Is Data Really a Barrier to Entry? Rethinking Competition Regulation in Generative AI

Abstract: The rise of generative artificial intelligence (AI) technology has raised concerns that a few large companies or digital platforms will corner the market. By controlling vast troves of data necessary to train foundation models, these firms could raise entry barriers, stifle competition and innovation, and ultimately harm consumers. Contrary to this fear, thriving data markets allow developers to access training data on an open-source or commercial basis, making it so that data access is not currently a barrier to entry in the generative AI market. In fact, tech giants with access to high volumes of user data face stiff competition from independent firms and startups that can outcompete larger rivals through superior algorithms or user experiences. Technological advancements, such as synthetic data and engineering workarounds that lower data demands and training costs, also suggest that concerns about data scarcity limiting AI innovation are overblown. Additionally, the industry’s shift towards more specialized foundation models and tailored AI applications also demonstrates that access to the “right” kind of data rather than vast troves of user data is paramount, further eroding the purported advantage of large tech giants. Conversely, preemptive regulations and remedies proposed to combat future data scarcity or monopolization would likely harm AI development and innovation. Current US antitrust law instead provides a flexible, pragmatic framework for policing exclusionary conduct by firms controlling key data inputs should these concerns materialize in the future.

The rise of generative artificial intelligence (AI) has prompted global competition authorities to raise concerns that the sector is highly susceptible to monopolization. They argue that large digital platforms benefit from economies of scale and network effects advantages, giving these platforms a significant competitive advantage that could lead to anticompetitive outcomes through their control over data.  This understanding has prompted authorities to suggest preemptive antitrust interventions. Our research indicates not only that these concerns are unjustified and overstated but that preemptive antitrust interventions could harm innovation and competition.

Generative AI foundation models (FMs) use machine learning software algorithms to ascertain and predict statistical relationships between data point inputs in order to produce outputs, such as answers to questions, video, audio, images, and more.[1] These models typically rely on and are refined by[2] training on large datasets consisting of billions or even trillions of parameters.[3][4] Proponents of preemptive antitrust intervention in generative AI, including the US Federal Trade Commission (FTC),[5] theorize that a few firms with access to large troves of proprietary and user data or the resources to acquire that data[6] may develop an unassailable advantage in generative AI technology. They may then use this advantage to curtail competition and innovation by restricting their rivals’ access to data[7] and by charging users higher prices without being disciplined by competitive pressure.[8]

Concerns about large firms and digital platforms’ exclusive access to data are also driven by fears about future data scarcity. Increasingly powerful and versatile foundation models designed to perform complex applications are expected to require higher volumes of data and access to increasingly specific kinds of data to further improve.[9] In addition to larger and more diverse datasets, a larger and more diverse set of human feedback (such as user rankings of application or model outputs) may also be required to fine-tune larger, increasingly complex models.[10] Although many leading FMs were trained on publicly available data scraped from the internet, a high proportion of publicly available data has already been used in training,[11] and an increasing proportion of public data now consists of AI-generated outputs that can degrade model quality when used in training.[12]

Antitrust regulators like the FTC theorize that data scarcity will increase the advantage that large tech firms and their digital platforms have over rival AI model providers and developers due to those firms’ access to vast volumes of training data from users of their other services.[13] Regulators argue that this exclusive access to data would increase tech giants’ ability to harm consumers and would stifle innovation by restricting competition, thereby further entrenching their monopoly.[14]

Contrary to these fears, developers of FMs and the applications that use them have access to a range of data sources, including a competitive market of open-source and commercially available proprietary datasets that are open to all. Similarly, data owners have an incentive to sell their data to big and small firms alike. The history of developing digital platforms and generative AI also demonstrates that the advantages that a firm enjoys from having a large existing database of user data is generally superseded when rivals have superior software and algorithms, thus indicating that access to data is not an ‘unassailable advantage’ over the competition. The right kind of data for a model or application’s intended function is generally more important than raw data volume. Future models and applications will thus likely have more specific uses and so will be trained more efficiently with fewer data. Advancements in the ability to create and correct synthetic data also promise to help to meet data volume demands.

Despite the failure of their concerns to materialize thus far, competition researchers worried about future data scarcity for training FM models have proposed a range of solutions that could do more harm than good. For instance, mandates on digital platforms to share their data with rivals and bright-line prohibitions on mergers or business practices that may give firms more control over data, but that may otherwise benefit competition and innovation. These proposed fixes would impose costs on the creators and developers of large digital platforms that FM developers rely upon for deploying and refining their models, and for licensing data used to train the models. They could also reduce competition by deterring the creation and improvement of both the platforms themselves as well as FMs built by their owners. These proposals are thus unlikely to be warranted as ex post antitrust remedies for exclusive dealing or monopolization are adequate for addressing potential abuses of market power. There are also regulatory proposals similar to the European Union’s Global Data Protection Regulation (GDPR) that would treat individual digital platform user data as a user’s property right. These too are flawed and would worsen data scarcity issues and digital platform functionality while erecting further cost barriers to competition in both the digital platform and generative AI development space. Current legal frameworks and remedies recognized by antitrust courts provide an appropriate, flexible and adequate framework for addressing data monopolization issues should they arise.

This paper will critically appraise current and potential future concerns about data scarcity and data monopolization in generative AI markets and whether these concerns warrant antitrust intervention. It will also argue for the efficacy of current antitrust law remedies in addressing these concerns. To this end, Section I will analyze what counts as a “barrier to entry,” and whether that term applies to data in the generative AI space under current or foreseeable future market conditions. Section II will evaluate the likelihood of increasing demand for training data in generative AI innovation in relation to evidence of emerging trends, techniques, and innovations that are likely to reduce data demands and increase the cost efficiency of smaller datasets. The use of synthetic data and other engineering workarounds as well as the trend towards smaller datasets for more targeted AI applications are included in this discussion. Section III will examine how current popular regulatory proposals aimed at addressing the perceived role of data as a barrier to entry could potentially harm competition and innovation. Section IV will analyze whether existing antitrust laws and remedies can sufficiently address future barrier of entry issues around data or the potential abuse of market power over data. 

Section I: Data as a Barrier to Entry in the Present

Upfront entry costs exist in virtually every business and industry. If potential market entrants choose not to compete with established incumbents due to low expectations of success or long-term profitability, this outcome could be optimal for consumer welfare and competition. In such cases, the market may already be functioning efficiently with the existing number of firms rather than the limited number of firms or market structure being an indicator of inefficiency or barriers to entry.[15] The potential entrant’s scarce resources would be more profitably invested elsewhere, such as in other markets or business lines. For instance, in a market consisting of only a few firms that benefit from scale economies, the potential reduction in output from limited competition may be offset by the increased output resulting from the cost efficiencies of the incumbent firms.[16] There would thus be no benefits to consumers from additional market entry. Rather, there may be a loss of consumer welfare if external interventions lower entry costs, as this could lead to misallocation of resources away from where they could be more productively used. 

Some researchers describe the advantages that incumbent firms enjoy in generative AI due to data network effects as “a sizeable entry barrier, particularly for small and medium-sized (SME) enterprises.”[17] However, it is unclear whether an incumbent’s ready access to data that would-be challengers lack is necessarily a “barrier to entry” in the sense that warrants antitrust intervention. 

The price of acquiring data for training generative AI models is a capital cost. Competition policy researchers Dirk Auer and Geoffrey Manne note that “[b]ecause data is used to improve the quality of products and/or to subsidize their use, the idea of data as an entry barrier suggests that any product improvement or price reduction made by an incumbent could be a problematic entry barrier to any new entrant. This is tantamount to an argument that competition itself is a cognizable barrier to entry.”[18] From an antitrust standpoint, an entry cost should be considered a barrier to entry only if it delays or hinders market entry in a way that reduces consumer surplus or welfare compared to the potential effects of any policy, regulatory, or legal intervention intended to address it.[19] Where capital markets function efficiently and without high levels of uncertainty, scale economies and capital requirements (including data) are not a barrier to entry for potential market entrants regardless of size.[20]

There are several ways that new entrants can obtain data or the means of constructing datasets when entering generative AI markets. New entrants can access capital funding through private and public markets, and independent LLM developers raise hundreds of millions of dollars or even multibillion dollars through institutional investors and venture capitalists.[21] Developers can purchase, license, or construct training datasets through web scraping or through a thriving market of open sources as well as closed (proprietary) sources.[22] Secondary data providers, including data brokers, data warehouses, and synthetic data providers, have entered and continue entering the market to meet growing demands.[23] Rightsholders of copyrighted works can also license these works to AI developers for training or fine-tuning models.[24] Digital platforms like Meta have trained their AI models partially on public posts from users of their services, including Instagram and Facebook.[25] These public posts can also be scraped by anyone.

Commensurate with the availability of training datasets and despite concerns about potential data scarcity, the market for generative AI foundation models as of today remains fiercely competitive and dynamic.[26] Current trends indicate an increase in competition and market entry rather than a decline. In September 2023, OpenAI’s GPT-4 (which powers ChatGPT), held the vast majority of the market share and had a commanding technological lead over alternatives from big tech platforms like Google and Meta.[27] By November 2024, 16 different AI labs had produced models that exceed GPT-4’s capabilities and performance benchmarks.[28] Though some of these models, including Google’s Gemini and Meta’s LLaMA, have been produced by established tech giants, OpenAI’s GPT remains the market leader based on usage. As of January 2025, ChatGPT (59.5 percent) and Microsoft’s Copilot (14.3 percent), both of which were built off the GPT LLM, collectively hold a 73.8 percent market share among generative AI chatbot tools.[29] Google’s Gemini is in third place with a 13.5 percent market share. In terms of quarterly user growth, however, ChatGPT falls into fourth place behind ClaudeAI from start-up Anthropic, Perplexity (which was built off Meta’s LLaMA LLM), and Phind (a tool for software developers, which was built off both GPT and original LLMs). LLM start-ups continue to successfully enter the market, even outside of the text generation segment. For instance, RunwayML has thrived in video generation, and MidJourney and Stable Diffusion thrive in image generation.[30]

Continued and successful market entry by a range of firms of varying sizes that compete head-to-head or serve different generative AI market segments indicates a fiercely competitive and rapidly evolving generative AI and LLM market. Notably, generative AI markets feature a mix of independent start-ups, start-ups benefiting from commercial partnerships with larger entities (such as Amazon/Anthropic and Microsoft/OpenAI), and large tech firms. This contradicts the claim that digital platforms with large troves of user data accumulated through network effects and scale economies possess an unassailable competitive advantage. Simultaneously, partnerships with large platforms have benefited start-ups by providing access to the platform’s user base, scale economies, and network effects, all of which help to further develop and refine products. These partnerships also foster the exchange of capital investment, cloud storage, access to computing power, and other resources.[31] Evidently, superior algorithms and preferred user experiences remain a far more significant competitive advantage than access to data, further indicating that data access is not a barrier to entry.[32] The top four LLMs ranked by performance benchmarks as of November 2024 are all either newly launched or have been updated within the preceding three months, which indicates that training data cost or scarcity has not been a barrier to entry or innovation.[33]

The range of firms of different sizes that compete with one-another, including those born out of commercial partnerships between firms of different sizes, also contradicts the theory that large tech giants raise anticompetitive concerns when they move into adjacent AI-related lines of business and use advantages from their existing business lines, including access to platform data. Some fear that tech giants’ advantages allow them to restrict competition, and hold an unassailable advantage.[34] Instead, the opposite is observed. The rise of OpenAI’s ChatGPT has driven competition and innovation in the adjacent search engine market. Although Google remains the dominant market leader,[35] it is in the face of increasing competition. This is despite Google’s monopoly over search indexing and user data across billions of customers on its many mobile and web applications and tools.[36] Thirteen million American adults used generative AI as their primary tool for online searches in 2023, a figure expected to exceed 90 million by 2027.[37] In the face of this threat, Google and Bing responded to ChatGPT’s rise in 2023 by integrating generative AI tools into their search engines.[38] OpenAI expanded into search tools by launching its GPT FM-powered SearchGPT tool in October 2024.[39] Since then, it has grown exponentially at a rate that far outstrips the growth of Google’s Gemini AI-powered search tool.[40] This demonstrates that generative AI startups are not only competitive against large tech giants, they are also able to expand into adjacent markets, including those currently dominated by one or more tech giants. This is possible even when startups do not begin with access to the vast user data, network effects, and scale economies of these giants.

Section II: Data as a Barrier to Entry in the Future

It is doubtful that generative AI markets, despite being initially competitive and disruptive, will inevitably converge on just a few players with concentrated and stable market shares. As shown, the history of generative AI markets until today contradicts antitrust regulators’ concerns about data scarcity, data network effects, and economies of scale giving an unassailable advantage to tech giants. Concerns about eventual market convergence are instead based on trends observed in mature digital platform markets, such as search and social media, where a few companies with greater user-network effects provide services with high switching costs for users, thereby creating a potentially unassailable advantage. Notably, even these claims about digital markets in general have been repeatedly contradicted in practice.[41] Regardless, the three main features that supposedly make digital platform markets potentially prone to convergence on a few players, including the role of data, do not apply equally to the market for generative AI tools and models.[42] These features are economies of scale, network effects, and switching costs.

First, convergence on a few players is more likely for digital platform markets as they are characterized by economies of scale that advantage incumbents who have already paid high fixed costs and face relatively lower variable costs in the long run, unlike potential new entrants who face high entry costs.[43] Thus, only the most efficient digital platform competitors will remain after others exit the market. Although building a generative AI foundation model does entail high fixed costs, such as pretraining and acquiring data and computing resources, variable costs are nonetheless more significant for market participants than they are for many digital platforms as models must be continuously tested and refined to ensure that they remain fit for their intended purpose and adapt to new information and user demands.[44] The availability of open-source datasets for building new models drastically lowers fixed costs,[45] as do open-source or commercially licensable pre-trained foundation models for building new applications. Large digital platforms with stable and significant market shares also have an incentive to lower costs for generative AI entrants. For instance, Meta has opted to release the code for its LLaMA foundation models on an open-source basis in order to lower entry costs and attract AI developers to integrate new applications into (and thus improve) its social media platform and “ecosystem”.[46] xAI (the AI division of social media platform X) has also released its Grok foundation model on an open source basis.[47] This supports the theory that fixed capital costs do not constitute a barrier to entry for AI models because would-be entrants have equal access to the technology relied upon by incumbents.[48] This is further affirmed by the fact that a majority of popular generative AI foundation models that have entered the market, including OpenAI’s GPT-4 and xAI’s Grok, were trained entirely on data publicly available on the web.[49]

Second, digital platform markets are more likely to experience convergence on one or a few providers due to user network effects because users seek providers that are popular with other users or whose services are improved by their larger existing user bases.[50] For instance, the quality of Google’s search results is refined through data from its large user pool,[51] and Meta is popular as a social media platform due to its volume of existing users. However, users do not attribute value to generative AI models or applications based on the number of other users.[52] Training models on larger datasets of users and their feedback may enhance model quality due to having a broader sample to “learn” from and calibrate responses to. Models that possess more data parameters also benefit from large datasets as more data points allow for the drawing of more inferences about the interrelationships among the model’s parameters.[53] However, the role of raw user numbers declines significantly with the degree of specialization in the model or the application’s function because improvements in the model or application’s output depend more upon the dataset’s quality and suitability for purpose than its size.[54] For instance, an AI tool or model developed for facial recognition in a large American city will produce more accurate results when trained on a dataset of 100,000 racially and gender-diverse faces than when trained on a dataset of 1,000,000 faces of a single gender and/or race. Thus, the appropriateness and scope of a specific training data set for its intended use are more crucial for generative AI development and model training than the data set’s size or the number of users from which it may be harvested. This is why a chatbot trained on a larger set containing incomplete or contradictory information—information lacking context or misinformation—will generally produce lower quality and less accurate responses to queries than one trained on a smaller albeit representative data set that lacks acontextual information, contradictory data inputs, or misinformation. Indeed, multiple studies have found that models trained on smaller high-quality datasets can outperform larger models on accuracy and efficiency.[55] In addition to requiring less data due to having fewer data parameters, “smaller language models” and those tailored for specialized tasks also generally have less computational power and memory requirements.[56]

Technological advances, like synthetic data, are also reducing the need for original data and how much it costs to train and fine-tune models. Synthetic data replicates real-world data by making a sample of artificial data that is representative of an existing dataset’s features.[57] It can “combat the data scarcity, privacy concerns, and algorithmic biases commonly used in machine learning applications.”[58] Although synthetic data comes with the same limitations as the sample of real-world data from which it is derived,[59] and cannot completely replace real-world data, models trained on a combination of synthetic and real-world data have outperformed models trained on larger datasets of real-world data on specific functions.[60] Overreliance on synthetic data for machine learning can result in “model collapse,” whereby the accuracy or quality of the model’s results decline over time.[61] However, synthetic data techniques and the quality of synthetic data are also likely to be refined over time and with the benefit of technological improvements. 

Decentralized AI is another innovation that promises to reduce entry barriers to AI development by making data more accessible. At the time of writing, most major AI models (like OpenAI’s GPT, Meta’s LLaMA, Google’s Gemini, Anthropic’s MistralAI, etc.) are built on centralized servers that host datasets and code. They are controlled by centralized corporations that hire workers for machine learning, data engineering, data cleaning, and coding. These corporations, or those they enlist, are also responsible for acquiring data, computation power (hardware), cloud storage, energy, and other necessary inputs. By contrast, decentralized AI involves using blockchain technology to host AI models and their underlying datasets and other components across a decentralized, collaborative network of servers (or “nodes”), thereby allowing for open, real-time, and continuous collaboration between individuals.[62] Digital assets, such as cryptocurrency tokens, are provided as rewards to those who contribute to the network by providing data, computing power, coding, machine learning, data cleaning, application integration, or hosting services that support the project.[63] “Subnets” are another way that decentralized AI can foster competition and counter potential data entry barriers to generative AI development in the future. These are smaller networks of collaborators within the larger decentralized blockchain network that focus on specific projects. For instance, Bittensor is an open-source decentralized AI protocol maintained and developed by the Opentensor Foundation.[64] The protocol is controlled and governed through community decisions across servers on its blockchain and any individual can contribute, with the foundation providing light oversight.[65] It includes 64 active subnets as of February 2025.[66] These include “Subnet 42 Real-Time Data,” which was launched by AI startup Masa, and provides data access with the goal of “democratizing” AI development.[67] These and other innovations are likely to continue reducing potential competitive barriers that could result in the future from data scarcity.

Besides technological innovation, human creativity and ingenuity also foster new means for lowering machine learning and data costs. For instance, the Chinese open-source AI foundation model DeepSeek delivers a comparable performance to leading models such as OpenAI’s GPT despite being trained at a significantly lower cost.[68] DeepSeek’s success in lowering training dataset and computation demands has been attributed to “a combination of many smart engineering choices including using fewer bits to represent model weights, innovation in the neural network architecture, and reducing communication overhead as data is passed around between GPUs.”[69] The public release of DeepSeek’s code and model weights means that these techniques can be replicated by other developers, signaling that further creative fixes in this space that lower costs and reduce entry barriers are possible.

Even when a large, high-quality dataset exists for model training or application refinement, the value of adding more data decreases after a certain point. A facial recognition model trained on a population-representative sample of 1,000,000 faces may not witness significant improvements in its performance with additional inputs of faces, and additional faces would not allow it to attract customers from a competitor application trained on its existing dataset that has a better algorithm, provides a better user experience, or is more cost or energy efficient to run. Even when additional data inputs would significantly improve model performance, this may still not be enough to outweigh a competitor’s improved algorithm, user interface, or cost efficiency. If two generative AI competitors have enough of the right kind of data, improvements in these other areas would give a greater competitive advantage in determining competitive outcomes. This is supported by studies that find that even for digital platforms, data network effects and additional data have diminishing marginal returns.[70] Furthermore, the cost of data cleaning and categorization (a necessary part of machine learning and a process that improves model and application performance given the same dataset) increases with the size of a dataset.[71] Beyond a certain point, the performance gains from increasing the dataset size may be outweighed by the rising costs of cleaning, categorizing, and processing the additional data.[72]

Finally, digital markets are sometimes prone to convergence on a few players due to the costs incurred by users when they attempt to switch to new or existing competitors.[73] In the case of foundation models, developers could incur costs by having to migrate to alternative foundation models. However, tools and platforms already exist that mitigate these costs by allowing for foundation model-agnostic application development.[74] For instance, LLM-agnostic software architecture, a framework adopted by some companies, allows for generative AI prompts to flow to different FMs depending on the application and its specific requirements. To be included, applications must be coded in a way that divorces their underlying logic from that of any specific or single FM.[75] Model switching is also facilitated by large open source platforms, such as the Hugging Face website, which hosts over 300,000 models and 250,000 datasets as of August 2024.[76] Given these factors and ongoing developments, it is unlikely that access to data will become an antitrust barrier to entry in the foreseeable future.

Section III: Analyzing Proposed Regulatory Fixes: Bright-Line Rules and Data Rights

Some policymakers and regulators believe that a proactive regulatory approach—imposing controls and guardrails—is necessary to ensure society benefits from AI technology while mitigating risks such as data misuse and anticompetitive concerns related to data control.[77] However, such regulations should only be imposed if their benefits to competition and consumers exceed their costs (inadvertent or otherwise) relative to a world where AI technology is allowed to develop without the proposed regulation.[78] Regulations should only be imposed upon markets where there is a “clear market failure, [and] also strong grounds for believing that government intervention will yield net welfare-superior results to non-intervention.”[79] Potential harms of regulation, including those introduced in the name of promoting competition, include regulatory capture that benefits vested interests while raising rather than reducing entry barriers for new or less politically and economically-resourced entrants.[80] Regulations also typically impose significant fixed costs, which place a disproportionate burden on smaller competitors in an industry or market relative to their larger peers as larger firms experiencing scale economies can spread the costs of compliance.[81]

AI is a rapidly evolving technology and generative AI markets are characterized by rapid growth.[82] Many of the markets around AI technology and its resulting products have yet to form, making the potential adverse consequences of overregulation hard to predict. Since AI innovations will power productivity in virtually every major industry and market, the adverse economic consequences of overregulation could be especially dire.[83] There could be a tremendous net loss in societal and consumer welfare from a regulation that retards the emerging technology’s exponential growth rate, even if the regulation is successful in reducing certain harms to some degree.[84]

A better approach is “permissionless innovation,” whereby new innovations aren’t assumed to cause theoretical harm unless the innovator can prove otherwise beforehand.[85] This approach allows new and rapidly evolving technologies to develop while surgically addressing harms as they emerge. For instance, the relatively light-touch approach taken by Congress in the 1996 Telecommunications Act, aimed at addressing concerns about the internet’s unfettered development, has been credited with enabling “the tremendous innovation and growth that we are still experiencing today.”[86] Preemptive regulation, on the other hand, crowds out market responses that tend to be more efficient and do not cost taxpayers the exorbitant administrative costs of enforcement.[87] When regulations do end up being adopted, they are more likely to be effective while reducing inadvertent damage to competition and innovation where the harms they target are clearly defined rather than vague.

The European Union’s [EU’s] attempts to regulate the use of data and reduce privacy harms in digital markets demonstrate the adverse impacts of imposing proscriptive regulations on innovative sectors. In 2018, the EU instituted the General Data Protection Regulation (GDPR), which requires companies and organizations to guarantee users the right to access and erase data. The GDPR also mandates positive opt-out style consent for certain data collection and use and that users can request their data and move it between different platforms, that is, render it “portable.”[88] When the GDPR was instituted, users did begin opting out of tracking and data collection, which led to less data collection and use by businesses, which hampered competition and innovation among online businesses. Web visits and revenue declined, while online search quality suffered and became less efficient due to reduced access to data, which limited the ability to tailor results to user preferences.[89] As competition in web advertising and user preference tracking decreased, the number of new firms, innovative applications, and venture capital investment also decreased.[90] Difficulties in acquiring data or using it for targeted services like advertising due to the GDPR disproportionately affected small businesses[91] and newer or potential entrants that lacked an established brand presence predating the GDPR. Because of the GDPR, the productivity of EU firms reduced relative to their American counterparts, and European tech start-ups saw reduced access to investment for research and development, further hampering innovation.[92]

The GDPR’s regulatory and compliance costs have also been significant,[93] with firms’ privacy office budgets increasing by an average of 29 percent to meet compliance requirements.[94] Additionally, American companies adhering to GDPR and similar privacy laws, such as the California Consumer Protection Act (CCPA), incur an estimated cost of $480 per individual whose data they retain.[95] These costs are more easily borne by tech giants and the firms they own: Google, Amazon, and Meta find it easier to obtain consent from individual users.[96] The costs and their relative burdens have further driven increases in the market concentration of tech giants and the largest firms at the expense of smaller players.[97] The long-term impact of the GDPR on innovation has yet to be fully measured due to the dynamic nature of the affected markets.  

The GDPR is also likely to adversely impact European AI development going forward. For instance, Italian regulators banned ChatGPT in 2023 after determining that it had no basis to collect and analyze user data under the GDPR.[98] This prohibits ChatGPT (and potentially other AI firms) from obtaining data from Italian users for fine-tuning (or refining) applications. Similar regulations that impose restrictions on data use in the name of regulating AI will also likely adversely impact AI quality and innovation[99] and harm competition against entrenched incumbents by disproportionately burdening newer and smaller rivals. 

In addition to harming competition, regulations intended to protect or otherwise benefit consumers (such as by improving privacy standards and control over data) can also harm them directly. As summarized by University of Florida law professor Jane Bambauer, “[c]ompanies [bearing regulatory compliance costs] respond by raising prices for consumers—a cost that may well be worth it to some people and in some contexts, but probably not as a general rule. And this ignores the costs of inconvenience not only in the form of the time required to click through and manage consents but also in terms of the degraded service that results from a less customized experience. For example, the introduction of GDPR seems to have resulted in consumers having to use 21 percent more search terms and access 16 percent more websites before making their online transactions.”[100]  So called “protective” regulations on AI tools are thus likely to degrade user experience and product quality while raising the costs for consumers accessing paid services.

Additionally, there must be a cost-benefit calculation for regulations that intend to counter the convergence of the generative AI sector on a few foundation model developers. It may be the case that such convergence and high levels of concentration may come to be the most efficient market structure for competition over time, as efficiencies and technical improvements beyond a certain point may be possible only where a firm can take advantage of network effects and scale economies that arise from concentration.[101] For instance, if a single firm develops a superior foundation model that outcompetes all others by attracting a large and exclusive user base on a single platform, that firm can leverage user interactions to refine and improve its generative AI model. In this situation, breaking up the firm to create more competitors and offer consumers “more options” could backfire by creating multiple inferior products. Additionally, such an intervention could discourage both the firm’s successors and potential competitors from investing in developing a superior product, fearing that achieving market dominance would again trigger regulatory intervention.

Since at least the 1966 Grinnell case, the US Supreme Court has recognized that market concentration isn’t an unqualified bad and can instead be a sign and outcome of vigorous competition rather than an indicator of consumer harm or lack of competition.[102] The same holds true when a firm is able to develop a superior product due to merging with or acquiring another firm or its assets, including data.[103] For a merger or acquisition to be deemed illegal, there must be a substantial likelihood that it will reduce competition in the relevant market.[104] This determination requires showing that the acquisition causes anticompetitive harm when weighed against likely procompetitive benefits.

Concerns about market power stemming from data acquisition in digital markets like generative AI have prompted some scholars to call for the companies that hold data to share this resource with rivals and to allow users to freely move data between competing services.[105] These data sharing or data portability mandates come with both potential benefits and costs for competition. On one hand, allowing developers and other foundation model creators to access a firm’s competitively valuable data—either for free or for a lower price—could foster competition by reducing the cost to compete and build comparable models and applications. This access would also encourage firms to compete on factors beyond data acquisition and use, such as enhancing user experience or modifying how data is used. This approach would reduce other potential purported harms of data gathering and scraping from websites, services, and databases, such as privacy and copyright infringement. 

Conversely, however, mandated data sharing would also reduce incentives for firms to gather data for developing superior AI models in the first place, since the expected profits from gathering and monetizing the data would be reduced by rivals being able to free-ride off the same training data. Thus, model developers who might otherwise have scraped or purchased the training data for their own FMs from these firms and application developers who would have built applications off models trained on those superior data sets would be left worse off. In other words, there would be no sense in mandating the forced sharing of a resource if the result would be to discourage that resource from being created in the first place. For instance, data sharing and portability mandates would reduce the incentives for digital platforms, such as search engines or social media or e-commerce sites, to invest in developing and improving the platform since they could no longer profit from the exclusive ability to monetize their data through means such as targeted advertising.[106] Platforms where user experience suffers from reduced incentives to invest in improvement would attract fewer users, and the aggregation of a large pool of user data and interactions would thus be further stymied. Rather than helping other firms, this outcome would erode their ability to get the data that they might otherwise have obtained through scraping it or licensing it from the platform owner. To reiterate, bodies of useful data may never come to exist if data sharing is enforced.[107]

Additionally, data sharing and portability mandates risk compromising user privacy and leaving sensitive data vulnerable to cybercriminals. Competitors that gain access to the data may not have the same privacy standards and protections or the same cybersecurity and anti-hacking resources as the service that originally collected the data. Services that attract users by promising better privacy protections and more user control over the data would be unable to guarantee these standards once the data is shared with other firms. Therefore, the overall risks to user privacy and data security from mandated interoperability and data portability could outweigh any benefits to privacy gained from companies shifting their attention away from data collection.

In sum, regulatory, policy, and legal fixes have been proposed in response to concerns about anticompetitive behavior in the generative AI space. These concerns center on a fear of market failures resulting from just a few firms cornering data that give their technologies an unassailable advantage. However, proposed remedies to anticompetitive behavior, such as breaking up firms or platforms or imposing interoperability or data portability mandates on them, would come with substantial costs that may outweigh their benefits. The impact of new rules for AI innovation and development is hard to predict and a negative impact that even slightly reduces the growth rate of technological development can have an exponential negative economic impact on innovation and competition in the long run. Policymakers should thus err against imposing regulations: (1) that are not surgically targeted against identifiable anticompetitive harms; and (2) where any inadvertent harms to competition and innovation from the regulation are not relatively predictable and cannot be minimized. 

For business practices and data acquisitions that are not clearly harmful or have both competitive and anticompetitive effects with uncertain outcomes, a “permissionless innovation” approach should be adopted. This means allowing firms to experiment freely unless clear harms can be identified. Where harms are identified, a new rule can be conceived if enforcing existing ones under current law and policy ex-post are found to be lacking and where the benefits of the newly proposed rule are likely to outweigh its costs relative to the status quo.

Section IV: Sufficiency of Existing Antitrust Remedies in Addressing Anticompetitive Data Concerns 

As the preceding sections make clear, antitrust enforcement and policy in high-innovation areas like generative AI should focus on evidence of anticompetitive conduct rather than hypothetical theoretical future harms.[108] Regulators should carefully weigh the likely net positive impact of any policy on competition, innovation, and consumers against unregulated outcomes. New regulations ought only be considered where current ones and laws are insufficient. In this regard, the United States’ existing antitrust laws already provide a dynamic and flexible framework for policing anticompetitive harms. These laws apply to issues arising from the use, misuse, or control of data, enabling the outright prohibition and punishment of inherently anticompetitive conduct, as well as the flexible, case-by-case assessment of business conduct and practices that may or may not have net anticompetitive impacts. The Supreme Court has criticized “[l]egal presumptions that rest on formalistic distinctions rather than actual market realities.”[109] In dynamic and rapidly evolving fields like AI where the freedom to experiment and innovate is crucial, and where today’s “market reality” may not reflect that of tomorrow, outright prohibitions on business conduct based on its form rather than its anticompetitive substance can be especially damaging to innovation and competition.

Similarly, because enforcement actions and antitrust litigation impose substantial costs and uncertainty on the targeted parties, enforcers should carefully choose when to bring cases against business conduct or mergers with ambiguous impacts on competition and innovation. The risk and threat of multi-year, multimillion-dollar potential litigation costs themselves could deter innovators from procompetitive conduct.[110] To be clear, this is not to suggest that enforcement actions should only be brought where the anticompetitive implications of a business practice can be assumed with absolute certainty.[111] That is a rare occurrence,[112] and such an approach would result in an unacceptable number of anticompetitive behaviors and deals. It would also allow firms that genuinely abuse their market power to further entrench their positions. However, the high cost of false positive errors in identifying and blocking allegedly anticompetitive practices and deals (even relative to the costs of permitting false negatives) should be acknowledged in contemplating enforcement actions and policy.[113] No cure should be worse than the disease.

In that light, consider the policing of data-driven anticompetitive conduct under current American antitrust law, which implements a flexible “rule of reason” framework. This framework weighs legitimate business justifications or procompetitive benefits of a business practice against its anticompetitive harms on a case-by-case basis to block practices that are net harmful.[114] Except for some presumptively anticompetitive practices like rigging bids and price-fixing that are deemed illegal, American antitrust law punishes unreasonable restraints of trade under Sherman Act Section 1 and exclusionary conduct or attempted monopolization under Sherman Act Section 2 through the “rule of reason.”[115] This framework offers courts the flexibility to tailor rulings and remedies to the circumstances rather than imposing bright-line rules that could inadvertently stymie competitive or pro-innovation business practices. Similarly, proposed mergers are illegal under Clayton Act Section 7 when they may substantially lessen competition or tend to create a monopoly.[116] This occurs when mergers enhance a firm’s incentive and ability to violate Sherman Act Section 1 or Section 2, either unilaterally (such as by restricting rivals’ access to a key input such as data) or through restrictive agreements with other firms to achieve the same outcome.[117] Firms that expand market share or gain a monopoly through a superior product, business acumen, or luck won’t be penalized under these laws.[118] Those that engage in practices that cannot be justified by any other reason besides excluding rivals from competing will attract punishment.[119] When a business practice or merger carries both procompetitive and anticompetitive implications, enforcers can seek binding consent decrees from the parties that limit the practice or that remedy anticompetitive concerns as a condition of settling the case or approving the merger.[120] This process ensures a faster and more cost-effective resolution of complaints while giving businesses the freedom to experiment.

Antitrust claims under the rule of reason rest on legal arguments about how a merger or business practice may harm competition and consumers, known as theories of harm. Three theories of harm recognized by the US Supreme Court under the antitrust laws could be applied to a company’s exclusionary conduct or attempted monopolization using data or foundation models trained on data. These are (1) raising rivals’ costs, (2) refusing to deal with rivals, and (3) tying or bundling. 

A dominant generative AI firm would raise the costs to compete incurred by its rivals and would thus be liable for monopolization if it acquires exclusive rights to a proprietary dataset or foundation model from a downstream supplier (or acquires the supplier itself) with the intent to restrict access for competitors, thus harming competition. The defendant would not be liable for exclusionary conduct or attempted monopolization if they could show procompetitive justifications for the practice,[121] or if the alleged key input could be substituted (such as by sourcing an alternate supplier) at low or no cost.[122] Legitimate justifications could include investing in the development and refinement of the model, in further data gathering, or investing in developing the platform through which it makes the model or dataset available.

Similarly, while firms can normally freely choose who they deal with and under what terms,[123] a firm can be held liable for “refusal to deal” if they sell a key input over which they hold a monopoly (such as a specific dataset that cannot be reasonably replicated) to another firm without a legitimate business justification.[124] This would be the case where the monopolist firm is shown to have sacrificed short-term profits for the sake of a long-term anticompetitive end.[125] Such cases are generally hard to win for plaintiffs relative to raising rivals’ costs claims. However, refusal-to-deal claims have better prospects of success when the defendant firm has previously abandoned procompetitive cooperation with its rival or rivals. This behavior gives courts a baseline from which to judge the refusal-to-deal’s anticompetitive effects and intent. Courts may also be reluctant to mandate that a firm must deal with its rivals because administering the duty to deal and setting its terms of dealing (such as price) can be a difficult exercise they may lack the expertise to undertake.[126] Thus, enforcers are likely to have greater success in prosecuting the same cases based on the raising rivals’ costs theory.

In some instances, if a firm were to condition the purchase of a product over which it has “market power” on the purchase of another one of its products, such as by only selling them together, that would constitute illegal exclusionary conduct.[127] Such arrangements are known as “tying” or “bundling” and generally have legitimate business justifications.[128] For instance, it makes sense to bundle multiple software applications on an operating system if they integrate and work more effectively together or if the bundled applications function as a single software ‘ecosystem’ with an integrated user experience under the operating system.[129] Similarly, firms that bundle foundation models or datasets with software applications, other datasets, or additional tools can be expected to argue that they are selling an integrated ecosystem of products that deliver enhanced functionality (or other benefits like data security) for users. They may argue that refinement of the model through additional feedback requires or is facilitated by bundling multiple applications built off the model. They may also argue that additional profits from the sales of the tied product help fund improvements in the tying product. For instance, the exorbitant additional costs of further data gathering and fine-tuning through machine learning might be funded through sales of additional software or web tools that integrate with the foundation model. In the case of an open-source foundation model, the provider may argue that requiring developers who build applications off the platform to license additional software before integrating their applications into the main platform may be necessary to be able to provide the model’s source code for free.[130] Regardless, if a proprietary dataset or foundation model is sufficiently unique to grant the provider market power, then a sales arrangement that forces users or application developers to make additional purchases without any legitimate business purpose may be anticompetitive monopolization. Allowing courts to make such determinations without bright-line prohibitions on potentially procompetitive tying and bundling arrangements facilitates innovation and competition.

The three theories of harm recognized by US courts under existing antitrust law provide a strong legal farmwork to hold wrongdoers liable for anticompetitive abuses, such as abuses of control over data in generative AI markets. This framework functions without requiring preemptive interventions or new regulations. Importantly, it also allows businesses the flexibility to experiment with and justify procompetitive practices.

Conclusion

Current evidence and trends indicate that a thriving market of open-source and licensable proprietary foundation models and datasets will continue to foster competition in the generative AI space while mitigating concerns about anticompetitive behavior. Evidence from the current and foreseeable state of the generative AI market indicate that these concerns have not played out. Instead, trends and evidence indicate that tech giants and large digital platforms possess no inherent unassailable competitive advantage due to their unique access to data. Indeed, commercial arrangements between these giants and smaller firms are mutually beneficial, driving innovation and competition by combining their respective synergies and granting start-ups access to capital investment. Technological innovations, such as synthetic data, are also likely to further erode potential competitive barriers by reducing the costs of gathering and using data. These developments support the case against preemptive antitrust enforcement or bright-line competition rules that limit business conduct in the name of mitigating concerns about data-driven anticompetitive behavior. Such rules would hinder innovation and technological development in the AI space, threatening to cause significant economic loss in key innovation-driven sectors where the United States competes with China. 

In contrast to potential new rules, including bright-line prohibitions, existing antitrust law offers ample means of prosecuting anticompetitive conduct, including attempts to monopolize or exclude rivals from competing. By applying a flexible and pragmatic ‘rule of reason’ framework, antitrust courts in the United States can judge business practices and proposed mergers on a case-by-case basis, weighing anticompetitive implications against procompetitive effects or legitimate business justifications for the deal or practice. As generative AI is a rapidly developing space impacting rapidly evolving markets in every economic sector, competition enforcers should continue to closely monitor and study these markets to identify changes or trends that could affect competitive conditions. Any regulatory or legal reforms to address future problems as they emerge should undergo a rigorous cost-benefit analysis prior to being undertaken. 

Notes