Regulators keep warning that AI markets are about to be captured by Big Tech. The awkward fact is that AI markets keep refusing to cooperate. Several years into the generative-AI boom, the sector still looks less like a coronation than a street fight: OpenAI, Google, Meta, Amazon, Anthropic, Perplexity, Mistral, xAI, and others are battling across models, applications, distribution, infrastructure, and enterprise services.
As I argued recently, “durable market power and demonstrable competitive harm remain elusive.” The market has not simply “tipped” toward Google, Meta, Amazon, Apple, or Microsoft. If anything, the most visible consumer-AI leader is OpenAI, which is reportedly preparing for an initial public offering.
Anthropic appears to have an edge in enterprise AI services, while Google and Microsoft benefit from distribution and infrastructure tied to their legacy businesses. Those advantages make them serious contenders, but hardly inevitable winners. On closer inspection, the AI ecosystem looks less like a market already captured by “Big Tech” than one defined by entry, rivalry, experimentation, and rapid technological change.
Despite those developments, a more pessimistic narrative continues to push for aggressive antitrust enforcement—and even direct state intervention—to “shape” AI markets. Advocates warn that AI could otherwise “supercharge ‘digital feudalism.’” That framing now runs through much of contemporary AI policy. The goal is no longer merely to police markets after anticompetitive conduct occurs, but to intervene ex ante—before the fact—to prevent a feared future of “private control” and “rent extraction.”
In his 2024 paper “The Case Against Preemptive Antitrust in the Generative Artificial Intelligence Ecosystem,” Jonathan Barnett argues that this turn toward enforcement in AI markets reflects a broader “preemptive approach” to antitrust. Under that approach, regulators presume certain practices by large technology firms are anticompetitive and place the burden on those firms to prove otherwise. Barnett’s warning is straightforward: in emerging markets, premature intervention risks suppressing “innocuous or efficient business practices” before regulators have enough evidence to assess their competitive effects.
That concern is especially acute in early-stage markets, where uncertainty is high and business practices that initially appear exclusionary may turn out to be competitively neutral—or affirmatively procompetitive. In AI markets, that includes product integration, minority investments, partnerships, licensing arrangements, and cloud-computing agreements that help firms assemble the complementary assets needed to compete.
Together with Dirk Auer, I have similarly warned that:
… overenforcement in the field of generative AI could engender the very harms that policymakers currently seek to avert. Indeed, preventing so-called “big tech” firms from competing in these markets (for example, by threatening competition intervention as soon as they build strategic relationships with AI startups) may thwart an important source of competition needed to keep today’s leading generative-AI firms in check.
This post examines three recent examples—from the European Union, Italy, and Brazil—that illustrate the common logic of this preemptive approach and suggest the warning is becoming increasingly urgent.
Emergency Remedies for Hypothetical Harms
One of the clearest examples is the European and Italian scrutiny of Meta’s integration of AI into WhatsApp.
The Italian Competition Authority (ICA) first opened an investigation in July 2025 into Meta’s decision to integrate Meta AI into WhatsApp. The ICA focused on the fact that Meta AI appeared prominently in the WhatsApp interface and was integrated into the search bar, while users allegedly had limited ability to remove or hide the feature.
The ICA later broadened the investigation to examine WhatsApp’s Business Solution Terms, which excluded Meta AI’s competitors from the platform beginning Oct. 15, 2025. Specifically, the terms barred rival AI-chatbot services from operating on WhatsApp. The ICA also opened proceedings to consider interim measures—that is, temporary remedies imposed before the underlying investigation has concluded.
The European Commission opened a parallel investigation, but moved faster and escalated further. After Meta updated its WhatsApp Business Solution Terms in October 2025 to exclude third-party general-purpose AI assistants effective Jan. 15, the Commission launched a formal Article 102 investigation into Meta’s restrictions on third-party AI assistants accessing WhatsApp. Article 102 of the Treaty on the Functioning of the European Union (TFEU) prohibits dominant firms from abusing their market position.
On Feb. 8, the Commission issued a Statement of Objections, a formal document setting out its preliminary view that Meta’s exclusion violated European Union antitrust rules. The Commission framed the conduct as an abuse of dominance in the “consumer communication applications market.”
Meta responded by restoring access for third-party AI assistants, albeit subject to a fee. The Commission remained unsatisfied. On April 14, it issued a Supplementary Statement of Objections arguing that the fee-based approach was “in effect equivalent to the previous access ban.” The Commission also notified Meta of its intention to impose interim measures requiring the company to restore third-party access under the pre-October 2025 terms, pending the outcome of the broader investigation.
Why call this preemptive? One could reasonably object that both the ICA and the Commission acted only after Meta restricted access to WhatsApp. The deeper problem is that both authorities appear to rely primarily on a theory of harm: that vertical integration—combining a platform with complementary services—creates incentives to disadvantage rivals.
But abuse-of-dominance cases require more than a plausible theory. They require evidence of dominance and evidence of harm to competition.
Even at this preliminary stage, the market definition looks implausibly narrow. As Dirk Auer has explained, consumers reach and use communications apps and AI assistants through a wide range of channels, often switching fluidly across platforms. Once those competitive dynamics are taken seriously, the relevant market likely becomes much broader, making it harder to sustain the claim that WhatsApp holds a dominant position.
The evidence of competitive harm also remains speculative. As Giuseppe Colangelo has warned:
It is uncertain that integrating Meta AI into WhatsApp would materially harm competition in AI assistants, especially given the success of rivals such as ChatGPT. ChatGPT achieved rapid adoption through cross-platform integration and partnerships enabling users to shop, book services, and perform other tasks within a single interface. By contrast, Meta AI’s market share remained minimal—about 0.2% during April 2024–March 202567 and below 1% in January 2026—and developer adoption was low.
The Commission’s choice of remedy makes the case look even more premature. Interim measures are an extraordinary tool in EU competition law. They are rarely used and traditionally require evidence of serious and irreparable harm before the main investigation concludes.
Yet the Commission seeks emergency intervention in a market where rival AI assistants—including ChatGPT, Claude, Gemini, Copilot, Perplexity, and Grok—have expanded rapidly through their own apps, browser integrations, operating-system partnerships, and enterprise channels. The Commission has not shown that exclusion from WhatsApp has actually foreclosed rivals from entering or expanding in the market.
Imposing emergency remedies based on asserted rather than demonstrated harms risks freezing the competitive process itself. It could even sideline Meta by stripping away one of the few advantages it plausibly holds in AI competition: distribution.
Brazil’s ‘Just in Case’ Merger Control
Brazil offers another example of this same preemptive impulse, this time in merger control.
In August 2024, Brazil’s competition authority, the Administrative Council for Economic Defense (CADE), opened proceedings into three major AI partnerships: Amazon/Anthropic, Microsoft/Mistral, and Google/Character AI. CADE said the goal was to “understand if these acquisitions, which would not require mandatory notifications, are to be investigated due to potential competitive harm.”
CADE’s own announcement acknowledged that opening proceedings did not necessarily mean the transactions were notifiable or raised competition concerns. The authority identified three possible outcomes: dismissal of the case, approval of the transaction, or initiation of formal merger-review proceedings to assess possible competitive harms.
Under Brazil’s merger rules, to be notifiable, one party to the transaction must exceed BRL 750 million (roughly $148 million) in turnover, while the other must exceed BRL 75 million (roughly $14.8 million). According to CADE’s March 31 technical notes for Amazon/Anthropic, Microsoft/Mistral, and Google/Character AI, the large technology firms each cleared the higher threshold. The AI startups did not.
In any mature merger regime, that should end the inquiry.
Instead, CADE’s General Superintendency referred each case to the Administrative Tribunal under the principle of in dubio pro societate. The authority argued broadly that “digital ecosystems present challenges for competition authorities” and that “the possibility of configuring associative contracts justifies, at least, the precautionary notification of certain contracts as acts of concentration.”
That reasoning stretches the principle beyond recognition. In dubio pro societate—roughly, “when in doubt, favor society”—is a controversial principle borrowed from criminal law. Whatever its proper scope there, it presupposes at least some meaningful doubt about harm. There should be conflicting evidence, or at least concrete indications of risk. In these cases, there appears to be neither.
To be sure, most merger-control systems include “call-in” powers, sometimes called “residual jurisdiction.” These mechanisms allow agencies to review transactions that fall below mandatory-notification thresholds. But such authority is meant to be exceptional, not routine.
The International Competition Network’s Recommended Practices for Merger Notification and Review Procedures emphasize precisely that point. They stress the limited nature of residual jurisdiction and note that, “[w]hen a jurisdiction maintains residual jurisdiction, it should take steps to address the desire of the parties to the transaction for certainty.”
That concern matters because merger thresholds exist for a reason: they provide legal certainty to firms contemplating transactions. Transplanting the controversial criminal-law principle of in dubio pro societate into merger control risks turning AI partnerships into inherently suspect arrangements under Brazilian competition law.
As Dirk Auer and I have argued elsewhere, the opposite is often true. Many of these partnerships may be essential mechanisms for AI challengers to scale. Developing advanced AI models is extraordinarily expensive. It requires computing power, specialized chips, engineering talent, distribution channels, and capital. Partnerships help firms assemble those complementary assets and compete against larger incumbents.
When Ex Ante Rules Age Overnight
The third example is the European Commission’s Digital Markets Act specification procedure concerning Alphabet’s data-sharing obligations under Article 6(11) of the Digital Markets Act (DMA). This is not traditional antitrust enforcement. It is explicit ex ante regulation.
The DMA imposes upfront duties on large digital “gatekeepers,” rather than waiting for a case-by-case finding of anticompetitive conduct. One could therefore argue that regulation is inherently preemptive and that this sort of intervention poses no special problem. But the case illustrates the broader concern running through this post: a pessimistic view of competition can produce flawed diagnoses and premature interventions in regulation as well as antitrust.
Even regulatory intervention requires threshold conditions. It is far from clear that this is an appropriate case for applying the DMA.
According to the Commission, Article 6(11) requires Alphabet to provide certain Google Search data to third-party online search engines on fair, reasonable, and non-discriminatory (FRAND) terms. In practice, FRAND rules are supposed to ensure access on terms that are not exclusionary, discriminatory, or opportunistically expensive. The Commission’s proposed specification would define how that obligation operates, including rules governing data access, anonymization, pricing, and eligibility.
Critically, the Commission’s preliminary measures would extend eligibility to “AI chatbots with online search engine functionalities,” even when search is merely one feature within a broader service.
My skepticism does not stem solely from the fact that Google would be compelled to share data. Data can represent a legitimate competitive advantage, although mandated access may be justified in some circumstances. The deeper concern is that data-sharing obligations can become tools for engineering competitive outcomes, rather than preserving the competitive process.
As Geoffrey Manne, Dirk Auer, and I argued in the International Center for Law & Economics’ (ICLE) recent submission to the consultation, the proposed measures “risk shifting Article 6(11) from a data-access obligation to a tool for delivering competitor success.” The effectiveness of the measures should be understood as creating opportunities to compete, not guaranteeing market-share gains for particular competitors.
More fundamentally, including AI chatbots within the scope of these obligations presupposes that Google can leverage dominance in search into dominance in AI-chatbot markets. That assumption looks increasingly detached from market realities.
The evidence points in the opposite direction: AI tools increasingly threaten Google’s position in search, rather than the reverse.
Google’s global search share fell below 90% for the first time since 2015 during the final quarter of 2024, a decline widely attributed to users shifting toward AI-native alternatives. Early in 2024, Gartner projected that traditional search-engine volume would decline 25% by 2026 because of AI chatbots—a forecast that, so far, appears remarkably prescient.
Meanwhile, Perplexity, which operates without privileged access to Google Search data, processed 780 million queries in May 2025, up 239% year over year, and reached a $20 billion valuation. ChatGPT now reportedly handles roughly 12% of Google’s daily search volume and has reached 900 million weekly users—again, without access to Google’s search data.
Perhaps most tellingly, during the remedies phase of United States v. Google, Apple executive Eddy Cue testified under oath that Google search volume on Safari had declined for the first time in more than two decades. Cue attributed the shift directly to users turning to AI tools for information discovery.
The U.S. court that conducted the most rigorous evidentiary review of Google’s search position reached much the same conclusion. Judge Amit Mehta found that “tens of millions of people use GenAI chatbots, like ChatGPT, Perplexity, and Claude, to gather information that they previously sought through internet search,” and that generative AI represents “a nascent competitive threat” to Google’s dominance in search.
That assessment mattered. It informed Mehta’s decision not to impose structural remedies on Google, with the court noting that competition in AI “is plentiful.” After a full adversarial proceeding examining precisely these market dynamics, the court declined to treat Google’s search dominance as the key to AI competition. If anything, the evidence suggested the opposite: AI competition is emerging as a threat to Google.
The Commission’s specification procedure moves in the other direction. It extends search-data-sharing obligations into a market where Google appears to be under competitive pressure, not successfully leveraging dominance.
That is the broader problem with ex ante AI regulation—and, more generally, with regulatory regimes like the DMA. They can quickly become moving targets. The DMA was designed to avoid the perceived slowness of traditional antitrust enforcement. But AI markets expose the limits of that ambition. When technology and consumer behavior evolve rapidly, ex ante rules risk hardening assumptions that may already be obsolete.
Don’t Freeze the Race at the Starting Line
Taken together, these cases reveal a common pattern.
First, agencies are importing legacy-platform theories into AI markets. Meta’s integration of AI into WhatsApp is framed through familiar tying and leveraging theories. The DMA’s Article 6(11) specification procedure extends data-access concepts developed for search into AI-chatbot services with search functionality. CADE’s scrutiny of AI partnerships draws heavily on “killer acquisition” and nascent-competition theories. (On that point, I strongly recommend Selçuk Ünekbas’s excellent post.)
Second, agencies are intervening before AI markets have stabilized. Interim measures against Meta, DMA specification proceedings, and Brazil’s procedural referrals all reflect a fear that waiting for proof may mean acting too late. From a public-choice perspective, that anxiety is understandable. Agencies face strong incentives to intervene early in politically salient markets. Antitrust law, however, has traditionally imposed safeguards precisely because those incentives exist.
Third, the evidentiary center of gravity is shifting away from demonstrated harm and toward generalized risk management. The recurring vocabulary is telling: firms “may exclude,” “may foreclose,” create “future bottlenecks,” undermine “contestability,” or threaten “effective access.” None of those concepts is meaningless. But they become dangerous when they displace disciplined analysis of market power, foreclosure, efficiencies, and consumer welfare.
That is precisely why Barnett’s error-cost framework matters here. The error-cost approach asks which mistakes are more likely and more harmful: false positives, where lawful conduct is wrongly condemned, or false negatives, where harmful conduct is wrongly allowed. In early-stage markets, false positives can be especially costly. Premature interventions can suppress efficient or experimental business practices, freeze product design, discourage investment, and transform access obligations into competitor subsidies.
Ironically, many of these interventions are defended in the name of protecting dynamic competition. But dynamic competition requires experimentation. Firms must be free to pursue integration, partnerships, open- and closed-source strategies, application programming interfaces (APIs), vertical specialization, and distribution through existing services. If enforcers treat each of those strategies as presumptively suspect whenever a large platform employs them, they risk narrowing the very pathways through which AI competition may emerge.
None of this means agencies should ignore AI. The technology will reshape products and services across the economy. Competition enforcement in this space matters. But it should remain disciplined.
First, authorities should distinguish between power in legacy digital markets and power in AI markets. A firm that dominates messaging, search, e-commerce, mobile operating systems, or cloud infrastructure does not automatically dominate AI assistants, foundation models, or agentic services—AI tools designed to perform tasks on a user’s behalf. Leverage theories require evidence of leverage, not merely evidence of size.
Second, agencies should require actual evidence of foreclosure. Product integration, default placement, and preferential design may matter, but they are not inherently anticompetitive. The relevant question is whether rivals are truly denied access to efficient distribution channels or competitively necessary inputs.
Third, partnerships should be evaluated not only as potential threats, but also as potential engines of competition. AI development requires compute, chips, data, talent, capital, and distribution. Partnerships can help firms assemble those complementary assets, allowing startups to scale and incumbents to challenge current leaders. Scrutiny may be appropriate where there is exclusivity, control, reduced rivalry, or problematic information sharing. Partnership alone, however, should not become a red flag.
Antitrust should remain focused on consumer welfare and innovation, not on preserving existing market structures. If AI changes how users search, communicate, shop, write, and code, competition law should not reflexively protect yesterday’s distribution of traffic, bargaining power, or rents.
The risk is no longer that regulators will arrive too late to AI markets. It is that they will arrive too early, declare the race over, and start handing out medals before the runners have cleared the first turn.
