For nearly two years, the world’s biggest tech companies have said that AI will transform the web, your life, and the world. But first, they are remaking the humble search engine.
Chatbots and search, in theory, are a perfect match. A standard Google search interprets a query and pulls up relevant results; tech companies have spent tens or hundreds of millions of dollars engineering chatbots that interpret human inputs, synthesize information, and provide fluent, useful responses. No more keyword refining or scouring Wikipedia—ChatGPT will do it all. Search is an appealing target, too: Shaping how people navigate the internet is tantamount to shaping the internet itself.
Months of prophesying about generative AI have now culminated, almost all at once, in what may be the clearest glimpse yet into the internet’s future. After a series of limited releases and product demos, mired with various setbacks and embarrassing errors, tech companies are debuting AI-powered search engines as fully realized, all-inclusive products. Last Monday, Google announced that it would launch its AI Overviews in more than 100 new countries; that feature will now reach more than 1 billion users a month. Days later, OpenAI announced a new search function in ChatGPT, available to paid users for now and soon opening to the public. The same afternoon, the AI-search start-up Perplexity shared instructions for making its “answer engine” the default search tool in your web browser.
[Read: The AI search war has begun]
For the past week, I have been using these products in a variety of ways: to research articles, follow the election, and run everyday search queries. In turn I have scried, as best I can, into the future of how billions of people will access, relate to, and synthesize information. What I’ve learned is that these products are at once unexpectedly convenient, frustrating, and weird. These tools’ current iterations surprised and, at times, impressed me, yet even when they work perfectly, I’m not convinced that AI search is a wise endeavor.
For decades, the search bar has been a known entity. People around the world are accustomed to it; several generations implicitly regard Google as the first and best way to learn about basically anything. Enter a query, sift through a list of links, type a follow-up query, get more links, and so on until your question is answered or inquiry satisfied. That indirectness and wide aperture—all that clicking and scrolling—are in some ways the defining qualities of a traditional Google search, allowing (even forcing) you to traverse the depth and breadth of connections that justify the term world-wide web. The hyperlink, in this sense, is the building block of the modern internet.
That sprawl is lovely when you are going down a rabbit hole about Lucrezia de Medici, as I did when traveling in Florence last year, or when diving deep into a scientific dilemma. It is perfect for stumbling across delightful video clips and magazine features and social-media posts. And it is infuriating when you just need a simple biographical answer, or a brunch recommendation without the backstory of three different chefs, or a quick gloss of a complex research area without having to wade through obscure papers.
In recent years, more and more Google Search users have noted that the frustrations outweigh the delight—describing a growing number of paid advertisements, speciously relevant links engineered to top the search algorithm, and erroneous results. Generative AI promises to address those moments of frustration by providing a very different experience. Asking ChatGPT to search the web for the reasons Kamala Harris lost the presidential election yielded a short list with four factors: “economic concerns,” “demographic shifts,” “swing state dynamics,” and “campaign strategies.” It was an easy and digestible response, but not a particularly insightful one; in response to a follow-up question about voter demographics, ChatGPT provided a stream of statistics without context or analysis. A similar Google search, meanwhile, pulls up a wide range of news analyses that you have to read through. If you do follow Google’s links, you will develop a much deeper understanding of the American economy and politics.
Another example: Recently, I’ve been reading about a controversial proposed infrastructure project in Maryland. Google searches sent me through a labyrinth of public documents, corporate pitches, and hours-long recordings of city-council meetings, which took ages to review but sparked curiosity and left me deeply informed. ChatGPT, when asked, whipped up an accurate summary and timeline of events, and cited its sources—which was an extremely useful way to organize the reading I’d already done, but on its own might have been the end of my explorations.
I have long been a critic of AI-powered search. The technology has repeatedly fabricated information and struggled to accurately attribute its sources. Its creators have been accused of plagiarizing and violating the intellectual-property rights of major news organizations. None of these concerns has been fully allayed: The new ChatGPT search function, in my own use and other reports, has made some errors, mixing up dates, misreporting sports scores, and telling me that Brooklyn’s Prospect Park is bigger than Manhattan’s (much larger) Central Park. The links offered by traditional search engines are filled with errors too—but searchbots implicitly ask for your trust without verification. The citations don’t particularly invite you to click on them. And while OpenAI and Perplexity have entered into partnerships with any number of media organizations, including The Atlantic—perhaps competing for the high-quality, human-made content that their searchbots depend on—exactly how websites that once relied on ad revenue and subscriptions will fare on an AI-filtered web eludes me. (The editorial division of The Atlantic operates independently from the business division, which announced its corporate partnership with OpenAI in May.)
[Read: AI search is turning into the problem everyone worried about]
Although ChatGPT and Perplexity and Google AI Overviews cite their sources with (small) footnotes or bars to click on, not clicking on those links is the entire point. OpenAI, in its announcement of its new search feature, wrote that “getting useful answers on the web can take a lot of effort. It often requires multiple searches and digging through links to find quality sources and the right information for you. Now, chat can get you to a better answer.” Google’s pitch is that its AI “will do the Googling for you.” Perplexity’s chief business officer told me this summer that “people don’t come to Perplexity to consume journalism,” and that the AI tool will provide less traffic than traditional search. For curious users, Perplexity suggests follow-up questions so that, instead of opening a footnote, you keep reading in Perplexity.
The change will be the equivalent of going from navigating a library with the Dewey decimal system, and thus encountering related books on adjacent shelves, to requesting books for pickup through a digital catalog. It could completely reorient our relationship to knowledge, prioritizing rapid, detailed, abridged answers over a deep understanding and the consideration of varied sources and viewpoints. Much of what’s beautiful about searching the internet is jumping into ridiculous Reddit debates and developing unforeseen obsessions on the way to mastering a topic you’d first heard of six hours ago, via a different search; falling into clutter and treasure, all the time, without ever intending to. AI search may close off these avenues to not only discovery but its impetus, curiosity.
The issues with factuality and attribution may well be resolved—but even if they aren’t, tech companies show no signs of relenting. Controlling search means controlling how most people access every other digital thing—it’s an incredible platform to gain trust and visibility, advertise, or influence public opinion.
The internet is changing, and nobody outside these corporations has any say in it. And the biggest, most useful, and most frightening change may come from AI search engines working flawlessly. With AI, the goal is to keep you in one tech company’s ecosystem—to keep you using the AI interface, and getting the information that the AI deems relevant and necessary. The best searches are goal-oriented; the best responses are brief. Which perhaps shouldn’t be surprising coming from Silicon Valley behemoths that care, above all, about optimizing their businesses, products, and users’ lives.
A little, or even a lot, of inefficiency in search has long been the norm; AI will snuff it out. Our lives will be more convenient and streamlined, but perhaps a bit less wonderful and wonder-filled, a bit less illuminated. A process once geared toward exploration will shift to extraction. Less meandering, more hunting. No more unknown unknowns. If these companies really have their way, no more hyperlinks—and thus, no actual web.
theatlantic.com