The Competition and Markets Authority (“CMA”) has published an update paper as part of its review into ‘AI foundation models’ (“FMs”), which are general purpose AI systems capable of carrying out a range of tasks and that underpin various generative AI applications.

In September 2023, the CMA published a report (previously commented upon here) proposing a set of principles designed to guide the development of these markets and promote positive outcomes for competition, consumer protection, and the broader economy. Since the publication of this report, the development and use of FMs has continued to evolve at a rapid pace, with several industry announcements covering new model launches, investments, partnerships, and innovations. This swift advancement, coupled with the CMA’s ongoing market monitoring and stakeholder engagements, has led to growing concerns about potential risks to competition and consumer protection.

The update paper comes as the Chief Executive Officer of the CMA, Sarah Cardell, delivered a speech underscoring the transformative potential of FMs and their rapid uptake as a “paradigm shift” for societies and economies, while expressing the CMA’s apprehension about the growing influence of a handful of large technology firms across the FM value chain (predominantly, the GAMMAN^[1] companies) on competition and consumer protection. These firms, which already hold significant market power in key digital markets, have strong positions in both the development of FMs, including the supply of critical inputs like data, and in the deployment of models through key access points and routes to market.

The update paper echoes these concerns, drawing attention to the risk of the market developing a ‘winner-takes-all’ dynamic as a result of the small number of developers of FMs and the growing relationships and partnerships between them. The update paper also sets out a revised set of principles based on stakeholder feedback and market updates, an updated assessment of the FM market, and outlines further actions the CMA is taking and considering to promote fair, open, and effective competition.

The CMA identifies three particular interlinked risks to fair, open, and effective competition:

• Risk 1 – Restricting access to critical FM inputs

The CMA is concerned that firms controlling key inputs like compute power or data for developing FMs may restrict access to these inputs. This could prevent rivals from building competitive FMs that challenge the dominant players’ existing positions.

As an example, large technology firms with extensive data sets or AI compute infrastructure could theoretically limit access, making it harder for potential disruptors to develop advanced language models capable of challenging their core businesses, such as search engines.

• Risk 2 – Distorting choice in FM services

The major players could exploit their control over consumer access points like mobile ecosystems, search engines, and productivity suites to give preferential treatment to their own FM services or those of partners. This could impede rival FMs and downstream applications from competing fairly.

For instance, the CMA highlights in the report how Microsoft is integrating OpenAI’s language models into its Office software and Windows operating system through the Copilot feature. If businesses and consumers cannot easily switch to alternatives, this tight integration could entrench Microsoft’s position and stifle competition in the FM space.

• Risk 3 – Anti-competitive partnerships

Partnerships between big tech firms and FM developers risk reinforcing their market power throughout the value chain, resisting potential competitive threats.

Notably, the CMA has identified in its report over 90 such partnerships involving Google, Amazon, Microsoft, Meta, Apple and chip supplier Nvidia since 2019. While partnerships can of course promote innovation, the regulator is concerned that dominant players may use them to steer the market for their own benefit and limit the diversity of FM types.

To mitigate these risks and guide development, the CMA has proposed updates to its original set of principles, in order to further promote competition and consumer outcomes in the use of FMs. The principles as they now stand are:

1. Access – Ensuring ongoing, ready access to key data, compute, expertise and funding without undue restriction.
2. Diversity – Maintaining diversity in models and model types for businesses and consumers to choose from.
3. Choice – Enabling businesses and consumers to choose how they use FMs and switch between options without lock-in.
4. Fair Dealing – Preserving confidence that the best products will win out and preventing anti-competitive conduct like bundling or self-preferencing.
5. Transparency – Providing businesses and consumers with the right information about FMs’ uses and limitations.
6. Accountability – Ensuring developer and deployer accountability for development and outputs that they are in control of.

Some of the key changes to the principles include:

• Merging the ‘Flexibility’ and ‘Choice’ principles – The original ‘Flexibility’ principle has been merged into the ‘Choice’ principle. This unified principle addresses the overarching objective of enabling choice for developing, releasing, and deploying FMs.

• Expanding the scope of the ‘Diversity’ principle – The updated ‘Diversity’ principle has been broadened beyond just sustaining a diversity of FM models. It now also encompasses diversity of business models, routes for releasing models, and model types. This expanded scope aims to support the range of FM requirements across different use cases.

• Transparency for users – The revised ‘Transparency’ principle emphasises the importance of deployers clearly communicating necessary information to users. This allows users to make informed choices about FM-based services.

• Firm accountability – Amendments to the ‘Accountability’ principle stress the responsibility of firms across the value chain. They must embed mitigations and responsible practices to reduce potential consumer harm.

It is worth noting that while the AI White Paper published by Department for Science, Innovation & Technology (DSIT) proposes five overarching cross-sectoral principles for AI regulation (namely, safety, security, transparency, accountability, and redress), the CMA’s principles are intended to complement these broader principles, but with a specific lens on promoting competition and consumer interests. DSIT’s principles find parallels in the CMA’s guidelines around access to key inputs, diversity of models and businesses, choice for users, fair dealing free of anti-competitive conduct, transparency into model capabilities, and accountability across the value chain. The CMA’s principles also appear aligned with the ICO’s approach. Both regulators emphasise the importance of transparency and maintain core objectives of enabling informed choice and accountability across the AI value chain.

As the paper explains, the market is more likely to flourish if more firms have access to data, compute, and expertise. The CMA believes that adherence to its proposed principles will help create conditions necessary for fair, open and effective competition. The CMA also outlines a number of actions that it intends to take. These include:

1. Using its market investigatory and merger control powers to examine the conditions of competition in the provision of public cloud infrastructure and investigate the ‘competitive landscape’ in AI accelerator chips and other partnerships and arrangements in the FM ecosystem. It will also examine Microsoft’s partnership with OpenAI to understand its effect on competition.
2. Considering which digital activities relevant to the development of FM models might warrant investigation in the future, particularly “those digital activities that are critical access points or routes to market for FM deployment, such as mobile ecosystems, search, and productivity software”.
3. Using its new powers proposed under the Digital Markets, Competition and Consumer Bill to investigate, enforce rules and issue penalties.

As mentioned, the CMA draws particular attention to the extensive partnerships and strategic investments between the leading businesses in this area. Whilst the CMA recognises that such partnerships may be advantageous, it states that it is “vigilant against the possibility that incumbent firms may try to use partnerships and investments to quash competitive threats, even where it is uncertain where those threats may materialise”. The paper points out that the CMA is already taking action against such risks by monitoring these partnerships and ‘stepping up’ its use of merger control.

The paper also touches on risks to consumer protection that are inherent in the use of AI, such as hallucinations that mislead consumers. It warns that if unfair practices in AI-powered markets emerge, the CMA is ready to take action, and points to the new powers and penalties that are envisaged in the Digital Markets, Competition and Consumers Bill. But, there is a recognition that adopting too heavy-handed an approach will itself bring risks of setting insurmountable barriers for new entrants to the market, thereby only cementing the position of incumbents and reducing competition. Therefore, the CMA commits to working with the Government and other regulators to make “targeted and proportionate” interventions where necessary to ensure that the market can be as competitive as possible.

A further update is expected in Autumn 2024. In the meantime, the CMA plans to release a number of further publications, including: a paper on AI accelerator chips; research into consumers’ understanding and use of FM services; and a joint statement with the ICO on the interaction between competition, consumer protection and data protection in FMs.

As the FM landscape continues to evolve at pace, the existing dominant players will inevitably continue in their race integrate the technology, and the CMA (and other regulators) face pressure to keep pace. Ultimately, striking the right balance between fostering innovation and safeguarding competition and consumers in AI will likely require ongoing efforts in the development of policies and regulatory approaches. While the regulator’s ability to enforce its vision remains somewhat constrained for now, we will have to wait and see how forcefully the CMA will actually be able to wield its future toolkit.

The update paper can be read in full here.

References

[1] Google, Apple, Microsoft, Meta, Amazon and Nvidia.