Guiding the development of AI: the CMA’s initial report on AI Foundation Models

Written By

saskia king Module
Dr. Saskia King

Partner
UK

I am a partner in our Competition & EU Law team in London with over 18 years' experience at the cutting edge of UK and EU competition law and policy having worked at regulators, competition authorities, in academia and private practice, with a particular focus on regulated sectors such as payment systems as well as sport, retail, consumer, financial, technology and communications markets more widely.

toby bond module
Toby Bond

Partner
UK

I'm a partner in our Intellectual Property Group. Having studied physics at university, I'm fascinated by technology and the way in which it is reshaping our world.

On 18 September 2023, the Competition and Markets Authority (CMA) published its initial report on AI Foundation Models (FMs) following a review into the market for FMs from a competition and consumer protection perspective.

As part of the report, the CMA proposed a number of principles to guide the development of the market: accountability, access, diversity, choice, flexibility, fair dealing, and transparency. These principles aim to guide the ongoing development and use of FMs to: (i) help people, businesses, and the economy fully benefit from the innovation and growth FMs can offer and (ii) ensure that competition and consumer protection remains an effective driving force as the development and deployment of FMs evolves.The CMA has clarified these principles are not the finished article. The CMA will continue engaging with stakeholders and review their uptake with a further update due in early 2024.

What are Foundation Models?

FMs are a type of AI technology trained on vast amounts of data and which can generate new outputs in response to user prompts. Text, computer code and image generation are the most common applications at present, although other areas such as video and audio generation are not far behind. Unlike narrow AI models that are only trained to perform a single task, FMs can be adapted to a wide range of tasks and operations (On the opportunities and risks of foundation models). This is particularly the case for large language models (LLMs) such as LlaMa 2 (Meta), PaLM 2 (Google) and Claude 2 (Anthropic), which can generate text for a wide range of purposes. The flexibility of FMs in general and LLMs in particular, opens opportunities for applications in a wide range on industry sectors.

The CMA Review

The CMA launched its review in May 2023, in response to the UK government’s March 2023 White Paper on AI which asked UK regulators to consider how the innovative development and deployment of AI can be supported.Over the past few months, the CMA has been looking at three levels of the value chain: (i) development of FMs, (ii) user applications of FM and its use in other markets, and (iii) consumer experience when using these new AI tools (AI Foundation Models Review: Short Version, Paragraph 1.5). The CMA has engaged with over 70 stakeholders, including a range of FM developers, businesses deploying FMs, consumer and industry organisations and academics, and considered publicly available information including the latest AI research.

The Report

Themes:

The report focuses on the following themes (AI Foundation Models Initial Report, Paragraph 1.9):

  1. Competition in the development of FMs: especially regarding access to the key inputs required to develop FMs.

  2. Impact of FMs on competition in other markets: the report highlights the CMA’s concerns about firms unfairly gaining or entrenching market positions through leveraging their positions in adjacent downstream markets or in the upstream development of FMs, including as providers of key inputs to FMs.

  3. Consumer protection: particularly regarding the reliability and accuracy of information to consumers, and consumer understanding. The CMA emphasised it will bring consumer enforcement action where appropriate.

Principles:

The report then proposes the following principles (AI Foundation Models Initial Report, Paragraph 7.1):

  1. Accountability: FM developers and deployers are accountable for outputs provided to consumers;
  2. Access: ongoing ready access to key input such as data, compute, expertise and capital without undue restrictions;
  3. Diversity: sustained diversity of business models, including both open and closed source models;
  4. Choice: a range of deployment options for businesses so they can decide how to use FMs;
  5. Flexibility: to allow users to switch or use multiple FMs according to need;
  6. Fair Dealing: no anti-competitive conduct such as self-preferencing, tying or bundling, especially from vertical integration;
  7. Transparency: to allow consumers and businesses to make informed choices.

What could undermine these principles:

The CMA considers the following could undermine the above principles:

  1. Mergers or acquisitions which could lead to a substantial lessening of competition in markets for the development or deployment of FMs;
  2. Firms using their leading positions in key markets to block innovative challengers who develop and use FMs;
  3. Undue restrictions on firms’ ability to switch between or use multiple FM providers;
  4. The development of ecosystems that unduly restrict choice and interoperability;
  5. Firms with market power in FM development or deployment engaging in anti-competitive conduct such as the tying or bundling of products and services; and
  6. Consumers receiving false and misleading content from FM services that impacts or is likely to impact their decision-making.

What’s next?

In order to develop the proposed principles further, the CMA will now start a significant programme of engagement with a wide range of stakeholders, including consumer groups, FM deployers, new entrants, academics, government, fellow regulators and leading developers (including Google, Meta, and Anthropic). This will take place in the UK, US and elsewhere over the coming months.

The CMA is currently focused on a “collaborative approach” and will seek “views both on the report overall and on the principles themselves.” As this is only a market review – distinct from a market study (Market Studies and Market Investigations: Supplemental guidance on the CMA's approach, Paragraph 1.9)  or a market investigation – the CMA cannot compel stakeholders to respond. Therefore, any information provided to the CMA is voluntary.

If a stakeholder does choose to respond, it should consider carefully whether any information it provides to the CMA is confidential or commercially sensitive and make this explicitly clear. Although the CMA is under various statutory obligations to protect confidential information, it may in certain circumstances disclose the information to third parties (such as government departments, enforcement authorities, or even other parties providing information to the CMA) for the purpose of facilitating any further related work (CMA6 Transparency and disclosure: Statement of the CMA’s policy and approach, Paragraph 2.7-2.8). The CMA may also use the information provided to take enforcement action against businesses operating in the AI market. Not least, any information provided can also be shared within the CMA and potentially other competition or regulatory authorities.

The CMA will publish an update on the principles in early 2024, and how they have been received and adopted, and also reflecting on further developments in the market. The CMA has made clear it will bring both competition and consumer enforcement action where appropriate (AI Foundation Models Initial Report, Paragraph 5.84 and 6.6). It is therefore important for businesses to comply with their obligations under competition law and consumer protection law regardless of market developments.

Commentary

The CMA’s focus on AI

The CMA has already been focused on AI and its significance for some time, so it is progressively building on its internal know-how and understanding. Alongside the ICO and Ofcom, the CMA is part of the Digital Regulation Cooperation Forum, which recently published a paper on algorithmic systems and transparency. A major part of the CMA’s knowledge basis is a result of work conducted by the Digital Markets Unit (DMU) established in the CMA back in 2021. The DMU will oversee a new regulatory regime designed to regulate the most powerful digital firms once the CMA gains its new regulatory and enforcement powers under the Digital Markets, Competition and Consumer Bill currently going through Parliament.

The pace of development in AI continues to be fast and it is notable that the CMA and a number of key regulators are joined up and working collaboratively alongside Government.

More collaborative approach from the CMA

The CMA has been adopting a more collaborative approach to its stakeholders, particularly since Sarah Cardell’s appointment as the Chief Executive of the CMA in December 2022 (and whilst interim Chief Executive). For instance, the CMA highlights its open-door policy in its guidance on environmental sustainability agreements, whereby businesses can approach the CMA for informal guidance. A similarly collaborative approach was adopted in respect of the Google Privacy Sandbox investigation, where the CMA engaged extensively with Google on the design and development of its Privacy Sandbox.

Likewise, this initial report underlines the CMA’s more collaborative stance, particularly when dealing with new markets. The report highlights how “[o]verly burdensome regulation may make it more difficult for competition and innovation to flourish and at worst may lead to concentration and become a significant barrier to entry in its own right.”

At this juncture, the CMA is taking the approach that to devise meaningful, useful and robust principles, the CMA needs to work with digital and tech firms rather than against them.

For further information or assistance, or if you would like to discuss any issues, please contact Saskia King and Toby Bond.

VISIT OUR COMPETITION & EU HOMEPAGE

Authored by: Dr. Saskia King, Toby Bond, Quinn Liang.

Latest insights

More Insights
Curiosity line teal background

A Deep Dive into China’s Network ID Proposal

Nov 06 2024

Read More
mountain scape

European Union Artificial Intelligence Act Guide

Nov 06 2024

Read More

California’s AI bill vs. the EU AI Act: a cross-continental analysis of AI regulations

Nov 06 2024

Read More