AI: A chance for early regulation to avoid the mistakes of the past?

These are reflections on a discussion conducted under Chatham House rules; they are not minutes of that discussion. The views expressed are the views of the authors only and do not necessarily represent the views of the UK Competition Policy Forum participants or Compass Lexecon, its management, its subsidiaries, its affiliates, its employees or its clients.

Reflecting on the desire to regulate competition in AI markets

• Huge potential, many concerns. Artificial intelligence is already transforming large parts of the economy. It promises substantial benefits, including a potential response to anaemic productivity growth since the financial crisis, and the ability to reshape sectors from healthcare to education. At the same time, it brings risks: disruption to labour markets, pressure on the climate and energy systems, tensions with privacy and copyright regimes, and possibly – at least in some scenarios – an existential threat and the collapse of liberal democracy.

• Concern that elements of the AI stack are concentrated or may tip. Focussing on competition concerns, currently, the supply of hardware required for AI is concentrated (GPUs, networking, and memory chips), and markets in AI-related software are not. However, there are concerns that, if these downstream markets tip, they will become difficult to contest, and market power in those segments could then be extended through ecosystems of connected services.

• The chance for a digital “do-over”? Part of the regulatory impulse reflects a desire to avoid the mistakes of the past. Many feel that competition authorities’ response to the first generation of digital markets was slow and ineffective – whether through ex post enforcement, or through ex ante regimes that arrived only after markets and consumer preferences had ossified. AI therefore appears to offer a ‘second go’.

• Yet, regulators cannot simply apply old lessons to new questions. AI-related markets do not offer an exact parallel to the past. Some potential concerns look familiar, such as refusing access to established platforms. Others, such as algorithmic collusion, are genuinely new. Perhaps most difficult cases are the ones that appear familiar but may not be analogous at all. Integrating generative AI into search results, for example, is not obviously a rerun of Google Shopping, even it if superficially presents that way.

Reflecting on concerns about market tipping, before concentration emerges

• Market tipping is not always a problem, nor preventable. Currently, investment in AI is substantial and the market appears competitive, with $1 trillion of global capital expenditure expected in 2026. That may indicate a “race to market”, fuelled by the prospect of supra-competitive profits after it tips. However, some markets tip for sound economic reasons, particularly where network effects and economies of scale are strong. It is not obvious which AI-related markets exhibit these characteristics. For instance, the arms race between LLM chatbots may well tip, but currently, it is unclear whether these products are as prone to lock-in and excess momentum as operating systems or app stores were. Regardless, where the economic features that lead to tipping are present, maintaining (highly) fragmented markets may not be possible or desirable.

• Concern about to whom a market tips, and why. This concern is not that tipping will occur; rather, it is that a market may tip towards a company that does not win by virtue of its competitive merits. For example, it may distort the competitive process by creating artificial reasons that influence when, why, and to whom the market tips, such as tying AI services to other platforms or complementary services. This is exacerbated by concern that the “usual suspects” from the digital era may leverage their access to data, compute, talent, capital, or existing services. Once a market tips for the ‘wrong’ reason, correcting the outcome ex post would prove extremely difficult to say the least. Intervening ex ante may protect competition on the merits; yet, it risks distorting the competitive race itself, if not done carefully.

• Concern about adjacent markets being dragged down with a market prone to tipping. A related concern is contagion, or a domino effect – digital ecosystems can allow concentration in one market to spill over into adjacent markets. The problem is that a core service may be naturally prone to tip. But, in doing so, the concentration in that market spreads to adjacent services where the benefits of consolidation are, at best, less clear. Early intervention could prevent contagion, but it may be challenging to do well. Competition policy has not reached a settled view on the costs and benefits of ecosystem expansion. Moreover, the critical links between markets are often easier to identify with hindsight than ex ante. Further, the most likely remedy – mandated interoperability – should be feasible after markets have tipped.

• Early intervention carries risks as well as benefits. The first wave of digital markets may have suffered from early under-enforcement. But that teaches us nothing about the risk of early over-enforcement. There are reasons for caution. While the supply of hardware is concentrated, there is little evidence that existing competition tools would be inadequate, if required. The concern, so far, is how that concentration affects supply shortages and resilience. Downstream, markets for software applications remain nascent, to the point that defining the relevant markets that services and companies compete in is challenging: they may be too narrow, or too wide, or miss the target altogether.

Reflecting on the risks and benefits of regulation, after concentration emerges

• Past regulatory failures warrant humility, not nihilism or hubris. By general consensus Neither ex post enforcement nor ex ante regulation in the first wave of digital markets has been satisfactory. However, that does not mean regulation is futile. If natural monopoly positions emerge (even temporary ones), abandoning regulatory oversight altogether is not a credible position. At the same time, simply applying old approaches earlier will not guarantee better results, whether through DMA-style rules or a new round of ex post investigations. Any improvements will need to be grounded in principle, evidence, behaviour, and market realities.

• The merit of a “study first, act second” approach. One lesson from the past is that competition authorities need to understand emerging markets early. Both action and inaction can harm competition if based on incomplete or ill-informed understanding. Encouragingly, authorities appear to have adopted this approach, through market studies, consultations, and early monitoring of AI developments.

• Conduct in digital markets rarely lends itself to simple rules. Some practices, such as restrictive distribution agreements, often appear problematic in hindsight. Yet, in specific circumstances, they can also encourage investment, market penetration, or improve consumer outcomes. Similarly, the effects of self-preferencing, tying, or bundling can vary across markets and technologies. Therefore, whether regulation adopts an ex post or ex ante perspective, the imposition of blanket rules will tend to be, at best, a blunt tool.

• Ex post intervention is not inevitably narrow and late, even in rapidly changing markets. Early intervention does not necessitate a wide-ranging ex ante regime. Nor does the need to respond to evolving markets. Ex post investigation can be timely and well targeted – particularly if supplemented, for example, with interim measures – offering some of the benefit of an ex ante regime in a more targeted and proportionate form.

• Avoid mobilising for the Last War. Fixating on the past is no guarantee that one is prepared for the future. However, nor should history be ignored. The successes and failures of earlier digital regulation provide useful lessons, provided they are applied with care, and adapted to the specific features of AI-related markets.