hmX9sfiS (Discussing how the GDPR’s privacy regulations, which generate large compliance costs, affect innovation in the EU).

[80] See, e.g., Mikolaj Barczentewicz, Does the DMA Let Gatekeepers Protect Data Privacy and Security?, Truth on Mkt. (Apr. 4, 2024), https://truthonthemarket.com/2024/04/04/does-the-dma-let-gatekeepers-protect-data-privacy-and-security (Noting how the EU’s interoperability mandates work at odds with its privacy mandates, putting firms in a bind when considering how to design their business processes. This extends to AI developers insofar as strict standards that operate at odds with each other will ultimately make certain kinds of AI behavior prohibited).

[81] We begin to strain the meaning of “commercialization” in this sense, as this would essentially turn related AI research and development into a commons. Thus, it’s perhaps better to think of this more generally as a distribution method, as the incentives to fully develop the technology that attend commercialization processes will be rather attenuated.

[82] John R. Thomas, Compulsory Licensing of Patented Inventions, Cong. Research Serv. (Jan. 14, 2014), https://crsreports.congress.gov/product/pdf/R/R43266.

[83]Marianna Mazzucato, Marietje Schaake, Seb Krier, & Josh Entsminger, Governing Artificial Intelligence in the Public Interest, Stanford Cyber Policy Center (2022); see also Bryanna Devonshire & Nicolas Harris, Government Regulation of AI – When Is It Coming?, Seacoastonline (Aug. 24, 2023), https://www.seacoastonline.com/story/business/ 2023/08/24/government-regulation-of-ai-when-is-it-coming/70671063007.

[84] Part II — Standards Concerning the Availability, Scope and Use of Intellectual Property Rights, Art. 31, WTO, https://www.wto.org/english/docs_e/legal_e/27-trips_04c_e.htm (last visited Aug. 2, 2024).

[85] John R. Thomas, supra note 82.

[86] Satellite Television Extension and Localism Act of 2010, S. 3333, 111^th Cong. (2010).

[87] 17 USC § 115.

[88] William N. Monte, Compulsory Licensing of Patents, 25(3) Info. & Comm. Tech. L. 246 (2016).

[89] Id.

[90] Funding a Revolution: Government Support for Computing Research, National Research Council (1999), https://nap.nationalacademies.org/read/6323/chapter/1#vii.

[91] Id.

[92] Kristian Stout, Biden’s AI Executive Order Sees Dangers Around Every Virtual Corner, Truth on Mkt. (Nov. 1, 2023), https://truthonthemarket.com/2023/11/01/bidens-ai-executive-order-sees-dangers-around-every-virtual-corner.

[93] See John R. Thomas, Compulsory Licensing of Patented Inventions, Cong. Research Serv. (Jan. 14, 2014), available at  https://crsreports.congress.gov/product/pdf/R/R43266.

[94] Geoffrey A. Manne & Dirk Auer, From Data Myths to Data Reality: What Generative AI Can Tell Us About Competition Policy (and Vice Versa), CPI Antitrust Chronicle (Feb. 23, 2024), https://laweconcenter.org/resources/from-data-myths-to-data-reality-what-generative-ai-can-tell-us-about-competition-policy-and-vice-versa.

[95] Id.

[96] Id.

[97] Id.

[98] Id.

[99] Required Rulemaking on Personal Financial Data Rights, Docket No. 2023-CFPB-0052, CFPB (Jun. 11, 2024), available at https://files.consumerfinance.gov/f/documents/cfpb_personal-financial-data-rights_final-rule_2024-06.pdf.

[100] Giuseppe Colangelo, Open Banking Goes to Washington: Lessons from the EU on Regulatory-Driven Data Sharing Regimes, 54 Computer L. & Security Rev., https://doi.org/10.1016/j.clsr.2024.106018.

[101] See, e.g., Ivan Bosch Chen et al., A Study on the Application and Impact of Directive (EU) 2015/2366 on Payment Services (PSD2), Publications Office of the European Union (2023), https://data.europa.eu/doi/10.2874/996945 (Estimating that TPPs spent €35 million on problems linked to accessing APIs, and €140 million on maintaining legacy systems, due to APIs not working properly).

[102] See Miguel de la Mano & Jorge Padilla, Big Tech Banking, 14 J. Competition L. & Econ. 494, 503 (2018). It’s also important to note that none of this occurs in a vacuum. In some respects, open-banking rules are broached as a way to deal with the prevalent practice of “screen scraping” in order to enable different fintech services to work. See Laurent van Huffel, From Screen Scraping to Open Banking, BAI (Oct. 20, 2023), https://www.bai.org/banking-strategies/from-screen-scraping-to-open-banking.

[103] See Barczentewicz, supra note 80.

[104] Id.

[105] See Open App Markets Act, S.2710, 117th Cong. (2022), https://www.congress.gov/bill/117th-congress/senate-bill/2710/text.