CMA outlines initial concerns involving AI foundation models

A recently published CMA update paper (24 pages / 3.2 MB) identifies three potential risks to fair, open, and effective competition posed by AI foundation models – the models that are the underlying systems that power AI applications. Foundation models source input and provide outputs, such as text or images, based on human prompts.

The CMA is concerned that the businesses which control the main inputs for developing foundation models, such as computing power, data, or expertise, could prevent competitors from accessing those inputs. For instance, businesses could restrict access to data that would allow competitors to build their own foundation models. This could ultimately reduce choice for consumers using AI, with traffic diverted towards a small number of larger technology firms.

The second risk identified by the CMA may occur if technology businesses exploit their existing market positions to limit business or consumer choice of AI services. Business and consumer users should be able to choose the foundation model service they use. However, as the existing models are already integrated with certain technologies, such as mobile phones and preferred search engines, users may find their choices restricted. If left unchecked, the CMA is concerned that this could result in fewer choices, lower quality and even higher prices for AI powered products and services.

Finally, the CMA is concerned that partnerships involving some of the main participants in AI innovation could entrench existing positions of market power. As an example, it cited the potential for strategic investments made by firms into leading suppliers of AI foundation models to limit access points or routes to market. Its update report identifies an “interconnected web” of over 90 partnerships and strategic investments involving a number of large firms, posing a potential risk to competition by limiting consumer options for AI use. Although the CMA does acknowledge that such partnerships may potentially lead to “pro-competitive benefits” and therefore “each partnership must be assessed on its individual facts and merits”.

In essence, the three risks the CMA identifies concern (1) upstream and (2) downstream aspects of the AI foundation model value chain, and (3) partnerships that could reinforce existing market positions in both.

The update paper sets out six principles the CMA has formulated to help guide the AI foundation models sector towards positive outcomes both in the areas of competition law and consumer protection. The proposed principles are “access” to key inputs, “diversity” of foundation models and model types, “choice” for business and consumers in which foundation models they deploy and how, “fair dealing” where firms refrain from anticompetitive bundling, tying or self-preferencing, “transparency” for consumers and businesses about the risks and limitations of foundation models, and “accountability” of developers and deployers for the outputs of foundation models.  

These principles have been revised from seven initial principles the CMA published in September 2023. The previous “flexibility” principle is now reflected in the revised “choice” and “fair dealing” principles.

The update paper explains how the three areas of competition risk perceived by the CMA may be mitigated if firms align their business practices with the AI principles of access, diversity, choice, and fair dealing.  The principles of transparency and accountability, on the other hand, are particularly relevant to AI consumer protections issues, according to the CMA.

The CMA is also considering issuing proactive guidance to firms on how to comply with consumer law in AI-related markets if the CMA identifies uncertainty or particular issues that need clarification.

CMA Chief executive Sarah Cardell said that the authority is “keeping very close watch on current and emerging partnerships” to understand any impact on competition. CMA action may include the use of merger control powers to assess whether, and in what circumstances, any arrangements fall within the merger control rules and whether they raise any competition concerns.

To supplement its update paper the CMA also published a more detailed technical update report, which outlines stakeholder feedback received by the CMA during its latest work and relevant market developments relating to the six AI principles.

While the CMA is already using its existing competition law powers to examine AI related issues in areas such as cloud services and partnerships, the transformative Digital Markets, Competition and Consumers Bill (DMCC Bill) – once enacted – will substantially strengthen the CMA’s powers to intervene and promote competition in digital markets and to enforce consumer protection laws. 

In a recent speech commemorating the CMA’s 10-year anniversary its Chair, Marcus Bokkerink, noted that the CMA “will evaluate developments around AI foundation models” when prioritising which technology firm “digital activities” to regulate under the UK’s forthcoming ex ante digital competition rules. The new UK regime that the DMCC Bill will establish broadly mirrors the EU’s Digital Markets Act.

The CMA will continue its AI foundational models work throughout 2024 with plans to publish a further update on AI foundation models this autumn.