U.S. Federal Reserve Chair Kevin Warsh and several prominent economists have expressed optimism that artificial intelligence (AI) will significantly boost labor productivity. However, some experts caution that the widespread adoption of AI technologies, particularly large language models, may actually diminish output per worker.

Historically, the integration of new technologies has driven economic dynamism. The internet era offers a clear example: early systems like Gopher relied on complex, text-based methods for document sharing. This evolved into the more user-friendly World Wide Web with browsers like Mosaic, Netscape Navigator, and Microsoft Internet Explorer, alongside efficient search engines. Yet, the success of search engines also led to challenges; as companies like Google gained dominance, results became cluttered with low-quality links, and traditional media pivoted toward clickbait to draw traffic, ultimately undermining productivity and user experience.

A similar trajectory has occurred in the AI chatbot space. ChatGPT, launched in November 2022, attracted a million users within its first five days, thanks to an accessible interface that allowed users to ask questions in everyday language. Nevertheless, the tool often produced inaccurate or fabricated information, requiring users to spend additional time verifying answers—sometimes more than they would have with conventional keyword searches. Google's own response, Bard, released amid a strategic bid to protect its search market share, faced similar shortcomings and was eventually discontinued in favor of new features like an “AI overview” and an “AI Mode” embedded in traditional search results. Early user feedback indicates that these replacements still struggle with reliability, even for straightforward queries.

While large language models benefit from vast parameter sets—trillions, compared to earlier statistical algorithms—allowing greater incorporation of contextual factors, they rely heavily on statistical extrapolation drawn from past data patterns. This approach works effectively for relatively stable natural phenomena, such as protein folding, but is less suited for the dynamic and evolving nature of goods, services, and human knowledge. The confident presentation of answers by AI chatbots often leads users to accept information uncritically, unlike traditional search results that provide multiple sources and contextual cues, helping users to better assess reliability.

This extensive parameterization can exacerbate inaccuracies by increasing the likelihood of identifying spurious patterns or selecting irrelevant data from vast and uncurated datasets. In contrast, designers of earlier models could limit variables and data sources to improve accuracy and manage computational efficiency. Moreover, these AI models lack true semantic understanding or authentic human sense-making capacities; their conversational style is a linguistic imitation rather than genuine comprehension. The portrayal of AI chatbots as sentient conversational partners has fueled aggressive marketing and broad deployment, despite large language models being truly effective only in specific, narrower applications.

Industry leaders appear to be encouraging user dependency on AI tools, initially offering free trials such as Google’s AI features, with plans to monetize access later to offset substantial development and operational costs. Experts warn that the human tendency toward addictive behaviors, especially in the context of increasingly isolated social environments, may make users susceptible to overreliance on AI chatbots for companionship, potentially distracting professionals and diverting critical resources—including capital, power consumption, hardware, and entrepreneurial focus—from other innovative endeavors.

As these trends continue, some commentators urge a reassessment of overly optimistic predictions regarding AI’s impact on labor productivity, highlighting the potential economic and social costs of unchecked AI adoption.