We Need to Talk about AI and Job Loss: On Jevons and Moravec And the Value of Nuanced AGI Studies

AI systems are often discussed as if they will wipe out jobs wholesale once “artificial general intelligence” arrives. The central claim here is more cautious and more specific: even if AI becomes widely capable, job loss projections that assume a simple, uniform replacement story miss two economic/technical realities—Jevons’ Paradox and Moravec’s Paradox—that can keep demand for human work from collapsing and can make “easy for humans” tasks disproportionately hard for machines.

The argument starts by tightening the definition of artificial general intelligence (AGI). By the standard used, AGI would be broadly deployed and able to perform “almost all valuable work that humans do,” including physical services like yard work and car repair. Under that definition, today’s chatbots are far from AGI. Even if the bar is lowered to knowledge work—AI doing most economically valuable cognitive tasks—the common “doomers vs. panicked comments” narrative still predicts widespread unemployment and argues for universal basic income or token-based compensation schemes.

Instead, the discussion challenges the way labor-market studies are built. A key critique is that many forecasts treat AGI as a single plug-in variable—one commodity-like input—rather than a technology that arrives with uneven capability and a “ragged edge” across job families. Two deeper omissions then take center stage.

First is Jevons’ Paradox: when a resource becomes cheaper or more abundant, demand can rise rather than flatten. The example is coal in the 19th century—more coal didn’t just make existing use saturate; it unlocked new uses. The same pattern is illustrated with the internet (from an early “coffee maker monitoring” concept to countless new applications) and with renewable energy, where production keeps expanding despite repeated expectations that growth would taper. The claim is that AI will similarly become more useful as it gets cheaper and more available, creating new tasks that humans and organizations would not have done otherwise. Anecdotally, the speaker describes using chatbots at work and in personal life for tasks that would otherwise never get done—an “in practice” version of Jevons’ Paradox that many labor projections fail to model.

Second is Moravec’s Paradox: it’s often easy to teach machines what humans find difficult, while tasks humans find easy—like walking, catching a ball, or navigating social cues—are much harder to replicate reliably. The discussion uses classic game-playing as the “easy for machines” side (Deep Blue beating Gary Kasparov, then machines surpassing at chess and nearly solving go). For knowledge work, the “hard for machines” side is framed as internal human dynamics: negotiation, stakeholder management, timing, and aligning constraints across sales environments. Even if AI can outperform a human on factual recall in product management, it may still fail at the broader conversational and coordination work that depends on context, judgment, and agentic prototyping.

Together, the paradoxes support a more optimistic view of job disruption: AI may change tasks and workflows, but it doesn’t automatically eliminate economically valuable work. The closing note points to a small signal that even major AI organizations may not be planning for an immediate end to employment—OpenAI reportedly adjusted stock-grant policies so departing employees can continue receiving value tied to tenure and vesting, implying a longer horizon for normal employment patterns.

The takeaway is not that AI won’t displace workers, but that forecasting needs more nuance: demand effects (Jevons) and capability asymmetries (Moravec) should be treated as central variables, not afterthoughts.