Investment in artificial intelligence will continue its meteoric rise in 2025, powered by a narrative that AI is not just a tool but a new workforce. As Jensen Huang of NVIDIA recently put it, the software industry of the past was about creating tools like Excel, Word and web browsers, "tools… only so large" because humans had to operate them. By contrast, "AI is not a tool. AI is, in fact, workers that can actually use tools."

Imagine AI systems as digital colleagues who write code in VS Code, plan trips through browsers or act as invisible chauffeurs in robotaxis. This framing has justified capital expenditures on a scale usually reserved for national infrastructure. But before we accept this narrative, it is worth revisiting the deeper historical and economic forces at play. The distinction between a tool and a worker is not merely philosophical. It underpins the assumptions that drive trillion-dollar forecasts and the fears of a potential valuation bubble.

The modern tech industry's enthusiasm rests on this "AI is work" thesis: that AI systems will increasingly perform tasks in place of humans, not just support them. If AI is a workforce, then its addressable market is the value of human labor itself.

There is a rival framing, however, rooted in Steve Jobs' idea of the computer as a "bicycle for the mind". In that tradition, digital technologies are tools that amplify human capabilities instead of replacing them. The central economic question is which description of AI will turn out to be true in practice: is it primarily a new class of workers, as Huang suggests, or a more powerful generation of tools in the lineage of the bicycle for the mind?

Understanding which frame is correct matters because financial markets are acting on the assumption that AI is work. Capital flows into GPUs, data centers and model development are exponential. AI-related startups are hitting billion-dollar valuations even before generating revenue. Forecasts of future income resemble science fiction more than standard industry projections.

History offers a useful reminder. Every general purpose technology has attracted massive speculation, followed by a painful correction when the timing of returns failed to match investor optimism. In each case, as we pointed out in our recent book, Power and Prediction: The Disruptive Economics of Artificial Intelligence, the underlying technology proved transformative, but adoption took far longer than expected.

AI today has all the hallmarks of a normal transformative technology: it makes a scarce input (prediction) abundant; it requires system-wide reorganizations to achieve productivity gains; its applications are uncertain; and its diffusion depends on people learning how to use it effectively.

This last point is critical. AI systems today operate with a form of "jagged intelligence": they excel in some tasks and fail unexpectedly in others. The difficulty is not merely technical performance; it is that we still lack a shared understanding of what these systems are reliably good for. This uncertainty slows adoption and forces humans to remain central to all high-stakes decisions. Far from acting as independent workers, AI systems depend on human judgment to determine objectives, evaluate outputs and verify correctness.

Judgment, in this context, is not a vague leftover from human uniqueness. It is the essential input that determines how prediction becomes action. Someone must decide what errors are tolerable, what trade-offs matter and whether an AI-generated output should be accepted or rejected. That someone may be a consumer, a frontline worker or a committee of engineers embedding these decisions into products. But it is always a person. As long as that remains true, AI behaves more like a tool requiring skilled operators than like a worker performing tasks autonomously.

Recognizing AI as a tool does not diminish its importance. Tools can radically reshape industries and societies. But it provides a different benchmark for valuing AI companies. The historical revenues of successful toolmakers — from software firms to industrial equipment manufacturers — are orders of magnitude lower than the revenue one would expect if AI were truly replacing labor. The market's current projections implicitly assume the latter, even though evidence overwhelmingly points to the former.

These dynamics shape the risk of a potential valuation bubble by 2026. If investors continue to price AI companies as if they will capture the economic value of labor rather than the economic value of tools, valuations will outrun reality. Meanwhile, adoption curves are likely to follow historical patterns: slower, uneven and dependent on complementary innovations across organizations.

AI's impact on labor markets also illustrates the tool-versus-work divide. Early data shows not widespread displacement but a shift in the value of experience. Workers with strong judgment and domain expertise appear to become more valuable in AI-exposed occupations, while entry-level workers may find it harder to compete. This suggests complementarity, not substitution: AI increases the productivity of experienced workers, reinforcing the need for human oversight rather than eliminating it.

These realities introduce tensions between commercialization and regulation. Firms want to automate aggressively, both to capture cost savings and to justify their investment levels. Regulators, however, must consider the systemic risks of premature automation: misaligned judgment, unverifiable outputs and overreliance on technologies whose capabilities remain poorly understood. The pressure to deploy AI as "work" may collide with the evidence that AI still requires substantial human governance to function safely.

Investment enthusiasm has remained high in 2025. But a clear-eyed reading of the economics suggests a more measured future. AI is powerful, expanding the frontier of prediction and enabling new forms of digital production. Yet its value depends on people: their judgment, their adaptation and their capacity to integrate tools into meaningful systems of work. The risk of a 2026 correction is real if markets continue to conflate tools with workers. The more productive path forward is to recognize AI's strengths, acknowledge its limits and invest accordingly in the human capabilities that make any transformative technology truly valuable.

The author is the Jeffrey S. Skoll Chair of Technical Innovation and Entrepreneurship and professor of Strategic Management at the Rotman School of Management of the University of Toronto.

The views don't necessarily reflect those of China Daily.