For nearly two decades, searching was the default way people accessed information online. Users typed keywords, scanned lists of results, and moved across multiple sources to construct an understanding. This process was not always efficient, but it required active engagement. Recently, however, a different pattern has begun to emerge. Instead of searching, many users now ask.Generative AI systems, particularly large language models (LLMs), are increasingly becoming the first point of contact in information-seeking. Users pose a question and receive a single, synthesized response. Often, that response is treated as sufficient. This shift from "searching" to "asking" may appear to be a simple improvement in convenience. Yet empirical evidence suggests that it reflects a deeper transformation. Adoption of LLMs has been associated with a measurable decline in traditional web search behavior (Padilla et al., 2025), while conversational interfaces are increasingly shaping how users approach information-seeking tasks (Pham et al., 2024).This article offers an interpretation of this shift. Rather than treating generative AI as a more efficient search tool, I suggest that it represents a change in how knowledge is accessed, processed, and trusted. Drawing on recent empirical findings, I distinguish between search-based and answer-based cognition and consider how this transition affects cognitive engagement and epistemic authority. The goal is not to evaluate whether this shift is good or bad, but to clarify what is changing and why it matters.These patterns do not imply that search has disappeared. Rather, they point to a shift in how users allocate attention between searching and asking, depending on task and context.This shift also connects with recent discussions on the relationship between artificial intelligence, cognition, and epistemic judgment. Perc (2025) argues that LLMs may reproduce the appearance of evaluative judgment without replicating the contextual understanding, deliberation, and epistemic responsibility that characterize human judgment. Similarly, Perc and Ö zer (2025) emphasize that increasing reliance on generative AI may alter the cognitive practices through which users retrieve, evaluate, and synthesize information. These works suggest that the consequences of generative AI should not be understood only in terms of technical performance or ease of access, but also in relation to changing forms of human cognition and judgment.The shift toward generative AI is not only technological but behavioral. Padilla et al. (2025) show that once users adopt LLM-based systems, their reliance on traditional search engines declines. This suggests that generative AI is not merely supplementing existing practices, but in some cases replacing them. Similarly, Pham et al. (2024) find that AI-powered interfaces often become the first system users consult, shaping both the starting point and structure of inquiry. Importantly, these effects are likely to be task-dependent and may vary across domains and user groups.What distinguishes these systems is not just their ability to retrieve information, but their ability to present it in a fully synthesized form. Traditional search engines provide multiple, often competing results. Users must decide which sources to trust and how to integrate them. In contrast, generative AI systems compress this process into a single response. The answer is already assembled.This difference changes the role of the user. Under search-based interaction, users act as navigators. They explore, compare, and construct meaning across sources. Under answer-based interaction, users engage with a finished product. Their task shifts toward interpreting or accepting the response rather than actively assembling it.This transition is subtle, but it matters. It reshapes the structure of information engagement, moving from a distributed and exploratory process to a more centralized and immediate one.If interaction patterns change, it is reasonable to expect cognitive processes to change as well. Experimental work by Melumad and Yun (2025) provides some evidence for this. In their studies, participants who learned through LLM-generated explanations tended to develop shallower understanding compared to those who relied on traditional web-based search, even when the informational content was equivalent. These findings should be interpreted cautiously. They reflect specific experimental settings and do not imply that LLM use invariably leads to shallow understanding across all contexts.The difference lies in how the information is encountered. Search-based learning requires effort. Users must sift through multiple sources, evaluate relevance, and integrate partial information into a coherent whole. These activities promote deeper cognitive processing. In contrast, LLM-generated responses present a polished explanation upfront, reducing the need for such effort. This concern is consistent with recent discussions of the cognitive costs of generative AI. Perc and Ö zer (2025) suggest that when LLMs take over tasks such as retrieving, evaluating, and synthesizing information, users may lose opportunities to exercise the cognitive skills that are developed through active engagement with complex content.This does not imply that generative AI leads to poor learning outcomes in all cases. It does, however, suggest that efficiency comes with trade-offs. When cognitive effort is reduced, opportunities for deeper processing may also decline.This brings up a broader question about how generative AI systems are designed and evaluated. Current models are often optimized for speed, clarity, and user satisfaction. These are reasonable goals. But they may not fully capture how such systems influence learning and understanding. If AI systems increasingly mediate how people engage with information, then their impact on cognitive processes becomes a relevant consideration.Beyond cognition, the shift toward generative AI also affects how users assign trust. Hauswald (2025) introduces the concept of artificial epistemic authority to describe situations in which users treat AI outputs as reliable grounds for belief. This form of authority does not depend on the system possessing knowledge in a human sense. Rather, it emerges when users consistently defer to the system's responses. This characterization is not meant to deny user agency, but to describe a tendency toward increased reliance under certain conditions.There are indications that generative AI systems are beginning to occupy such roles. In educational contexts, for example, Jose et al. (2025) show that students often perceive AI-generated explanations as legitimate sources of knowledge. These systems do not simply provide information; they shape perceptions of credibility.Part of this shift is related to how AI systems present their outputs. Responses are typically fluent, coherent, and delivered with a unified voice. Unlike search engines, they do not require users to navigate conflicting sources. As a result, they can appear more definitive than they actually are. Perc (2025) describes this problem in terms of "counterfeit judgments," emphasizing that LLMs may produce outputs that simulate evaluative judgment without being grounded in human-like understanding or deliberative reasoning. This distinction helps explain why answer-based interaction can reshape epistemic authority: users encounter outputs that appear already evaluated, even though the evaluative process behind them differs fundamentally from human judgment.This has implications for how users evaluate information. In search-based environments, the presence of multiple sources introduces opportunities for comparison and doubt. In AI-mediated environments, these intermediate steps are often absent. The user is presented with a single answer, and the burden of evaluation is less visible. This is not to say that users blindly trust AI systems. However, the structure of interaction may encourage default acceptance, particularly when the system has previously produced useful responses.The transition from searching to asking is not easily categorized as either a problem or a solution. On one hand, generative AI systems offer clear benefits. They reduce the time and effort required to access information and can make complex topics more accessible. From this perspective, they represent an efficient reorganization of epistemic labor.On the other hand, the same features that make these systems efficient may also reshape how knowledge is constructed and evaluated. Reduced exposure to multiple sources, diminished need for comparison, and increased reliance on synthesized answers may affect both cognitive engagement and trust formation.These developments raise questions that extend beyond technical performance. For example, how should generative AI systems balance efficiency with opportunities for deeper engagement? Should systems be designed to present multiple perspectives rather than a single answer? How should uncertainty be communicated to users? These considerations are intended as possible directions rather than prescriptive requirements.These are not purely technical questions. They involve assumptions about how knowledge should be acquired and what kinds of cognitive practices should be supported. At present, these considerations are not always central to system design.It may be useful to think in terms of what could be called epistemic alignment. That is, aligning AI systems not only with user intent, but also with practices that support responsible belief formation. This does not require abandoning efficiency, but it does suggest that efficiency should not be the sole priority. In this sense, epistemic alignment also requires attention to whether AI systems complement or replace the cognitive activities through which users learn to evaluate information, form judgments, and revise beliefs.At the same time, it is important to avoid overstating the risks. Humans have always relied on external sources of knowledge, including experts, institutions, and technologies. Generative AI can be seen as part of this broader pattern. The key difference is the degree to which these systems integrate and present information, potentially obscuring the processes behind it.Given these competing considerations, a fully optimistic or pessimistic stance may be premature. The shift toward AI-mediated information environments is already underway. The more relevant question is how these systems will be integrated into existing practices and how their design will shape future patterns of inquiry.The claims made here are modest and interpretive, aimed at clarifying emerging tendencies rather than establishing universal conclusions. The growing tendency to ask rather than search reflects more than a change in user preference. Understanding this shift requires moving beyond a narrow focus on technological capability. Generative AI systems are not only tools for delivering answers; they are increasingly part of the infrastructure through which knowledge is formed. As such, their design and use have implications for cognition, trust, and the broader organization of epistemic practices.Whether this transformation ultimately enhances or diminishes human understanding remains an open question. What seems clear, however, is that the move from searching to asking is neither trivial nor purely technical. It is a change that warrants closer attention, not only in terms of what these systems can do, but also in terms of how they shape the ways in which we come to know.
Jun Woo Kwon (Tue,) studied this question.