Decision in US vs. Google Gets it Wrong on Generative AI

Kate Brennan is the Associate Director at the AI Now Institute. She is a former Google employee and a graduate of Yale Law School.

Last week, federal court Judge Amit Mehta issued a ruling in the US v. Google remedy case rejecting many of the Department of Justice’s bold, structural remedies to unseat Google’s search monopoly position. As I wrote in Tech Policy Press earlier this year, generative AI was the elephant in the room in this case given Google’s exceptional position to leverage its search monopoly to achieve dominance in the generative AI market. In fact, the DOJ’s proposed remedies were acutely aware of this, initially proposing that Google must, among other bold structural remedies, divest from any interests in rival AI companies.

Despite how much attention generative AI is given in the decision—Judge Mehta wrote that “The emergence of GenAI changed the course of this case”—the decision fails to seriously contend with the tremendous advantages Google holds in the generative AI market due to its search monopoly. It also fails to draw the necessary connections between Google’s search dominance and its control of AI across the entire AI stack.

Most troublingly, the decision overstates the threat that generative AI poses to Google’s dominant search position and understates the advantages that Google holds in the AI market, all but assuring that Google will continue to leverage its monopoly power in search in the AI market.

Right off the bat, the decision improperly positions generative AI as a magic fix to competition in the search engine market. The decision asserts that the generative AI space is competitive—citing circumstantial evidence from Google executives who are incentivized to make such claims. Regarding the competitive space that Gemini occupies, the decision cites Sissie Hsiao, Gemini’s former general manager, who said, “I would say I don’t think I’ve seen a more fierce competition ever in my 20-some years of working in technology.” From Eli Collins, vice president of Google’s DeepMind, “[Foundation model competition] is the most competitive market I’ve ever worked in.”

The decision’s actual market evidence of competition is meager at best, taking a sampling of noisy market signals and misinterpreting them as genuine competition. For example, the decision cites DeepSeek, the Chinese model released in December 2024, as evidence of a strong new entrant. Yet the emergence of DeepSeek in the generative AI market did not tell a story of a competitive LLM market but rather one that is trending towards commodification. By name-dropping models and AI firms without interrogating the larger market dynamics, the decision mistakes product launches for genuine competition.

In this current paradigm, where LLMs are becoming cheaper and easier to deploy, market dominance is determined by those players who are best positioned to turn LLM access into a profitable business model. This is precisely where Google reigns supreme, as Google maintains control over key structural chokepoints, from AI infrastructure to pathways to the consumer. Compare the decision’s nod to AI start-up Perplexity, which “continues to negotiate with other OEMs and browser developers” to make its product accessible on devices with Google who (checks notes) owns its own operating system, browser, and devices.

Other market signals cited by the court don’t stand up to scrutiny, either. The decision cites that generative AI firms have access to a significant amount of capital, presumably showing they can match Google’s capital expenditure. The decision’s primary evidence is OpenAI’s recent announcement to raise an additional $40 billion at a valuation of $300 billion. Yet the court fails to contend with the extraordinary burn rate of these companies, which their revenue streams have failed to catch up to: The Information reported this week that OpenAI anticipates burning through $115 billion by 2029, and underestimated its spend by an astonishing $80 billion.

What differentiates Google is its existing ecosystem of products, which have become infrastructural for most consumers. ChatGPT loses at least $2 for every $1 it spends, and, as Brian Merchant wrote for AI Now, the company fails to have a discernable or profitable business model. Now, as AI increasingly becomes more expensive to integrate into products, firms who are best positioned to succeed are those who have alternative revenue streams to offset costs, profit off of high compute costs, or those who can easily integrate AI into their existing product workflow. Google is able to take capital intensive risks on AI because of its ongoing search monopoly profits. By owning Google Cloud, Google can discount the costs of running its own models. And finally, Google can integrate AI into all of its existing products, tapping into a user base of billions. In fact, a Google executive testified in this case that Google has already integrated generative AI into every single one of Google’s products—again giving the company an impossibly strong upper hand on those AI start-ups listed as competition.

Finally, when it comes to disrupting Google’s search market share, I have written that claims that ChatGPT is meaningfully chipping away at Google’s search market share fail to consider how dominant Google’s entrenched position is. This decision only adds support to this claim. The decision fails to find any concrete evidence that generative AI products have had a significant impact on general search engine usage. Instead, the court finds that AI overviews have strengthened Google’s search position, with search queries increasing 1.5 to 2%. This is not to say that how people access information across the internet may change over time. Instead, it is a fundamental mistake to talk about changes in information access without documenting how the structural chokepoints to access that information—from the cloud to data centers to browsers to assistants—remain firmly in Google’s control.

Taken cumulatively, this evidence swayed Judge Mehta to reject the bold—and frankly sensible—remedies recommended by the DOJ, including some of the more obvious behavioral remedies. For example, Judge Mehta denied a prohibition on search related payments to distributors on the assertion that these payments are made more palatable by the ostensible threat generative AI poses to search competition. Writes Judge Mehta, “These companies already are in a better position, both financially and technologically, to compete with Google than any traditional search company has been in decades.”

But none of Google’s generative AI competitors have a customer base that approximates Google’s scale—and Google is already pushing AI onto these consumers in every single product it owns. The opinion wholly overlooks Google’s second most strategic asset when it comes to AI dominance, cloud infrastructure, which was made possible because of the reinvested revenue and scale achieved through search. And almost incredulously, the decision dismisses the data advantages Google has in training its LLMs. The court finds that Google holds no technical advantage in the LLM foundation model market and even goes so far as to say that Google does not use search data to train its LLMs—despite contradicting evidence directly from a Google executive who publicly discussed how Google’s data access gives it an edge in training LLMs. And even if it is true that Google doesn’t directly use its endless trove of user data, it is able to use the fruits of its search monopoly (revenue) to buy as much exclusive data as it wants, edging out other rivals.

Gesturing towards the importance of generative AI in the search engine market then dismissing its actual effects is a dangerous precedent. It is true that tech markets are being shaped by generative AI. But in this case the court failed to accurately examine the broader AI market and the effects of consolidated power. We do not have to gaze into a crystal ball, as Judge Mehta suggests, to know what will happen next: Google will continue to enjoy the fruits of its search monopoly to secure its success in the generative AI market—including trillions of real-time search queries to train its models, unparalleled access to consumers across dozens of products to push AI onto, and billions of dollars in revenue to invest in risky capital investments.