CMA publishes update paper on AI Foundation Models

The question

What is the state of play for the market of AI Foundation Models (FM) and what issues in the market is the Competition and Markets Authority (CMA) most concerned about?

The key takeaway

The market for AI FMs is growing rapidly with a fast expansion of publicly available models with strengthened capabilities and greater integration in other digital products and services. This growth is being driven in large part by the technology incumbents who have an upstream advantage in access to the vital resources for FM development. As a result, the CMA is developing a rigorous programme of scrutiny and guidance for these technology firms that may underly future AI regulation. The report has produced a set of guiding principles for FM development firms to adhere to, with a view to ensuring a fair and competitive market that yields the optimal development of the innovative technology for businesses and consumers.

The background

FMs are an AI technology of large machine learning models trained on vast data sets which can be adapted to a wide range of tasks. The technology has rapidly emerged in the last five years with broad applications, including chatbots, search engines, code writing assistants, image generators and healthcare research. Its fast evolution has created much debate about how it should be regulated, so in May 2023 the CMA launched an investigation into the wider FM market and the risks arising for competition and consumer protection. The CMA has now published its latest update paper of its review, which builds on the initial review with continued monitorship of how the market landscape of FMs has developed and refines its set of guiding principles to minimise the risks.

The development

The report identifies a number of evolving trends in the FM market. Demand is fast growing among both consumers and businesses, as the report found that 46% of large businesses now use AI for improving operations and product/service delivery, whilst 31% of UK adults use generative AI for personal, educational or professional purposes. This is matched in supply by a rapidly expanding variety of FMs available publicly, with over 120 FMs released since September 2023 bringing the overall number to over 330 known FMs globally. These new releases are increasingly capable and multimodal, able to handle text, audio, image and video inputs of increasingly larger amounts of data. They are also being integrated into numerous other digital products and services, such as productivity software or social media, which is fast multiplying its use-cases and uptake among consumers and businesses alike. 

The development of these FMs is still found to hinge on a few critical factors, which the CMA fears will create a bottleneck for competitive development. These include:

• AI chips: the availability of AI accelerator chips remains crucial yet constrained, although signs suggest market diversification. While Nvidia continues to lead, competitors like Intel, AMD and cloud service providers (such as Amazon, Microsoft and Google) have released or announced future releases of AI chips
• compute: access to computing resources via the cloud remains key, with strategic partnerships and investments with cloud service providers (including Amazon, Microsoft and Google) often enabling this access to compute for developers
• data: training and fine-tuning data is essential for FM development and developers are agreeing deals with owners of proprietary data sources, such as publishers and online forums, to gain access. Alternatively, developers are employing synthetic data or granting website owners opt-in controls to allow their content to train FMs
• expertise: unable to fill the demand for technical expertise through partnerships, a high-pay war for talent among AI researchers has emerged between competitor developers. 

The CMA’s “strongest concerns” for the FM market are those few large technology incumbents whose strong upstream positions in these areas provide key market powers. This significant advantage is seen as a core risk to fair, open and effective competition. The concern is that:

• these firms control and can restrict access to the main “critical inputs” to develop FMs: training data and technical expertise. The CMA’s concern relates to the industry as a whole. It is investigating the competitive landscape of both public cloud infrastructure services and AI accelerator chips, the key infrastructure and hardware behind FM development
• they already have powerful positions in consumer or business-facing markets where they can distort choice of FM services. The report also considers how consumers may be harmed through FMs, such as through hallucinations and flaws which mislead consumers
• they can establish partnerships with each other to reinforce this market power throughout the value chain. The report found that the FM value chain is therefore already becoming more interconnected through vertical integration deepening relationships between firms.
• To address these risks, the CMA delivered an updated set of principles to guide the market’s development to more fair and effective competition, with a greater emphasis on monitoring the market’s integration through acquisitions and partnerships. The CMA also intends to produce guidance to firms. Joint research on consumers’ use of FM services will be published with the Digital Regulation Cooperation Forum, and the CMA plans to collaborate with the ICO on a joint statement of the interaction between competition, consumer protection and data protection in FMs.

Why is this important?

The report highlights another area of scrutiny that the CMA is likely to focus its new emboldened powers on, following the passing of the Digital Markets, Competition and Consumers Act (DMCCA) into law. The DMCCA granted the CMA’s Digital Markets Unit extensive regulatory powers to issue specific conduct requirements on the largest technology firms and make interventions on conduct with potentially adverse competition. This report summarises some of the key competition concerns which the CMA has about these digital firms, providing an insight into their likely enforcement priorities ahead of the DMCCA coming into full force. Large technology incumbents investing in partnerships across the FM development should therefore pay close attention to the CMA’s AI principles and align their practices with these standards as much as possible.

Any practical tips?

FM developers and large technology firms investing in the FM space should review their strategy against the CMA’s AI principles to ensure it is meeting their standards for competition and consumer protection. For more discussion about FM models and related AI implications, please refer to our article “What is a foundational model?“ published in RPC’s AI Guide.

Summer 2024