US Federal Reserve Chair Kevin Warsh is convinced, and prominent economists agree, that AI will meaningfully increase labour productivity. Don’t bet on it. In fact, wider adoption of these models could reduce output per worker.
Widespread adoption of new technologies is a cornerstone of economic dynamism. And in the internet era, accessible user interfaces and affordability have been indispensable enablers. Until the early 1990s, technical wizards relied on rigidly structured, text-based tools like Gopher to exchange documents on the internet. Then, the more flexible World Wide Web emerged, spawning browsers—Mosaic, Netscape Navigator, and Microsoft’s Internet Explorer—and search engines that were quick, easy-to-use, and cheap.
AltaVista, the first “full-text” searchable index of the World Wide Web, went from receiving 300,000 hits on its first day in 1995 to more than 80 million daily hits. But AltaVista defaced its results with ugly banners. So, when Google launched its search engine in 1998, users like me fell for its design, without regard to the quality of its results. Google’s success crushed rival search engines and traditional print publishers, and its engineers miraculously maintained the efficiency of computing infrastructure to ensure that searching for information remained quick, easy, and cheap.
But success also then undermined user experience and productivity. As Google became a monopolist, search-engine optimizers flooded results with useless links. Traditional media resorted to clickbait. This flotsam forced users to learn new skills: how to select keywords, use Boolean operators (like “AND”), and recognise clickbait. Google further degraded the experience by displaying “sponsored” results at the top of the page.
AI chatbots have followed a similar path. ChatGPT attracted a million users in its first five days after its launch in November 2022. As with Google’s initially clean search page, the chatbot’s interface was compelling: users could ask questions in plain English, no Boolean operators required. Unfortunately, ChatGPT turned out to be a mendacious talking horse. Filtering out its fabrications took more time than conducting keyword searches on Google, reducing productivity. Bard—the chatbot that Google, facing an existential threat to its search dominance, rushed to release—was equally disappointing.
Now, more than three years later, Google has shut down Bard and replaced it with both an “AI overview” above its traditional search results and an AI Mode option—a chatbot that urges users to “ask anything.” In my experience, though, the results remain maddeningly unreliable, even for simple queries.
Compared to traditional statistical models, including Google’s pioneering algorithms, large language models appear to offer compelling advantages. With trillions of parameters, LLMs can incorporate contextual factors that earlier models had to ignore. And LLMs do not merely inform; they use metaphors and humour in what the philosopher Ludwig Wittgenstein called “language-games” that reassure, explain, flatter, assert, and cajole.
Like their earlier counterparts, however, LLMs rely on statistical extrapolation, assuming futures that repeat the past. That is fine for natural phenomena like protein folding, but not for ever-changing goods and services. Small modifications to the layout of a laptop battery, for example, can make instructions on how to replace it useless. Despite this, users—like the participants in Stanley Milgram’s notorious shock experiments—reflexively obey a chatbot’s confident answer. By contrast, traditional search links are better at displaying the outdatedness of information and the unreliability of sources. This gives users more discretion to exercise their judgment, which reduces misfires and costs.
Indiscriminately including trillions of parameters magnifies LLMs’ extrapolation problems by increasing the likelihood that they will find nonexistent patterns or select irrelevant or unreliable answers from exhaustive, uncurated catalogues. Designers of traditional statistical models can restrict variables and data sources to avoid that outcome while controlling computational costs.
LLMs’ reliance on statistical extrapolation also makes their conversational interfaces linguistic chimeras. Because LLMs cannot replicate human sense-making and have no real contextual understanding of intent, they do not use language as a “form of life,” to borrow from Wittgenstein again. The pretense that they do has enabled AI peddlers to push chatbots indiscriminately, especially as the narrower applications for which LLMs are actually useful cannot justify multi-trillion-dollar investments.
AI pushers are going all out to addict people. Google’s AI overviews and AI Mode are like a free taste. Once users are hooked, hyperscalers will surely charge high prices to recoup their steep development and operating costs.
One might hope that even the most powerful firms cannot dupe consumers into submission. Despite seemingly unlimited resources, Google has had at least as many flops as hits. Meta’s very name memorialises a failed multi-billion-dollar bet on virtual reality. But the marketing blitz that has turned junk-food addiction into a booming business provides a disheartening counter-example.
Hyperscalers could make LLM addiction as widespread as physiological addictions by exploiting humans’ deep-seated desire for companionable conversation. As social ties fray, chatbots offer a simulacrum, potentially turning serious professionals into lovesick teenagers. And as with other addictions, LLMs are already usurping resources—including capital, electricity, chips, and entrepreneurial energy—that other deserving innovations need. As this continues, techno-optimists should rethink their rosy view of AI’s productivity promise.
The writer is Professor of Health Policy at Columbia University’s Mailman School of Public Health. ©Project Syndicate
