How AI-Driven Search May Reshape Democracy, Economics, and Human Agency

In May 2024, Google unveiled a new way to search the web: the AI Overview, which sits just above the classic PageRank cascade in Google Search. It presents you with direct answers to queries, summaries of the content found in the links below, and a carousel of sources. According to Google, AI Overviews (and AI Mode) promise speed, accuracy, directness of answers, effortless convenience, and an ability to “do more than you ever imagined.”

The rollout sparked both excitement and unease, prompting researchers, publishers, and policymakers to ask how this shift might reshape the economics, governance, and epistemic norms of online search. Last week, Google issued a blog post to reassure the public about these changes. While offering few details, the company reported that overall click volume from Search has remained stable year-over-year, that the proportion of “high-quality clicks” has grown, and that AI results are designed to “highlight the web” rather than replace it. If accurate, these are encouraging numbers for some publishers and creators.

And yet, traffic metrics — whether stable, declining, or improving — capture only part of the story. Preliminary studies voice concerns over the degree to which this innovation cedes ever-greater algorithmic control over information curation to Google, its effects on the web-link–based economy, and the ways in which it might undermine users’ ability to verify, diversify, and weigh the merits of search results. Acknowledging these issues doesn’t mean harkening back to an imagined past in which similar issues were absent from web search in general, or from Google’s search engine in particular. But there are many who see differences in kind, as well as in degree, from the pre-generative model-based online search environment. Indeed, we see a variety of specific concerns spanning the economic, political, social, and cognitive domains.

Algorithmic governance – and power

Algorithms have long shaped public decision-making, but generative AI pushes that influence into a new, less accountable register. When generative models stand between users and the web’s knowledge, they become de facto mediators of civic life—powerful yet unelected actors whose authority no legislature or courtroom has formally conferred. This legitimacy deficit is heightened by two forms of opacity: model opacity and institutional opacity.

In the first place, even if researchers understand the broad architecture of large-language models, that system-level clarity does not reveal why any given query surfaces one fact, image, or viewpoint while omitting another. Model opacity has three practical consequences:

• First, it erodes presentational privacy: people lose the ability to manage how they appear online because they cannot see—much less correct—the composite portrait the model presents to others.
• Second, it frustrates the right to explanation. When an opaque summarizer suppresses a resume snippet or ranks a negative news item prominently, the affected person lacks an intelligible account of that individual decision and therefore lacks any practical route to adapt future behavior or seek redress. In short, the very opacity that powers seamless answers also deprives users of both self-curation and meaningful accountability.
• Third, where AI Overview answers present flawed information (let’s say, for example, about rules for a given public park), there is no ready way to “correct” the record, which may have a complicated, opaque origin in the first place, despite references links being provided.

PageRank-era search already posed privacy and transparency challenges, but generative summaries mark a categorical shift: they replace a visible marketplace of links with a single, pre-digested verdict for the user. That opacity blocks the traditional checks on informational power—click data, outside audits, public scrutiny—and turns the search provider into an unaccountable (and unelected) gatekeeper of knowledge. In effect, a search engine moves from a contestable relevance-ranking system to an “answer authority,” widening the gap between those who set the algorithmic dial and everyone subject to its judgments.

In the second place, institutional opacity is the opacity introduced by a company’s decision to keep key details of its systems—data sources, fine-tuning methods, evaluation protocols—safe, secure, and proprietary. History shows why fuller transparency matters: journalists gained access to outputs from the COMPAS recidivism tool and uncovered racial bias that had gone unnoticed in courtrooms nationwide. In response to such episodes, most leading AI developers—including Google—now publish high-level “system cards” which aim to surface major vulnerabilities in frontier models. These voluntary measures are laudable; they signal an awareness that public oversight is part of the social license to innovate. Still, they stop short of releasing the underlying data, evaluation benchmarks, or long-term performance logs that independent researchers would need to replicate a COMPAS-style audit. For Google’s AI Overviews, that leaves press releases and blog posts like the one issued last week as the primary windows into how the system prioritizes, filters, and frames web knowledge. When the sole mediator of both the search results and the evidence about those results is the same company, the public is asked to judge a black box largely by the keyhole that the box itself has chosen to open.

In short, model opacity obscures the internal logic behind each answer, while institutional opacity withholds the data and documentation needed to probe that logic from the outside. Together they force the public to rely on the goodwill of frontier labs like Google that now stand between citizens and the web’s knowledge.

Economic extraction

This shift towards LLMs in search is driven, in large part, by staggering economic incentives and a potential for widespread consolidation of wealth in the information economy. Anyone who has used OpenAI’s ChatGPT or Anthropic’s Claude for troubleshooting coding tasks or brainstorming dinner recipes knows that getting a direct, tailored answer rather than a link cascade saves a lot of time and is often much more helpful. These use cases and many others explain how Anthropic skyrocketed from a $1 billion annualized revenue in January to $4 billion valuation just five months later. OpenAI is even further down this track, reportedly with $12 billion in annualized revenue. It is understandable that Google feels intense pressure to take advantage of its own formidable “Gemini” foundation model for tasks in Search: competitors will step in to meet user demand if Google is not quick enough to do so. And in theory, Google should have little trouble doing that. Gemini is formidable precisely because it is a natural outgrowth of Google’s dominance as a search engine, specifically as the avatar of enormous data collections, stored in massive datacenters and ripe for use as training data.

And yet, critics fear that—despite Google’s claims to the contrary—as ever more money fills AI lab coffers, less and less money finds its way back to publishers and content creators. These changing dynamics threaten what has been described as the internet’s “grand bargain,” in which bloggers, publishers, illustrators, news agencies, etc., allowed search engines to index their websites under “fair use” in order that these engines could direct monetizable traffic their way, and everyday Internet users “agreed” to the collecting and selling of information as the default for access to digital services and spaces. As more users rely on AI overviews and perform “zero-click” searches, click-through rates may plummet. Some say that they already have, noting drops of up to 55% in the last three years, with more precipitous falls expected in Google’s roll-out of AI mode.

Google denies the legitimacy of such reports, but without published data or research on these trends, it is difficult to impartially access these competing claims, and reasonable to suppose that Google is extending its ability to answer whole classes of queries like “when is the next full moon” without involving traditional sources of such information. This movement builds on Google’s earlier move toward “direct answers,” which were found to push users away from competitors and toward Google’s own services. Scooping up this sector of Search—in spite of the attention of regulators such as the Federal Trade Commission (FTC)—suggests a restricted future for publishers, and a wider (and more profitable) playing field for Search. How could that be, if Google would no longer be directing as many users to the sites where they would formerly have seen ads? Whatever answer Google gives will certainly be centered on its new AI tools. One obvious possibility is that advertisers will pay to have ads immediately following – or perhaps even integrated directly into – Google’s AI Overview results.

And yet, preventing Google from using LLMs to replace search will not solve the problem, as its competitors will gladly carve up its market share. Publishing was by no means a perfect, fair market before. But as search companies potentially lean away from practices that encourage users to click on links, thus impoverishing the companies that relied on link-based traffic, they may also impoverish certain content creators, media creatives, and user-centric discussion platforms supported by publishers. It may be that “average,” larger sites see only modest changes in traffic, but a great deal of destruction may happen at the web’s edges and margins.

Effortless, frictionless search

If companies are driving us toward AI-dominated search, does this mean users genuinely benefit from and prefer AI-generated responses? There is no doubt that many users do favor these responses—even in high-stakes contexts—because they are fast, direct, and convenient. However, it is important not to confuse user preference with what best supports their ability to learn, verify, and understand. In fact, the very friction and effort that users often dislike in traditional search may actually help them discover more relevant or reliable information than the seamless answers provided by LLMs.

Especially in the earliest days of search engines, a chance headline or an unfamiliar domain could easily redirect an inquiry and expose users to unexpected perspectives. The PageRank algorithm and similar ones, along with the new habits of interacting with search engines that they cultivated, have reduced that possibility. Generative AI-based summarization takes us even further from “browsing” in the sense of desultory wandering: it tends to offer a single, ostensibly comprehensive answer. Generative models reproduce the statistical imprint of their training data. For that reason, canonical, Anglophone, and majority viewpoints are privileged in the currently dominant LLMs, while minoritized, emerging, or dissenting perspectives are comparatively muted. The resulting impression of completeness dampens users’ motivation to consult additional sources, tacitly endorsing the system’s framing and, in the process, constraining the pluralism on which democratic deliberation depends.

AI-generated summaries don’t just narrow what users see—they also shape how users think. Earlier search interfaces fostered what Nobel laureate Daniel Kahneman called System 2 thinking—deliberative, effortful reasoning that fortifies epistemic resilience—whereas AI overviews create a frictionless sequence of query, receipt, and tacit acceptance. Should this pattern solidify, users risk losing proficiency in evaluating provenance, identifying bias, and reconciling conflicting evidence—capacities on which democratic deliberation depends. Framed in an authoritative register, such outputs tend to drop qualifiers like “might” and “may,” and thus risk dampening users’ critical reflexes even further, producing an epistemic vulnerability in which individuals encounter content whose lineage is opaque yet feel diminished incentive to interrogate its accuracy. As such, convenience, velocity, and a veneer of certainty can eclipse reflection and the productive role of skepticism.

These harms are accentuated by the current interface architecture of AI overviews. The familiar “ten blue links” interface made informational provenance transparent—any assertion could be followed back to its originating URL. By compressing multiple sources into a single, citation-light paragraph—where references are omitted or relegated to seldom-clicked drop-downs—the interface invites undetected errors, omissions, and novel syntheses, particularly in domains where precision is crucial (e.g., health, finance, or civic data). This loss of transparency is especially troubling at a time when the very same models prove liable to introduce subtle, hard-to-detect distortions—varying by language, culture, and context—and when users may be less able than ever to verify, contest, or even notice these errors.

Finally, since Google is already selling ads against AI overviews, it is unclear how advertising money, specifically promotional payments from third-parties, may ultimately infect AI overview answers, on all manner of queries. After all, Google began as a company disavowing paid links in search results, only to embrace them in order to create perhaps the greatest money-printing machine in history. There has always been a big industry around SEO for brands, politicians, etc., and a similar industry is emerging around generative AI.

Overview of the future

The transformation of search from a collection of ranked links to AI-generated summaries represents more than a technological upgrade—it marks a significant shift in how information flows through society. The three concerns examined here—algorithmic opacity, economic extraction, and epistemic erosion—are not isolated problems but interconnected features of a system that prioritizes efficiency over accountability, convenience over verification, and corporate profit over democratic discourse.

This convergence demands immediate attention from policymakers, technologists, and users alike. We need transparency requirements that make AI decision-making processes auditable, economic frameworks that ensure fair compensation for content creators, and interface designs that preserve opportunities for critical engagement with information. Most urgently, we need public awareness of what we lose when search becomes frictionless: the habits of skepticism, the diversity of perspectives, and the distributed authority that has long characterized the web's knowledge ecosystem.

The stakes extend beyond individual search queries to the very foundations of informed citizenship. As AI overviews become the default gateway to information, we risk creating a generation of users who consume knowledge without question, publishers who cannot sustain quality journalism, and a public sphere increasingly shaped by the statistical patterns embedded in large language models. The potential for economic manipulation and self-dealing by search companies under this new paradigm are immense – and should be monitored. The efficiency gains are real, but so too are the democratic costs. How we navigate this trade-off will determine whether AI-powered search serves as a tool for enlightenment or a mechanism for epistemic capture.