The digital landscape is undergoing a profound transformation with the emergence of conversational AI search engines, tools powered by large language models (LLMs) that promise to redefine how users access and interact with information online. In February 2024, Fast Company [1] highlighted this burgeoning trend, noting the rapid proliferation of applications capable of retrieving, summarizing, and presenting information from the Internet in a conversational format. This development marks a significant departure from the keyword-based search paradigm that has dominated the digital experience for decades, largely shaped by Google’s ubiquitous model. Academic and scientific research institutions are at the forefront of this wave of experimentation, exploring the capabilities and implications of these advanced search mechanisms [2]. The enthusiastic reception surrounding this new modality of information discovery underscores a collective anticipation for an alternative that moves beyond traditional link-centric results.
The Evolution of Information Retrieval: From Keywords to Conversation
For over two decades, the internet search experience has been largely synonymous with typing keywords into a search bar and sifting through a ranked list of links on a Search Engine Results Page (SERP). This model, pioneered and perfected by companies like Google, revolutionized information access, making vast swathes of human knowledge discoverable. Users developed a specific mental model: articulate a query, receive a directory of potential sources, and then independently evaluate and navigate those sources to extract the desired information. This process, while incredibly powerful, often involved a degree of cognitive load, requiring users to discern credibility, synthesize information from multiple pages, and tolerate irrelevant results.
The advent of large language models (LLMs) and generative AI, exemplified by the widespread adoption of tools like ChatGPT, has fundamentally shifted expectations. Users are now accustomed to interacting with AI in a dialogue format, receiving direct, synthesized answers rather than just pointers to information. This ‘conversational turn’ in AI has naturally extended to search, giving rise to engines like Perplexity AI and Andi. These platforms leverage the LLM’s capacity for natural language understanding and generation to interpret complex queries, scour the web for relevant data, and then compose coherent, summarized responses. This represents not merely an incremental improvement but a paradigm shift in the fundamental interaction model for information retrieval, moving from a directory-based system to a direct-answer system.
A New Mental Model for Search: Beyond the SERP
Conversational AI search engines are not merely augmenting existing search interfaces; they are introducing a fundamentally different mental model for how users conceive of and interact with search. While platforms like Perplexity AI, as illustrated in Figure 1 (referencing the original article’s Figure 1 depicting Perplexity’s homepage), retain familiar elements such as an input field and a central display for results, the underlying interaction is profoundly different. Unlike classic search engines that return a list of links, Perplexity provides a synthesized answer, drawing snippets from multiple sources, much like a knowledgeable assistant.
This shift is precisely what Figure 2 (referencing the original article’s Figure 2 contrasting mental models) aims to articulate: the user’s mental model transitions from that of a librarian pointing to books, to a conversational chatbot directly answering questions. This aligns closely with the user experience popularized by ChatGPT, where generative AI applications thrive on conversational interactions. Andi, another emerging search engine (Figure 3 in the original article), takes this conversational interface to an even greater extreme in its visual layout and interaction patterns. While its information architecture still somewhat resembles a classic search engine by presenting a clear list of sources, its primary mode of engagement is through dialogue.
These innovations are prime examples of a "technology push" [3], where the immense capabilities of a new technology—generative AI—drive the development of novel digital experiences. The market is responding to the sheer power and potential of LLMs, creating products that showcase what this technology can achieve. However, as with any technology push, the ultimate benefit to the user, and the long-term strategic viability, remain critical questions. The rapid deployment of new products driven by technological novelty sometimes risks overlooking human-centered design principles and the potential for unintended consequences.
Enhancing the User Experience: Immediate Benefits and Usability Gains
The immediate and undeniable advantage of conversational AI search engines lies in their ability to dramatically improve the usability and efficiency of the search experience. The query example often seen on platforms like Perplexity (e.g., in Figure 1 of the original article) demonstrates key elements that contribute to a more pleasant and rewarding user journey.
Firstly, the format of the search result is a direct, coherent text, akin to a web page article itself. This contrasts sharply with the traditional SERP, which typically offers a list of links that might contain the answer. With conversational AI, the answer is served directly, compiled and summarized from various sources. This eliminates several steps in the user’s cognitive process: the need to evaluate multiple search results, make educated guesses about which link is most relevant, open new tabs, and then scan through potentially lengthy texts to find the specific piece of information.
The user’s fundamental goal remains constant across both classic and AI-powered search: to find an answer within a text. However, the AI model significantly streamlines this process. By directly providing the answer, these engines effectively perform the initial sifting, evaluation, and summarization tasks on behalf of the user. This simplification closely mirrors the natural Q&A pattern inherent in human conversation, addressing a crucial usability heuristic: matching the system to the real world [4].
Furthermore, AI-powered search engines often enhance the user experience by implicitly or explicitly adhering to other Nielsen’s usability heuristics. For instance, they aim to provide:
- Visibility of system status: By often indicating that they are "generating an answer" or "searching," they keep the user informed.
- User control and freedom: While providing direct answers, many also offer follow-up questions or options to explore sources, giving users agency.
- Error prevention: By synthesizing information, they attempt to reduce the likelihood of users landing on irrelevant or misleading pages.
- Recognition rather than recall: Presenting direct answers means users don’t need to remember complex query syntax or navigate intricate site structures.
- Flexibility and efficiency of use: They cater to both novice users seeking simple answers and expert users who might appreciate the quick synthesis before diving deeper.
- Aesthetic and minimalist design: Often, the conversational interface is clean and focused, reducing clutter.
- Help and documentation: Though often implicit, the conversational nature itself can act as a form of "help" by guiding the user’s interaction.
While these engines undoubtedly "tick the right boxes" for improving the immediacy and convenience of the search journey, the crucial question remains: does this optimization for convenience necessarily make them the ideal standard for the future of search, especially when considering broader societal and cognitive implications?
Broader Implications: Trust, Explainability, and the Erosion of Human Agency
The profound efficiency offered by conversational AI search engines comes with a set of complex implications, particularly concerning trust, AI explainability, and the potential impact on human agency and critical thinking. When using tools like Perplexity, it becomes starkly apparent how much decision-making responsibility is being delegated to the AI. The generative technology selects, extracts, and summarizes content, crafting what appears to be an authoritative answer. However, the application often falls short in communicating why certain sources were chosen over others, or how the summary was derived. This lack of transparency neglects a crucial aspect of AI explainability [5], which, in turn, directly impacts the trustworthiness of the system.
Francesca Rossi, IBM’s Global Ethics Leader, eloquently articulates this challenge: "[AI raises] some concerns, such as its ability to make important decisions in a way that humans would perceive as fair, to be aware and aligned to human values that are relevant to the problems being tackled, and the capability to explain its reasoning and decision-making" [6]. The risks associated with untrustworthy AI systems are increasingly becoming a pressing concern in the enterprise sector, driven by impending regulations, the specter of substantial fines, reputational damage, and legal liabilities. However, the individual internet user, often referred to as an "Internaut" – a highly skilled, habitual user of the internet – may not immediately perceive these risks in the same vein.
Roberta Katz, an expert on Gen Z, draws a compelling analogy: "There is a saying about architecture that has long seemed true to me: first you make the building and then the building makes you. A light and airy house can create one kind of emotional and behavioral response for its residents and, similarly, a dark and heavy home can create an altogether different environment" [7]. Katz extends this observation to digital architectures, highlighting how the design of our applications and IT systems significantly shapes the behavior and cognitive patterns of their users.
Applying this principle, a significant concern with tools like Perplexity is the potential for users to become overly accustomed to and accepting of convenient, ready-made answers, without pausing to critically question their accuracy, bias, or even outright veracity. While these tools typically provide links to source documents, several critical questions remain difficult for the user to ascertain:
- The credibility and bias of the selected sources: How reliable are the underlying websites or publications from which the information is drawn?
- The completeness and balance of the summary: Has the AI presented a comprehensive overview, or has it selectively emphasized certain perspectives while omitting others? This issue is compounded by the "black box" nature of many LLMs, where the internal workings of their decision-making processes are not transparent.
While the first issue (source credibility) is also a challenge with traditional search engines, the second (completeness and balance of AI-generated summaries) arises directly from the deeper delegation of decision-making inherent in generative AI. Even if these tools were to provide full explainability, the sheer ease of consuming pre-digested answers could, over time, subtly erode users’ critical thinking skills. This mirrors concerns raised about how text messaging, for instance, might have contributed to a decline in formal writing proficiency.
From a purely usability standpoint, the process of sifting through Google’s SERP, patiently comparing and piecing together information from various articles to construct an answer, is undeniably more tedious and cognitively demanding than quickly consuming a concise, authoritative-sounding AI summary. However, these more laborious tasks serve as a vital "exercise" for our analytical and creative faculties. They inherently foster a degree of skepticism and critical evaluation, potentially discouraging an uncritical over-reliance on AI models. If the prevailing mental model of a search engine evolves into that of an infallible oracle that "always knows the right answer," the implications for our collective ability to critically discern truth from falsehood could be enormous. This might be inconsequential when searching for mundane information like purchasing a new pair of trousers or removing a grease stain. However, for more critical domains such as policymaking, scientific research, or understanding civil society issues, the consequences of an uncritical acceptance of AI-generated answers could be severe and far-reaching.
Beyond cognitive impact, there are also significant economic implications for content creators and the broader digital ecosystem. Kevin Roose of The New York Times posed a pertinent question: "If AI search engines can reliably summarize what’s happening in Gaza or tell users which toaster to buy, why would anyone visit a publisher’s Web site ever again?" [8] This highlights a genuine concern about the potential for AI search to disintermediate content creators, drastically reducing web traffic to original sources and undermining traditional advertising-supported business models. The value proposition of generative AI search, while beneficial for the end-user, could inadvertently starve the very content ecosystem upon which it feeds.
Interestingly, not all AI search experiences are designed to maximize this disintermediation. Andi’s architecture, for example, while featuring a chatbot-like user interface, does not stray as far from the traditional mental model of search. As seen in Figure 3 (referencing the original article’s Figure 3 showing Andi’s UI), the brief answer provided is often a direct snippet from the first relevant source, such as Wikipedia, rather than a fully generated summary. Crucially, the results prominently display a list of all sources with direct links, appearing on the left side of the page. Conceptually, this model, despite its modern layout, is not fundamentally dissimilar from a traditional SERP. Andi’s design thus encourages users to click through to the source websites, demonstrating how thoughtful UX and UI design can mitigate some of the negative economic impacts by eliciting more desirable user behaviors and supporting the original content ecosystem.
Navigating the Future: The Imperative of Trustworthy AI Search
The most pertinent question regarding the future of information retrieval is not whether AI will be the future of search—as it almost certainly will be, given its capabilities to enhance user experience and accelerate knowledge acquisition. Rather, the critical inquiries revolve around how this future will be shaped:
- How can AI search engines be developed ethically and responsibly to maximize benefits while mitigating risks?
- How can they be designed to foster, rather than diminish, critical thinking and human discernment?
- How can they ensure fairness, transparency, and accountability in their operation?
A study published in Nature [9] vividly illustrates the conflicting perspectives on AI science search engines among researchers. Some lauded their utility and accuracy, while others expressed profound distrust and frustration over the inconsistency of AI’s retrieval performance. This underscores that trust remains a central impediment to widespread AI adoption. Addressing this deficit requires a concerted focus on two key areas:
-
Enhancing Technical Robustness and Transparency:
- Improved Source Attribution and Citation: Clear, verifiable links to original sources, indicating why a source was chosen and its relative authority.
- Mitigation of AI Hallucinations: Technologies like Retrieval-Augmented Generation (RAG) [10] are crucial. RAG grounds LLMs in verifiable external knowledge bases, significantly reducing the propensity for AI to generate false or misleading information, often referred to as "AI hallucinations" [11]. This ensures that answers are not merely plausible but factually accurate and traceable.
- Transparency in Summarization and Selection: Developing mechanisms to explain how a summary was constructed from disparate pieces of content, including highlighting key arguments or contradictory information from sources.
- User Feedback Loops: Implementing robust systems for users to flag inaccuracies or biases, contributing to the continuous improvement of the AI model.
- Ethical Data Sourcing: Ensuring that the training data for LLMs is diverse, unbiased, and ethically obtained, avoiding the perpetuation of societal biases.
-
Fostering User Literacy and Critical Engagement:
- Educational Initiatives: Empowering users with the knowledge and skills to critically evaluate AI-generated content, understand its limitations, and recognize potential biases.
- Interface Design for Criticality: Designing interfaces that subtly encourage users to investigate sources, compare information, and question conclusions, rather than passively accepting direct answers. Andi’s approach of prominently displaying sources is a step in this direction.
- Transparency Statements: Clear disclosures about the generative nature of the content and the potential for errors or biases.
As new regulations emerge globally, aimed at ensuring safer and more ethical uses of AI, the fundamental question of whether a "good" search experience (in terms of speed and convenience) is also the "right" experience (in terms of societal and cognitive impact) becomes increasingly relevant. Within a context where individuals are delegating a substantial portion of their decision-making and information synthesis to LLMs, design efforts must extend far beyond merely building a super user-friendly conversational interface. The paramount focus must shift towards ensuring accuracy, fostering trustworthiness, and guaranteeing the comprehensiveness and ethical grounding of search outputs [12]. The future of AI search demands a delicate balance between unparalleled convenience and an unwavering commitment to intellectual integrity and human empowerment.
