Greetings, I’m Yves. Throughout his extensive academic career, Amar Bhide has focused on the genuine aspects of entrepreneurship and innovation, distinctly separate from the extractive practices often seen in the financial services sector. He expresses skepticism about the ability of AI to fulfill its lofty productivity promises. Instead, he believes that AI chatbots are designed more to create psychological dependency among users than to provide genuinely beneficial services. As Amar pointed out in a recent email:
Experts and champions of opinion tend to overlook the addictive nature of these tools. There’s a prevailing belief among both skeptics and advocates that AI will enhance productivity, yet discussions often veer away from the larger issue of widespread distraction.
As the well-known Marx-aligned phrase goes, from Chico in *Duck Soup* and not from Karl in *Das Kapital*, “Who you gonna believe? Me or your own eyes?”
Will you trust the experts or your own observations?
By Amar Bhidé, Professor of Health Policy at Columbia University’s Mailman School of Public Health, and the author of the upcoming book Uncertainty and Enterprise: Venturing Beyond the Known (Oxford University Press, 2024). Originally published at Project Syndicate
Kevin Warsh, the Chair of the US Federal Reserve, is confident in the belief that AI will significantly boost labor productivity, a view echoed by several prominent economists who share similar sentiments. However, I urge caution in placing such bets. The broader integration of these AI models may, in fact, lead to decreased productivity per worker.
The introduction of new technologies has historically been a fundamental ingredient of economic innovation. During the internet’s formative years, user-friendly interfaces and cost-effectiveness were critical factors. Before the 1990s, tech-savvy individuals depended on rigid, text-based tools like Gopher for document sharing. Then the World Wide Web revolutionized the landscape, leading to more adaptable browsers such as Mosaic, Netscape Navigator, and Microsoft’s Internet Explorer, which offered user-friendly and economical access.
In 1995, Alta Vista, the first full-text searchable index of the internet, surged from receiving 300,000 hits on its launch day to over 80 million daily hits just two years later. However, the site’s cluttered advertisements tarnished user experience. Consequently, when Google debuted its search engine in 1998, many users, myself included, were drawn in by its simplicity, regardless of the actual quality of its search results. Google’s dominance not only obliterated competitors but also traditional print publishers, thanks to its engineers who managed to maintain computational efficiency, keeping information searches swift and inexpensive.
However, success also brought about a decline in user experience and effectiveness. As Google solidified its monopoly, search engines became inundated with irrelevant links due to the rise of search engine optimizers. Traditional media began utilizing clickbait to maintain viewer engagement. This overflow forced users to develop new skills, including keyword selection and the use of Boolean operators (like “AND”) while also learning to identify clickbait. Moreover, Google worsened the user experience by placing “sponsored” results at the forefront of search results.
Similarly, AI chatbots have embarked on a comparable trajectory. ChatGPT attracted five million users within just five days of its November 2022 launch. The chatbot’s intuitive interface allowed users to pose questions in plain English, without needing Boolean operators. Regrettably, ChatGPT turned out to be a deceptive tool, requiring more time to filter its inaccuracies than to conduct keyword searches on Google, ultimately decreasing productivity.
Now, several years later, Google has retired Bard, replacing it with an “AI overview” positioned above traditional search results and an AI Mode that encourages users to “ask anything.” Unfortunately, from my experience, the reliability of these results remains frustratingly low, even for straightforward inquiries.
When compared to earlier statistical models, such as Google’s pioneering algorithms, large language models (LLMs) seem to exhibit enticing advantages. With trillions of parameters, LLMs can integrate contextual elements that prior models overlooked. Additionally, LLMs don’t simply provide information; they weave metaphors and humor into their responses, engaging in what the philosopher Ludwig Wittgenstein termed “language-games” that can reassure, explain, flatter, assert, and persuade.
Nevertheless, like their predecessors, LLMs rely on statistical extrapolation based on past patterns. While this approach is suitable for natural phenomena—like protein folding—it falls short for the continuously evolving landscape of goods and services. Even minor adjustments to the design of a laptop battery can render replacement instructions obsolete. However, users often unthinkingly accept a chatbot’s confident responses, similar to participants in Stanley Milgram’s controversial shock experiments. In contrast, traditional search results better convey the outdatedness of information and the unreliability of sources, providing users with the freedom to exercise their judgment, thereby reducing errors and costs.
Including trillions of parameters amplifies LLMs’ extrapolation issues, as it heightens the chances of identifying non-existent patterns or irrelevant, unreliable answers from extensive, uncurated databases. Designers of traditional statistical frameworks can limit the variables and data sources to circumvent such pitfalls, all while managing computational expenditures.
The dependence of LLMs on statistical extrapolation further complicates their conversational interfaces. Since LLMs lack true contextual comprehension and cannot replicate human cognitive processes, their language use becomes a mere imitation rather than a genuine “form of life,” to echo Wittgenstein again. The false belief that they understand context allows AI marketers to promote chatbots broadly, especially since the narrower applications for which LLMs are authentically beneficial cannot justify the enormous investments involved.
Advocates of AI are determined to cultivate dependency among users. Google’s AI overviews and AI Mode serve as a trial experience. Once users are hooked, major tech players will inevitably impose high fees to compensate for their considerable development and operational expenses.
While it’s easy to hope that even the most influential companies cannot deceive consumers indefinitely, history teaches us otherwise. Google has experienced as many failures as successes, and Meta’s legacy serves as a reminder of its multi-billion-dollar miscalculation in the sphere of virtual reality. However, the marketing surge that has transformed junk food into a thriving industry presents a sobering counterargument.
Tech giants could easily make LLM dependency as prevalent as physical addictions by leveraging users’ innate desire for social interaction. As social connections weaken, chatbots offer a superficial substitute, potentially transforming even dedicated professionals into individuals enamored with these tools. Moreover, similar to other forms of dependency, LLMs are currently claiming essential resources—capital, electricity, chips, and entrepreneurial initiative—that could otherwise support more deserving innovations. As this trend persists, proponents of technology should reconsider their optimistic view regarding AI’s potential to enhance productivity.
