The EC’s Competition Policy Brief on competition in Generative AI: a new framework for enforcement
On 19 September 2024, the European Commission (EC) published its latest Competition Policy Brief,[1] focusing on competition in Generative AI (GenAI). The publication of the brief highlights that competition enforcement in GenAI is at the top of the EC’s agenda.
Relevance and scope of the Policy Brief
Despite not necessarily reflecting the official EC position, [2] the Policy Brief offers a concrete reference framework by outlining possible relevant market definitions, theories of harm, and potential efficiency gains to be considered by the EC when assessing whether and how to enforce competition in GenAI markets.
The EC’s Policy Brief is clear evidence that competition risks in GenAI markets will be central to the EC’s agenda and that the EC will use all enforcement tools at its disposal to ensure that markets remain dynamic.
Theories of harm
The Brief identifies a number of potential anticompetitive concerns that may materialise in GenAI markets, and notes that it will remain particularly, but not exclusively, vigilant of five possible types of competition risks:
- Upstream foreclosure: the risk that incumbents with access to key components required to develop GenAI models may grant access to competitors on an exclusive basis or prevent access altogether. Examples include:
- Acqui-hires, i.e. hiring of talent which may lead to the de facto dissolution of a potential competitor by absorbing the competitor’s expert talent;
- Dominant players entering into licensing agreements to secure high-quality data from sources, preventing others from training or fine-tuning their models on that data.
- Downstream distortion: incumbents offering GenAI Foundation Models [3] (FMs) may use their market power to limit choice or distort competition when distributing AI applications. Examples include:
- Tying a search engine to a FM owned by the same player;
- Preferentially displaying an incumbent’s own FM on their platform (akin to self-preferencing on app stores or search engines).
- Horizontal agreements between competitors which may reduce competitive constraints and allow for the unlawful exchange of sensitive information.
- Pricing policies adopted by vertically integrated players to margin squeeze competitors.
- Killer/reverse-killer acquisitions aimed at eliminating emergent competition so that the incumbent acquirer retains market power.
Generally, the Brief emphasises competitive concerns relating to barriers to entry and expansion in GenAI. This mirrors the CMA’s recent concern [4] over bottlenecks in critical inputs (such as GPU chips and cloud services) where incumbents have own-access or preferential access. For new entrants, securing key components risks being prohibitively expensive or near impossible.
However, the Brief also recognises that agreements between players in the GenAI market may lead to pro-competitive efficiency gains, such as combining skills which lead to better products and/or reducing costs when the supply of a critical input is limited. This is also in line with the CMA’s view that partnerships and investments by and between players may be a competitive risk or benefit depending on the circumstances. [5]
Market definitions
When investigating abuse of dominance under Art 102 TFEU, the EC must define the relevant market. The Policy Brief outlines factors that may be taken into consideration in investigations in the AI sector.
In particular, the Brief suggests distinguishing upstream and downstream markets:
Upstream markets comprise those that an AI developer requires to purchase necessary inputs such as high-quality pre-training data sets; cloud capacity; data centre services; GPU chips; and AI engineering talent.
Downstream markets are those for the sale/supply of GenAI Foundation Model [6] (FM) services, which may be further segmented into e.g. consumer-facing and business-facing models; cloud-based and on-device models.
As downstream markets are nascent and not yet fully monetised, the Brief recognises that revenue will not necessarily be a proxy of market power. Accordingly, it offers potential alternative metrics. These include: activity volume (e.g. number of users/interactions processed); costs of processing capacity (a proxy for performance quality); and downloads of the model.
Last, in line with the EC’s recently revised Market Definition Notice on digital ecosystems, [7] the Brief highlights that network effects, switching costs and other ecosystem dynamics may also be considered.
Comment
Policy briefs, despite their unofficial status, have played an important role in marking the start of competition enforcement in particular areas. By way of example, in May 2024 the EC published a brief focusing on the risks posed by wage-fixing and no-poach agreements to competition in the labour market. Only two months later, in July 2024, the EC launched its first investigation on no-poach agreements, against Delivery Hero and Glovo, under Art 101 TFEU.[8] This new Policy Brief similarly is likely to represent an important step forward in the EC’s enforcement of competition law in the GenAI sector.
As illustrated by the joint statement on competition in GenAI FMs published in July 2024 by the EC, CMA, Federal Trade Commission and Department of Justice, [9] competition authorities are collaborating closely and are pro-actively assessing the GenAI sector and the tools at their disposal to tackle risks arising out of market concentration and distortions of competition.
In identifying key competition risks, the Brief makes clear that competition enforcement tools like merger control and the implementation of the Digital Markets Act (DMA) are complementary but distinct, for example because the DMA will not cover competition concerns in GenAI services for non-designated companies.
Accordingly, we can expect the EC to continue its pro-active assessment of the sector; adopt a framework akin to that set out in the Brief; and use all the tools at its disposal (antitrust, merger control and the DMA) to enforce competition in GenAI markets.
Footnotes
[1] Kowalski et al, European Commission, Competition in Generative AI and Virtual Worlds, Competition Policy Brief No 3/2024, doi 10.2763/679899
[2] The Policy Brief does not necessarily reflect the official EC position, only of its authors. It is also not exhaustive in dealing with potential AI competition risks, as it expressly focuses on competition dynamics in GenAI markets, and not on AI as a possible tool to facilitate anticompetitive conduct such as through algorithmic collusion.
[3] The Policy Brief appears to use GenAI models as shorthand for the entire sector as well as its supply chains, despite then referring to “generative AI foundation model services”. However, it is important to note that there is a distinction between GenAI models (which the brief defines as AI models that can process various inputs (text, images, sound) depending on the data they were trained on and produce new content) and foundation models (FMs). FMs are trained on very broad data at scale and are designed for generality of output so that they can act as the base for several applications. They can then be adapted via fine-tuning or when developers “build over” FMs. For example, GPT-4 is an FM. ChatGPT and Microsoft Copilot are GenAI applications built on fine-tuned models which use GPT-4 as their base.
[4] See our commentary here: Hausfeld | Will the UK shift its regulatory approach to AI in 2024?
[5] Ibid.
[6] The Policy Brief appears to use GenAI models as shorthand for the entire sector as well as its supply chains, despite then referring to “generative AI foundation model services”. However, it is important to note that there is a distinction between GenAI models (which the brief defines as AI models that can process various inputs (text, images, sound) depending on the data they were trained on and produce new content) and foundation models (FMs). FMs are trained on very broad data at scale and are designed for generality of output so that they can act as the base for several applications. They can then be adapted via fine-tuning or when developers “build over” FMs. For example, GPT-4 is an FM. ChatGPT and Microsoft Copilot are GenAI applications built on fine-tuned models which use GPT-4 as their base.
[7] European Commission’s Notice on the definition of the relevant market, fn. 10, p. 40 https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=PI_COM:C(2023)6789&qid=1726475579651
[8] See our commentary here: Hausfeld | Competition and labour markets: Competition authorities get to work
[9] Joint statement on competition in generative AI foundation models and AI products - GOV.UK (www.gov.uk)