AI CHALLENGES IN COMPETITION LAW

HOW ARE REGULATORS RESPONDING?

Aymeric de Moncuit, Daniel Vowden, Kathryn Lloyd, Nadia Dhorat, Ora Nwabueze
and Sarah Wilks of Mayer Brown discuss the key risks that competition authorities
in the EU, the UK and the US may think that AI could pose to competitive markets,
and the tools that they may seek to use in this area.

The transformative potential of AI, together AI encompasses a range of data-based digital tools and not be too slow to act, at
with its unprecedented global growth, tools, from problem-solving algorithms to the same time, there is a risk that intervening
has made competition authorities around large language models (LLMs) (see box “AI before building up sufficient understanding
the world take stock. AI offers significant basics”). AI is “intelligent”, which means reduces innovation and healthy competition.
benefits to businesses and consumers, that it can create and generate new content
such as reducing human error as well while being able to learn and refine its This article explores how legislators and
as enhancing efficiency and innovation. output over time, making it less predictable competition authorities are responding to
However, some competition authorities may and less subject to human oversight. While these perceived risks, and the key challenges
believe that AI also brings new competition competition authorities are familiar with that AI may pose to competitive markets in
concerns, not least because the largest, electronic communications, algorithms and different areas of competition law.
and therefore most popular, language digital ecosystems, and have developed
models need so much data and computer extensive expertise in the data sector more REGULATORY APPROACHES
power. Competition authorities therefore broadly, AI may be perceived as presenting
see themselves as, on the one hand, facing new challenges. Competition authorities around the world
the challenge of enabling consumers to have been seeking to better understand AI
experience the benefits of the emerging One concern is that the rapid progression of AI and are assessing how best to respond to
markets that are enabled by AI while, on the technology means that traditional regulatory the challenges of regulating it (see News brief
other hand, minimising the associated risks tools may be too slow to protect competition “International developments in AI governance:
of potentially anti-competitive behaviour effectively. Although regulators have said that same goal, different paths”, www.practicallaw.
and practices. they are keen to learn lessons from earlier com/w-041-5134). Regulators have stressed

© 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine. 1
 the need to actively regulate and manage
the development of AI markets, rather than AI basics
allow for digital self-regulation. But different
jurisdictions are taking divergent approaches While there is no commonly accepted definition of AI, it can be described as the
to regulation. For example, while the EU has development of computer systems that perform specific functions or tasks that would
created an overarching AI law in the form normally require human intelligence. Narrow AI refers to the performance of tasks that
of the AI Act, the UK is proposing a “pro- are linked to the performance of a particular function. Generative AI is a type of AI
innovation” sector-specific approach (see system that can be used to create new content that reflects the general characteristics
News brief “White paper on regulating AI: is of the training data on which it was trained.
a pro-innovation approach enough?”, www.
practicallaw.com/w-039-2427). Foundation models are large AI models that are trained on multiple data formats
that can be used as the basis for many different types of tasks and operations. Large
New regulatory rules, such as in the EU’s language models (LLMs) are a type of foundation model. LLMs are primarily focused
Digital Markets Act (DMA) and the AI Act, on natural language understanding and generation, and are trained on a large amount
and the UK’s Digital Markets, Competition of text data in order to understand, summarise, generate and predict new content.
and Consumer Bill (DMCC Bill), are expected
to be the main basis for handling AI issues in
competition law, especially in relation to the On 6 September 2023, the Commission DMCC Bill
largest technology companies in the EU and formally brought six companies within the In the UK, the DMCC Bill will introduce the ex
the UK. In a webinar held on 28 September scope of the DMA by designating them as ante regulation of firms designated as having
2023, Mr Alberto Bacchiega, Director gatekeepers: Alphabet, Amazon, Apple, strategic market status (SMS) and enhance the
for Digital Platforms at the Directorate- ByteDance, Meta and Microsoft (https:// CMA’s investigative and enforcement powers
General for Competition, emphasised that ec.europa.eu/commission/presscorner/detail/ in relation to competition and consumer
the European Commission (the Commission) en/ip_23_4328). These companies operate one protection (see feature articles “Digital markets
now has “all the tools it needs” to regulate or more core platform services and several of regulation: comparing the new EU and UK
AI (https://competition-policy.ec.europa.eu/ these services have AI technology embedded. regimes”, www.practicallaw.com/w-040-0659
about/reaching-out/lets-talk-competition/ Concurrently, the Director-General for and “New consumer protection regime: a
dma-compliance_en). In the UK, the Competition opened five market investigations dramatic increase in compliance risk”, www.
Competition and Markets Authority (CMA) has to assess whether other services should be practicallaw.com/w-041-8851). The DMCC Bill
launched a Data, Technology and Analytics scoped in, including a market investigation is designed to help guide the CMA to positive
unit to equip it to better understand, and into a generative AI chatbot that is integrated competitive outcomes, while mitigating
respond to, the competition questions that into Microsoft’s Bing search engine. However, perceived potential harm, in digital markets
are arising from emerging technologies, after an investigation lasting five months, the that are susceptible to AI influence.
including AI. Commission found that the service was not in
scope (https://digital-markets-act.ec.europa. MARKET STUDIES
Digital Markets Act eu/commission-closes-market-investigations-
The application of the DMA to AI is somewhat microsofts-and-apples-services-under-digital- In the UK, the CMA launched an initial review
ambiguous and, in comments made at the markets-act-2024-02-13_en). into AI foundation models on 4 May 2023
6th W@Competition Conference held on 22 (www.gov.uk/cma-cases/ai-foundation-
February 2024, Margrethe Vestager, the AI Act models-initial-review). The review is intended
Executive Vice-President of the Commission, In February 2024, the EU’s AI Act was to explore the opportunities and risks of
said that it will be interesting to see how the approved and, when formally adopted, will these models for competition and consumer
DMA works with AI, given the new market regulate a significant number of AI-related protection, and focuses on fairness, open
dynamics that AI introduces and the colossal issues in the EU (see News brief “Artificial markets and contestability: language that
speed at which it is advancing. intelligence: the dawn of a new legal era”, competition lawyers are familiar with.
www.practicallaw.com/w-031-0858). The AI Simultaneously, as part of its market
Only designated gatekeepers that provide Act will also have an impact on competition investigation into public cloud infrastructure
a core platform service are bound by the enforcement across the EU; for example, services that was launched on 5 October
DMA’s obligations. While a narrow reading the broad procedural powers provided to 2023, the CMA is considering the potential
implies that foundation models, which are a the relevant supervisory agencies, which impact of AI on how competition works in
form of generative AI, may not be in scope, include examining evidence and accessing the cloud services market (www.gov.uk/cma-
senior Commission officials have advocated a data and documents, can be transferred to cases/cloud-services-market-investigation).
more flexible application in order to regulate national competition authorities. In addition,
the surge in AI technologies. To this end, the enhanced transparency of AI systems In the EU, the Commission started 2024
two outcomes may be possible: regulating promoted by the AI Act necessitates the by launching two calls for contributions on
AI services where they are integrated with sharing of important information between competition in virtual worlds and generative
gatekeeper-designated services and/ companies. Competition authorities may AI, along with information requests to several
or amending the text of the DMA so that believe that this could potentially facilitate large digital companies (https://ec.europa.eu/
it expressly encompasses standalone AI collusive behaviour and increase abuse of commission/presscorner/detail/en/IP_24_85).
services. dominance concerns. National competition authorities within the

2 © 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine.
EU are also working in this area; for example, in the AI area (www.gov.uk/government/ Commission is keeping an eye on other
the French competition regulator launched speeches/consumers-competition-and- partnerships in the AI sphere and that it will
a public consultation on 8 February 2024 artificial-intelligence). For example, the CMA’s take account of the impact of AI in how it
looking at the practices of major digital 2023 decision on Microsoft’s acquisition of assesses mergers, as well as how AI may
companies in allegedly consolidating, or Activision Blizzard resulted in Microsoft lead to new kinds of algorithmic collusion
leveraging, their market power in the AI value divesting the online cloud streaming rights (https://europa.eu/newsroom/ecpc-failover/
chain (www.autoritedelaconcurrence.fr/en/ to Activision’s games to a competitor for pdf/speech-24-931_en.pdf).
press-release/generative-artificial-intelligence- the next 15 years (see feature article “UK
autorite-starts-inquiries-ex-officio-and- merger control: what’s in store for 2024?”, US developments
launches). www.practicallaw.com/w-042-2435). In In the US, the Federal Trade Commission
addition, on 8 December 2023, the CMA (FTC) issued orders to five companies
While these kinds of fact-finding exercises are began investigating the partnership between on 25 January 2024 that required them
not new, it is interesting that, at the moment, Microsoft Corporation and OpenAI Inc using to provide information regarding recent
competition authorities are giving their work its merger control powers (www.gov.uk/cma- investments and partnerships involving
a lighter touch feel than the full force of a cases/microsoft-slash-openai-partnership- generative AI companies and major cloud
traditional sector inquiry. There is a slightly merger-inquiry). service providers (www.ftc.gov/news-
different approach in the AI context, with events/news/press-releases/2024/01/ftc-
competition authorities showing significant Under the DMCC Bill, there are plans to launches-inquiry-generative-ai-investments-
flexibility in the form that responses are introduce an “acquirer-focused” merger partnerships). Specifically, the FTC is looking
sent in and asking strikingly open-ended control threshold targeting so-called “killer into whether investments and partnerships
questions. Nevertheless, the prevailing trend acquisitions”, which may be relevant to pursued by dominant companies risk
points clearly in the direction of stronger transactions where large acquirers buy distorting innovation and undermining fair
intervention. Indeed, in the past, these kind start-ups such as generative AI software competition; largely, the same issues that
of information-gathering exercises have been companies. As firms with SMS status will are concerning competition authorities in
a precursor to enforcement action, which be obliged to inform the CMA of relevant the EU and the UK.
there have already been signs of, such as transactions under the DMCC Bill, it is
in the area of merger control (see “Merger unlikely that acquisitions in digital markets The subpoena nature of these requests
control” below). will escape the increasingly bold interventions reflects the seriousness of the FTC’s interest
of the UK regulator. in this area, as well as more recent comments
The same seems to be true in the US. In a by the FTC’s chair, Lina Khan, who said that
speech published on 29 February 2024, EU developments in all of its work, the FTC is “making clear
US Assistant Attorney General Jonathan In the EU, Article 22 of the EU Merger that there is no AI exemption from the laws
Kanter made clear that, based on the US Regulation (139/2004/EC) (EUMR) and on the books. Firms cannot use claims
government’s 9 July 2021 executive order Article 14 of the DMA may be used to justify of innovation as cover for lawbreaking”
on promoting competition in the American merger control in the AI sphere. Article 22 (www.ftc.gov/policy/advocacy-research/
economy, the Department of Justice is of the EUMR allows for the examination of tech-at-ftc/2024/02/few-key-principles-
“laser focused on breaking up existing certain mergers that do not meet the EU excerpt-chair-khans-remarks-january-tech-
monopoly chokepoints across the economy merger control thresholds but affect trade summit-ai).
and preventing new ones before they arise”, between EU member states and threaten
and considering what enforcement policy to significantly affect competition. Article COLLUSION
and other tools can be used to open up 14 of the DMA imposes an obligation on
the market (www.justice.gov/opa/speech/ gatekeepers to inform the Commission about Both UK and EU competition law prohibit
assistant-attorney-general-jonathan-kanter- certain mergers. agreements, arrangements and concerted
delivers-remarks-22nd-international). To this practices that, by object or by effect, prevent,
end, the Department of Justice is currently The Commission has already shown itself restrict or distort competition in the relevant
undertaking a significant amount of work willing to analyse mergers in a forward- market (Chapter I, Competition Act 1998 (1998
with respect to AI and competition, including looking context and has also delved deeply Act); Article 101, Treaty on the Functioning of
numerous active investigations. into online ecosystems, both of which will the European Union (TFEU)). The restriction
be useful groundwork for future mergers of competition can arise from horizontal
MERGER CONTROL concerning AI; such as in Adobe/Figma and restrictions, such as price-fixing, or vertical
Booking Holdings/ETraveli Group (www. restrictions, such as exclusivity agreements.
Recent developments indicate that AI-related practicallaw.com/w-040-7939). Cartels are associations or arrangements
partnerships could fall within the ambit of between two or more competing companies
traditional merger rules. As part of its calls published on 9 January that discuss or exchange information about
2024, the Commission stated that it was their businesses, or make agreements about
UK developments also considering whether the partnership future conduct, in order to limit competition
In a keynote speech published on 1 November between Microsoft Corporation and OpenAI between them and to increase their own
2023, Marcus Bokkerink, chair of the CMA, Inc falls within the scope of its merger control prices or profitability. Cartel members may
said that the CMA will be vigilant in keeping powers. In a speech on 19 February 2024, take part in price-fixing, bid-rigging, output
an eye on potentially problematic mergers Margrethe Vestager emphasised that the quotas or restrictions, or market sharing.

© 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine. 3
 In collusion cases, competition authorities
must show that there was an agreement Algorithms and theories of harm
to collude; that is, a meeting of the minds.
Unilateral, individual decisions by companies On 19 January 2021, the Competition and Markets Authority published a research and
to use algorithmic pricing with no proof of analysis paper to explore how algorithms can reduce competition and harm consumers
agreement is legal. However, interest is (www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-
growing into whether AI could inadvertently and-harm-consumers/). It identified a number of different ways in which consumers
lead firms to become party to anti-competitive can be harmed from algorithmic systems, including:
practices, such as the use of algorithmic
pricing tools that leads to unlawful collusion • The personalisation of prices in a way that is opaque to the consumer.
(see box “Algorithms and theories of harm”).
• The manipulation of choice architecture or user journeys.
In this context, some competition authorities
may think that AI introduces the following • Algorithmic discrimination based on protected characteristics, including
unique challenges: geographic targeting and advert targeting.

• While companies that use AI are often • Unfair ranking and design, including facilitating the preferencing of others for
well aware of the risks of sharing commercial advantage and the use of dark patterns; that is, user interface designs
competitively sensitive information, that trick users into making unintended and potentially harmful decisions.
their AI may inadvertently be trained to
do this. • Exclusionary practices, such as self-preferencing, manipulating ranking
algorithms to exclude competitors and changing an algorithmic system in a
• AI offers much greater speed and ease gateway service that unintentionally harms businesses that rely on it.
of gathering, assessing and acting on
marketplace intelligence. • Collusion by pricing algorithms.

• There is potential for autonomous AI • Ineffective platform oversight, where a lack of transparency can make it difficult
programs to raise prices. to externally evaluate whether an algorithmic system is effective, and therefore
drive improvements.
• AI is better at enforcing agreements
than humans, since it can monitor the
market and “punish” deviations without oversight and override powers. A scenario that employer market power and worker outcomes
human input. may risk infringing UK and EU competition (www.gov.uk/government/publications/
rules is conscious parallelism, also known competition-and-market-power-in-uk-
At the same time, co-operation in AI markets as autonomous tacit collusion. This is where labour-markets). Interestingly, the French
is key for research feasibility and developing companies independently adopt a common competition authority’s current consultation
responsible AI systems and governance. course of conduct without any agreement in the AI sphere includes an explicit question
or communication. It may occur where about contractual clauses that limit the ability
A competition authority’s theory of harm regulators fear that a pricing algorithm has of highly qualified people in the generative
in AI collusion cases may be that an AI tool “learned” to collude without requiring other AI sector to be recruited by competing
replaces the risks of competition with the information sharing or existing co-ordination. companies (www.autoritedelaconcurrence.fr/
certainty of collaboration. In a transparent sites/default/files/2024-02/AI-questions_eng.
marketplace, AI may use other companies’ An additional concern for competition pdf).
prices as a variable in setting its own prices. authorities may occur in contexts where
Where other companies follow the same companies use the same pricing algorithm ABUSE OF DOMINANCE
approach, competition authorities may fear to set prices. This could be through using
that a practice of price signalling, or even a third party’s software or AI service, or by Chapter II of the 1998 Act prohibits companies
price-fixing, could develop. They may be delegating prices to a common intermediary, from engaging in any conduct that amounts
concerned that what may initially appear to creating a “hub-and-spoke” structure (see to the abuse of a dominant position in a
be unilateral behaviour could veer into grey box “Hub-and-spoke structure”). marker insofar as it may affect trade in the UK.
areas and, eventually, illegal co-ordination, Similarly, Article 102 of the TFEU prohibits the
with authorities making clear that price-fixing Due to the challenges involved in intervening abuse by companies of a dominant market
using AI is still price-fixing. Competition on AI markets, at this stage, competition position in the EU or a substantial part of
authorities are likely to keep a close eye on authorities may prefer to consider collusion the EU.
this kind of behaviour. in related markets first, such as labour
markets. The development of AI requires Competition authorities may consider
Another issue yet to be tested is whether significant human resources and competition whether companies that they think might be
an agreement could be inferred from the authorities are already working in this area, dominant in a market could use AI to increase
conduct of an autonomous tool, potentially as demonstrated by the CMA’s 25 January their market power and exclude competitors,
accompanied by a spectrum of human 2024 report on labour market concentration, for example:

4 © 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine.
 harm by competition authorities that are
Hub-and-spoke structure
keen to be seen as active in this nascent
market. Indeed, concepts that were used in
A hub-and-spoke structure is where horizontal collaboration occurs at the level of some of the Commission’s pharmaceutical
suppliers or retailers through the common use of a vertical participant, such as a cases relating to abuses of dominance where
common manufacturer or service provider. The “spokes” are the horizontal participants access to IP was key seem to have some
and the “hub” is the vertical participant. The hub facilitates the co-ordination of application to future potential AI dominance
competition between the spokes without direct contact between the spokes. This investigations, where the protection of IP
means that a horizontal cartel can be created based on indirect communication is proving critical in terms of the data sets
between the horizontal participants, making it difficult to prove that the horizontal used by LLMs.
participants engaged in concerted action. Competition authorities may therefore
consider that price-monitoring software and algorithms could facilitate the exchange However, defining the relevant market in AI
of competitively sensitive information without any communication between sellers. cases may prove a challenge for competition
authorities, and this is key in dominance
ntal collaboratio cases. The Commission’s new market
Horizo n
definition notice, which was published on 8
February 2024, includes updated guidance
Horizontal for defining digital markets (https://
participant ec.europa.eu/commission/presscorner/
detail/en/ip_23_6001). While this does not
apply to companies that are designated
as gatekeepers under the DMA, it is likely
to still assist competition authorities that
Horizontal Horizontal
are eager to step in when faced with digital
participant participant
companies abusing market power in this area.
Given the cumbersome and lengthy nature
Vertical of abuse of dominance investigations, the
participant forthcoming new regulations, such as the
DMA and the DMCC Bill, seem better suited
to interventions in this area.

Horizontal KEY TAKEAWAYS

Horizontal
participant participant Given the speed and vibrancy with which
AI is transforming markets, competition
regulators may believe that they need to act
quickly. All participants on the AI market
should therefore be careful to stay on the right
side of the constantly changing rules. Start-
ups and smaller companies should leverage
regulatory attention to their advantage. In
• Companies with successful AI foundation service more favourably in its general particular:
models may leverage these to impose search results compared to competing
high prices on downstream players for comparison shopping services (Google • Companies are free to gather public
access to their technology. Search (Shopping) AT39740; see News information and use it in making
brief “Google competition charges: competitive decisions. However, given
• A foundation model provider may stop, Commission raises the stakes”, www. increased regulatory scrutiny and the
or reduce, application programming practicallaw.com/2-610-2845). nascent nature of the law in this area,
interface (API) access. API is a software companies should evaluate whether
interface that allows two or more The classic theories of harm, such as self- any algorithms that are being used
computer programs or components to preferencing, tying and bundling, and the in areas where competitors are also
communicate with each other. use of data, may be invoked to govern active can be programmed on non-
behaviour in the AI sphere and competition competitively sensitive data only. In
• A company could employ algorithms in authorities may leverage past experience in addition, the underlying methodologies
a manner that favours its own products addressing allegedly dominant behaviour of algorithmic and AI tools should be
or services, along the lines of the related to algorithms and digital ecosystems. regularly reassessed from a competition
Commission’s decision in relation to Furthermore, since the categories of abuse compliance perspective.
allegations that Google LLC had infringed are not definitive, the technicality and
Article 102 of the TFEU by positioning and complexity of AI models could constitute • Engagement with regulators should
displaying its own comparison shopping fertile grounds for imaginative theories of be planned strategically. At the

© 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine. 5
 Related information
This article is at practicallaw.com/w-042-5695

Other links from uk.practicallaw.com/

Topics
Authorities topic/8-103-1172
Cartels topic/1-103-1161
Horizontal co-operation topic/3-103-1160
Information technology topic/5-103-2074
Market power and dominance topic/2-103-1165
Merger control topic/0-103-1166
Pricing topic/2-103-1170
Restrictive agreements topic/5-103-1178
Vertical agreements topic/9-103-1181

Practice notes
Competition regime: Article 101 0-107-3707
Competition regime: Article 102 8-107-3708
Competition regime: Chapter I prohibition 6-107-3691
Competition regime: Chapter II prohibition 4-107-3692
Demystifying AI (UK) w-008-5369
Digital Markets Act (EU): overview w-038-9380
Information exchange and EU competition law 3-525-5773
Information exchange and UK competition law 7-538-7446
Legal aspects of AI w-018-2338
Transactions and practices: EU mergers and acquisitions 4-107-3705
Transactions and practices: UK mergers and acquisitions 8-107-3690

Previous articles
AI governance, risk and compliance: shaping an unknown future (2023) w-040-0428
Cloud computing: trying to find the silver lining (2023) w-039-4053
Robo-advice: risks and regulation (2023) w-037-8150
AI and automated decision making: to regulate or deregulate? (2022) w-033-8467
AI and data protection: balancing tensions (2019) w-020-9713
Big data and competition law: with great opportunities come great risks (2019) w-020-7091
Challenges in the consumer sector: transformative technology (2019) w-020-3706
Algorithms, apps and AI: the next frontier in discrimination law (2018) w-013-8054

For subscription enquiries to Practical Law web materials please call +44 0345 600 9355

moment, competition authorities are it is clearer that the AI tool is learning with robots and codes, competition
keen to speak with large and smaller to focus on particular parameters. authorities are engaging with
participants on the technology markets Strong arguments and advocacy will businesses and individuals whose rights
as they seek to better understand how be essential. However, companies are protected by law. Safeguards and
AI works and consider its potential should plan carefully when and how to due process must be preserved, however
impact on a market and the relevant discuss these issues with competition “artificial” the intelligence might seem.
participants. For example, it may be both authorities, if at all, especially during
harder and easier to infer agreement these early stages of development. Aymeric de Moncuit and Daniel Vowden are
based on an AI learning tool compared partners, Kathryn Lloyd is a senior associate,
to a tool based solely on algorithms: • Companies must not let the protection Nadia Dhorat is a lawyer, and Ora Nwabueze
harder as the AI tool is unpredictable of their fundamental rights slip. While and Sarah Wilks are professional support
and not pre-programmed, and easier as AI may give the impression of dealing lawyers, at Mayer Brown.

6 © 2024 Thomson Reuters (Professional) UK Limited. This article first appeared in the April 2024 issue of PLC Magazine.