AI: A Brief Overview
Artificial intelligence (AI) has rapidly evolved in recent years, permeating various aspects of our daily lives. From virtual assistants (such as Siri, Alexa, or Google Assistant) that schedule our appointments to sophisticated algorithms that power recommendation systems, AI is transforming how we interact with technology and the world around us.
So, what is AI? AI involves the development of computer systems capable of performing tasks typically requiring human intelligence. These tasks include learning, reasoning, problem-solving, and perception. AI’s applications are vast, ranging from healthcare, where it aids in diagnosing diseases and developing new treatments, to finance, where it helps detect fraud and optimize investment strategies.
AI has three core components:
1. Data
High-quality and diverse data is the foundation upon which AI systems are built. It is akin to the raw material that AI models use to learn and improve. The more data an AI system has access to, the better it can understand patterns, make accurate predictions, and perform complex tasks. However, the quality of data is equally important. Biased or incomplete data can lead to biased or inaccurate results. In essence, the quality of an AI’s output is directly tied to the quality of its input.
2. Computational Resources
Powerful hardware, such as GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units)[i], are essential for training and running complex AI models. These specialized hardware accelerators can handle the intensive calculations required for AI tasks much faster than traditional CPUs (Central Processing Units). However, access to such hardware can be limited and its price too high, especially for smaller AI developers. The more scalable and cost-efficient alternative for them is access to cloud computing services, which provide access to powerful hardware infrastructure on demand. However, even with cloud computing, challenges such as latency, data privacy concerns, and potential vendor lock-in can hinder AI development, especially for smaller teams or projects with limited resources.
3. Engineering Talent
Skilled engineers and data scientists are the architects and builders of AI systems. Their expertise is crucial for designing effective algorithms, developing robust models, and ensuring the ethical and responsible deployment of AI. However, there is a global shortage of AI talent, which can hinder innovation and create competition between countries and companies for skilled workers.
As AI continues to advance, its potential to revolutionize industries and improve our lives is immense. But, it undeniably presents significant dangers. One major concern is the potential for AI to be used for malicious purposes, such as developing autonomous weapons or spreading misinformation. Additionally, as AI becomes more advanced, there is a risk of job displacement as machines become capable of performing tasks traditionally done by humans. Furthermore, the concentration of AI development and control in the hands of a few powerful players raises concerns about bias, discrimination, and privacy violations. Addressing these challenges will require careful consideration of ethical implications, robust regulation, and international cooperation.
Competition Law: Core Principles
Competition rules seek to foster a fair and equitable marketplace where all businesses have an equal opportunity to compete, and that consumers benefit from fair prices and a wide range of choices.
To reach this goal, competition rules prohibit certain practices, such as:
a. Entering into anti-competitive agreements:
These include strictly prohibited clauses or activities such as cartels (agreements between competitors to e.g. fix prices, restrict output, or share markets) or resale price maintenance, and other restrictive covenants (such as exclusive distribution agreements or non-compete arrangements, which, depending on the circumstances, may be lawful).
b. Abuse of dominance:
Players with significant market power need to play by stricter competition rules, which means they cannot engage in practices that could be allowed to smaller companies. For these biggies, sometimes perfectly legit (and legal) strategies and activities, such as pricing below cost, tying and bundling, or refusal to deal with certain customers, are a no-go.
c. Concentrations that significantly reduce competition:
Some M&A deals may lead to creating or strengthening a dominant player, who could impose higher prices, lead to less consumer choice, or other negative consequences. Through merger control procedures, competition authorities make sure such deals do not go through.
The Intersection of AI and Competition Law
The interplay between AI and intellectual property (IP) and AI and data protection is at the forefront of legal discussions, as AI directly and immediately affects these two sets of rules. On one hand, AI systems often rely on vast datasets to learn and improve, raising questions about data ownership and possible violations of data subject’s rights. On another hand, AI can generate creative content, blurring the lines between human and machine authorship, as well as use copyrighted materials available online for its training.
However, the impact of AI on competition is often more nuanced and requires careful analysis of specific market dynamics and regulatory frameworks. AI can both enhance and potentially distort competition, depending on its application and the specific circumstances.
AI can undoubtedly improve the competitive landscape by:
a. Increasing efficiency:
AI can improve efficiency and reduce costs for businesses, leading to lower prices for consumers. For example, AI-powered supply chain management systems can optimize logistics and reduce costs.
b. Driving innovation:
AI can incite innovation and the development of new products and services, benefiting consumers and the economy. For example, AI-powered medical diagnostics tools can improve healthcare outcomes.
c. Improving market transparency:
AI can increase market transparency by providing consumers with more information about prices, product features, and quality. For example, AI-powered price comparison tools can help consumers find the best deals.
This article further explores how AI can facilitate a wide range of anti-competitive practices, including restrictive agreements, abuse of dominance, and cleverly designed “partnerships”, which in reality could be concentrations in disguise.
AI and Restrictive Agreements
Can companies enter into restrictive agreements through the use of AI and thus breach competition rules? Apparently, yes.
Specifically, the use of algorithms by market participants can result in competition law infringements. Algorithms, the backbone of AI, are essentially sets of instructions that computers follow to perform specific tasks. In the AI context, algorithms enable machines to learn from data, recognize patterns, and make decisions.
Some competition authorities, such as the European Commission (“EC”)[ii], the UK Competition and Markets Authority (“CMA”)[iii], and, more recently, the French Autorité de la Concurrence (“Autorité”)[iv] have assessed how algorithms can negatively affect competition by, broadly speaking, facilitating the conclusion of restrictive agreements.
A particularly intriguing and complex phenomenon, “algorithmic collusion” has emerged as a topic of interest for competition authorities. Algorithmic collusion is a relatively new concept in competition law that refers to the potential for algorithms used by competing firms to collude and manipulate markets. This can occur when algorithms are programmed to react to each other’s behavior in a way that mimics human collusion, such as coordinating prices or limiting output.
Here are three main scenarios where algorithmic collusion can occur[v]:
- Algorithm Supporting the Existing Collusion: When companies have already agreed to collude, algorithms can be used to monitor, enforce, or conceal this agreement. Here, the algorithm does not cause anticompetitive behavior, but rather serves as a tool to enforce a “traditional”, preexisting agreement. An example of this is the Eturas case[vi], where the Court of Justice of the European Union (“CJEU”) concluded that the discount caps applied by Eturas and around 30 other Lithuanian travel agents through an online booking platform, infringe competition rules and amount to collusion. CJEU confirmed that the terms of use of online platforms can infringe competition rules, so platform administrators need to carefully review these to avoid the risks of facilitating collusive practices between the platform users. Another interesting case is the CMA’s Trod/GB eye decision.[vii] Here, these two entities configured an automated repricing software, so as not to undercut each other’s prices of online sports posters on Amazon UK website, which effectively means they entered into a cartel, which is a hard core competition infringement.
- Shared Third-Party Algorithms Result in Tacit Collusion: Competitors’ use of the same AI algorithm developed by a third party can lead to implementing similar pricing strategies and competitive conditions between these companies, even without direct communication between them. Still, whether this (in competition enforcement practice still hypothetical) scenario would lead to an antitrust infringement is questionable. The outcome would mainly depend on whether competing companies knew (or could have reasonably known) that they were using the same AI algorithm, and, if they did, that the use of that algorithm could result in anticompetitive conduct.
Competition authorities may specifically be concerned when competitors use the same pricing algorithm, either through a shared third-party software or AI service or by coordinating pricing through a common intermediary. This so-called “hub-and-spoke” arrangement, where sensitive information is shared by competitors through the same intermediary, can potentially lead to tacit collusion and thus competition infringement.
- “Rogue” Self-Learning Algorithms: Imagine two competitors using the same sophisticated AI algorithm for market intelligence that learns and evolves, and not reasonably knowing that they are using the same product. This algorithm could over time teach itself to – without any intervention from humans – align these competitors’ pricing strategies to ensure the best market outcome and thus infringe competition rules. While this scenario remains theoretical for now, rapid advancement could change that in the future. The question will then remain if current competition rules are fit for purpose – could our two competitors could be guilty of tacit collusion, or the clever algorithm is doing nothing more than intelligently adapt to competitor’s conduct in the market, which is perfectly legal? In any case, any competition authority trying to prove our two competitors’ liability for competition infringement under the currently applicable set of competition rules, solely based on the AI “gone rogue”, would really be tough.
Algorithmic collusion is therefore a complex issue with potentially significant implications for competition law enforcement. While it can be difficult to detect and prove, competition authorities are increasingly aware of the potential for algorithms to facilitate anti-competitive behavior. As AI continues to evolve, it is likely that competition law will need to adapt to address the challenges posed by algorithmic collusion.
In the meantime, watchdogs can freely focus on the more traditional markets, where detecting and proving infringements will be easier than in the AI markets per se. We have seen that one of the cornerstones of AI is a highly skilled engineering workforce. Some practitioners therefore predict the regulators could initially focus their enforcement on the labour markets, where restrictive practices will certainly be much easier to prove.[viii] Notably, earlier this year, the EC issued a Competition Policy Brief focused on labour markets.[ix] In this document, the EC reiterated that wage-fixing agreements between companies and no-poach clauses represent severe (by object) restrictions of competition. This applies to all markets, including the labour markets specific to AI-skilled workforce. Further, Autorité, one of the most active regulators overseeing conditions of competition in AI-related markets, specifically notes that no-poach agreements in the AI industry require careful monitoring. One instance that caught their eye is Microsoft’s recent hiring spree of employees from other AI companies. Specifically, the Redmond-based giant took over many of Inflection’s key employees including its two co-founders. It also briefly hired Sam Altman before he came back to OpenAI.[x] This conduct admittedly can be assessed (and was assessed) also through merger control rules (please see Section III below), but Autorité goes a step further and considers that the takeover of highly skilled individuals can also be seen as “as an attempt to exclude competitors from the sector.”[xi]
AI and abuse of dominance
While entering into restrictive agreements through AI activities is a developing area, competition infringement through abuse of dominance is a more immediate issue.
As we have indicated, AI needs vast amounts of data to train and improve, making this input critical for this industry. Those players that harvest the most data have the best AI models and consequently the highest market share. Further, AI needs super-strong GPUs and TPUs to function, and these are nowadays in short supply. Consequently, several companies producing them can set the terms of their sale to AI developers. Realistically, only a few of them (Big Tech?) can afford these units.
The rest will have to rely on cloud infrastructure, which provides access to powerful computing resources and simplifies the development process, but the question remains at what cost and under what terms. From this, we can see that big suppliers of either GPUs/TPUs or cloud infrastructure effectively set the rules of the game.
Either on the data or computer resources supply level, there is a realistic danger that the dominant few will be able to effectively create high barriers to entry into downstream levels, making it very difficult for smaller AI developers to effectively compete. In other words, where there is dominance, there is a realistic potential for its abuse.
Former EU Competition Commissioner and current EU Executive Vice President of the European Commission for A Europe Fit for the Digital Age, Margrethe Vestager, known for her legal battles (with somewhat mixed results) against BigTech, nicely summed up the potential competition-related issues in the AI sector:
“A major risk we see is big tech players leveraging their market power across different markets within their ecosystem. Concentration is especially high at the top of the value chain, where large foundation models are trained to be used in various applications. These models need vast amounts of data, computing power, cloud infrastructure, and talent, which only a few players have.
This could lead to practices like tying and bundling by dominant firms, blocking AI competitors from accessing essential resources, and preventing customers from switching. We need to keep a close eye on this. […].
Another risk we see is that big tech companies could make it difficult for smaller foundation model developers to reach end users. Whether alone or in alliances with preferred partners. So, we are closely monitoring distribution channels to make sure businesses and consumers still have a wide range of choices among foundation models.”[xii]
So far, regulators have not issued any decisions on abuse of dominance in any of the AI-specific markets, but they are certainly expected soon. For example, Autorité carried out an unannounced inspection (dawn raid) at the premises of NVIDIA, main supplier of GPUs needed for training the generative AI models.[xiii] In June 2024, Autorité openly stated that NVIDIA appears to have a dominant position on the market for GPUs needed to train the foundation AI models, while also expressing concern about the industry’s reliance on its CUDA chip programming software (apparently, the only compatible with these GPUs).[xiv] However, based on the publicly available information, Autorité has not yet opened a formal investigation against NVIDIA.
Moreover, we have seen what has been and can be done through algorithms in the restrictive agreements’ domain. In the abuse of dominance context, the possibility of competition infringement through the use of algorithms is even higher.
Below are a few interesting cases:
- Google Shopping Case: In this case (which Google finally lost on 10 September 2024), the EC fined Google EUR 2.42 billion for abusing its dominant position by using an algorithm favouring its own shopping service over competitors’.[xv]
- Amazon: Amazon operates both an online marketplace and sells products as a retailer on that same marketplace. EC effectively forced it to make commitments to address concerns about favouring its products and those of selected sellers in the “Buy Box”, a prominent product placement feature on its marketplace. The “Buy Box” is crucial for Amazon marketplace sellers, as most consumers only view and purchase products displayed there. The EC believes that Amazon abused its dominant position by setting unequal conditions for selecting the “Buy Box” winner, and thus artificially increased traffic to and purchases of products sold either by Amazon’s retail arm or preferred sellers.[xvi]
For these and similar practices, practitioners coined the term “self-preferencing”, which, broadly speaking, refers to a “platform favouring its own products and services over those of third parties that operate on the platform”.[xvii]
Another abusive practice potentially resulting from the use of algorithms by dominant firms is predatory pricing. AI-powered pricing algorithms could enable major tech companies to engage in predatory pricing, where they set prices below cost to eliminate competitors’ offers and thus lock in customers or groups of customers to themselves and away from smaller competitors. This can be particularly harmful as AI allows for virtually instant and targeted price adjustments. However, determining if a dominant company’s pricing strategy is predatory can be challenging, and regulators must balance the need to prevent anti-competitive behavior with promoting innovation and transparency in the AI sector.
Merger Control in AI markets
As AI technologies become more sophisticated and integrated into core business operations of many companies (including BigTech), the need for effective merger control in AI markets becomes paramount. Moreover, transactions involving companies active in various AI markets require scrutiny to ensure that the combination of resources and expertise does not lead to undue market dominance or the creation of insurmountable barriers to entry for competitors. Hence, it is no wonder that some of the key competition authorities have turned their eyes on the “partnerships” in the AI sector.
We have indicated that Microsoft’s takeover of most of Inflection’s employees was reviewed as a possible concentration. This deal also included a non-exclusive license for Microsoft to use some of Inflection’s IP. CMA assessed this transaction and ultimately concluded that the entire arrangement amounted to a notifiable concentration. Following a Phase I review, the UK regulator concluded that the takeover was in line with UK competition rules.[xviii] While the UK indeed has a somewhat distinct notion of concentration (“merger situation”) compared to the EU, this deal still met the applicable local criteria.
CMA also looked at the partnership between Microsoft and Mistral AI. This deal included making Mistral’s AI models available through the Microsoft Azure platform, investment by Microsoft of EUR 15 million in return for equity, and the possibility of future collaboration in R&D activities. Here, the CMA concluded that the partnership did not meet the criteria for a merger situation.[xix] Most recently, its review of the Amazon/Anthropic deal had the same outcome.[xx]
Probably the most widely known example of the merger control scrutiny of an AI-related deal was the review of Microsoft’s USD 13 billion investment into ChatGPT’s developer, OpenAI. As part of this deal, Microsoft got 49% of the shares in OpenAI and integrated OpenAI’s models, such as GPT-3.5, into its Azure cloud platform, making them accessible to developers and businesses through Azure OpenAI Service. In return for the investment, Microsoft took a non-voting observer seat on OpenAI’s board. All this led to increased scrutiny of the deal by the competition authorities in the EU, the US, and the UK. The EC extensively reviewed whether the investment amounted to acquisition by Microsoft of control over OpenAI, but ultimately concluded that was not the case, while still assessing if the exclusivity deal between the two warrants further antitrust scrutiny.[xxi] The US Federal Trade Commission and CMA are also still investigating this deal, and the outcomes are eagerly awaited. Interestingly, Microsoft ultimately gave up the observer seat in OpenAI to ease the regulatory pressure.[xxii]
Most recently, the EC agreed to, under the referral mechanism for concentrations that do not meet the EU turnover thresholds, the Italian Competition Authority’s request to review another deal that involves AI markets – namely the proposed acquisition of Israel-based Run:ai Labs Ltd by NVIDIA. Run:ai is active in the supply of GPU orchestration software which allows corporate customers to manage and optimize their AI compute infrastructure.[xxiii] We are looking forward to the results of this procedure as well.
All these cases highlight the regulators’ unwavering commitment to safeguarding fair competition and preventing excessive market concentration in the AI sector. These deals, often disguised as seemingly innocuous “partnerships” rather than notifiable concentrations, can have far-reaching implications for competition. As AI technology continues to advance, we anticipate even more stringent scrutiny of mergers and acquisitions, as well as partnerships, in this field.
Instead of Conclusion: A few practical tips
While we are eagerly monitoring the developments in the practice of the main competition authorities in the AI markets, here are a few practical tips in case your business uses, develops, or puts AI on the market. These suggestions can immediately help you ensure compliance with competition rules:
- Develop an antitrust compliance program specifically for AI-related activities.
- Secure that agreements to supply key inputs for your AI, whether that be a rich data set or cloud computing services, are non-discriminatory and competition law compliant – particularly where those agreements are exclusive in nature, and entered into with a dominant supplier.
- Assess regularly the potential competition law risks associated with your AI activities, keeping in mind the rapid development of this field and the increased regulatory scrutiny.
- Seek specialist legal advice to address any specific legal questions or concerns related to application of competition law to AI.
[i] According to Google, „ […] TPUs are custom-designed AI accelerators, which are optimized for training and inference of large AI models. They are ideal for a variety of use cases, such as chatbots, code generation, media content generation, synthetic speech, vision services, recommendation engines, personalization models, among others.“, https://cloud.google.com/tpu?hl=en
[ii] Algorithmic competition – Note by the European Union, 14 June 2023, a written contribution from the European Union submitted for Item 5 of the 140th OECD Competition Committee meeting on 14-16 June 2023, available at: https://one.oecd.org/document/DAF/COMP/WD(2023)17/en/pdf (accessed 21 October 2024).
[iii] CMA, Algorithms: How they can reduce competition and harm consumers (2021), available at: https://assets.publishing.service.gov.uk/media/60085ff4d3bf7f2aa8d9704c/Algorithms_++.pdf , p. 29-34.
[iv] Autorité de la Concurrence, Opinion 24-A-05 of 28 June 2024 on the competitive functioning of the generative artificial intelligence sector, available at: https://www.autoritedelaconcurrence.fr/sites/default/files/commitments/2024-09/24a05_eng.pdf.
[v] Algorithmic competition – Note by the European Union, para. 16
[vi] Case C-74/14 Eturas, ECLI:EU:C:2016:42.
[vii] CMA, Trod/GB eye, 12 August 2016, available at: https://www.gov.uk/cma-cases/online-sales-of-discretionary-consumer-products
[viii] Mayer Brown, AI Challenges in Competition Law: How Are Regulators Responding?, 2024, available at: https://www.mayerbrown.com/-/media/files/perspectives-events/publications/2024/04/ai-challenges-in-competition-law_mar24.pdf%3Frev=55168f8e10a64e458c3fc1ac7af179df
[ix] EC, Competition Policy Brief: Antitrust in Labour Markets, 2 May 2024, available at: https://competition-policy.ec.europa.eu/document/download/adb27d8b-3dd8-4202-958d-198cf0740ce3_en?filename=kdak24002enn_competition_policy_brief_antitrust-in-labour-markets.pdf .
[x] Autorité, Opinion 24-A-05 of 28 June 2024 on the competitive functioning of the generative artificial intelligence sector, p. 7 and p.64.
[xi] Ibid.
[xii] Speech by EVP Margrethe Vestager at the EC workshop on “Competition in Virtual Worlds and Generative AI“, 28 June 2024, transcript available at: https://ec.europa.eu/commission/presscorner/detail/en/speech_24_3550
[xiii] https://www.eenewseurope.com/en/NVIDIA-confirms-it-is-under-scrutiny-in-eu-us-and-china/
[xiv] Autorité, Opinion 24-A-05 of 28 June 2024 on the competitive functioning of the generative artificial intelligence sector, p. 8.
[xv] EC, Case AT.39740 – Google Search (Shopping), 27 June 2017
[xvi] EC, case AT.40703 – Amazon Buy Box, 20 December 2022, summary available at: https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:52023AT40462(01).
[xvii] Compass Lexecon, What Constitutes Self-Preferencing and its Proliferation in
Digital Markets, 8 December 2023, available at Global Competition Review’s website: https://globalcompetitionreview.com/guide/digital-markets-guide/third-edition/article/what-constitutes-self-preferencing-and-its-proliferation-in-digital-markets#footnote-054 .
[xviii] CMA, Microsoft Corporation’s hiring of certain former employees of Inflection and its entry into associated arrangements with Inflection, ME 7103/24, 4 September 2024, summary available at: https://assets.publishing.service.gov.uk/media/66d82eaf7a73423428aa2efe/Summary_of_phase_1_decision.pdf
[xix] CMA, Microsoft Corporation’s partnership with Mistral AI, Decision on relevant merger situation, ME/7102/24, 17 May 2024, available at: https://assets.publishing.service.gov.uk/media/664c6cfd993111924d9d389f/Full_text_decision.pdf
[xx] CMA, Amazon.com Inc.’s partnership with Anthropic PBC: Decision on relevant merger situation, 27 September 2024, accessible at: https://assets.publishing.service.gov.uk/media/6710ba44e84ae1fd8592f52c/Full_text_decision.pdf
[xxi] https://ec.europa.eu/commission/presscorner/detail/en/speech_24_3550
[xxii] https://www.euronews.com/business/2024/07/11/microsoft-drops-openai-board-observer-seat-amid-regulator-scrutiny
[xxiii] EC, Daily news 31/10/2024, accessible at: https://ec.europa.eu/commission/presscorner/detail/de/mex_24_5623
