A Gap In AI Adoption? Moravec And The AI Productivity Paradox

Businessman in a state of confusion as artificial intelligence takes over his job roles. Symbolises the challenges and adjustments in the modern workforce due to technological advancements

getty

A new report in the Wall Street Journal (WSJ) seems to highlight an AI productivity paradox where executives in the C-suite say AI Is making work efficient, while employees are less likely to say so. This highlights a fundamental observation from AI research dubbed Moravec’s Paradox. The paradox, summarized by Hans Moravec, states:
“It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”

Moravec’s paradox points to an inverse association between human and AI proficiency in cognition. This means that, tasks that humans perform intuitively and instantly, such as causal reasoning, can stump the most advanced AI model. In professional contexts, this means AI can often out-calculate a CEO but struggles to match the situational awareness of a frontline worker. We may be witnessing a “Management Information Paradox” wherein AI can be deployed to draft high-level strategic memos more easily than ensure consistency, reliability and auditability of a junior analyst, which requires error-free data processing. The reported gap, where 40% of executives save 8+ hours while 66% of workers save almost none, isn’t just a matter of training; it’s a reflection of the type of work each group performs.

“High-Level” vs. “Contextual” Logic and the AI Productivity Paradox

Moravec’s Paradox posits that abstract reasoning (logic, math, high-level strategy) is “easy” for AI because it is evolutionarily new and follows clear rules. For a senior level executive, day-to-day tasks often involve synthesizing long reports, drafting strategic emails, or simulating market scenarios. These are hard for humans but “easy” for LLMs because they exist entirely in the realm of symbolic logic and language. An executive can prompt an AI to “summarize these 50 pages of market research into three strategic pillars.” The AI performs this “intelligent” task in seconds, reclaiming hours of high-level cognitive labor. Non-management roles, by contrast, often involve “grounded” work: data entry that requires checking for real-world accuracy, coordinating between conflicting human schedules, or “fixing” minor errors in a process. The WSJ report highlights an AI Tax: workers often spend their “saved” time correcting AI errors. Because the AI lacks “common sense” (the “easy” part of the paradox), it might produce a perfect-looking spreadsheet that is factually hallucinated. The worker then must do the “hard” task of manual verification—a perceptual and contextual task the AI can’t help with. There is also the delegation dimension wherein executives could use GenAI to delegate the “thinking” part of a task downward. If an executive uses AI to generate a rough project plan in 5 minutes (saving them 2 hours), they then send it to a staffer to “finalize.” The staffer then spends 2 hours fixing the AI’s “hallucinations” and formatting errors, which require the very perceptual “common sense” that AI lacks. The executive sees the time saved; the worker absorbs the extra labor required to make the AI’s output usable.

This has broad implications for AI adoption. For example, in the world of high-stakes finance, AI excels at processing millions of data points to execute arbitrage trades or predict quarterly earnings (abstract mathematics). A hedge fund manager may need to deal with a black swan event such as how to diversify a major investment during a sudden geopolitical crisis. While AI is brilliant at “hard” tasks like calculating Value-at-Risk, it cannot easily distinguish between a “routine” political protest and a “regime-changing” revolution by watching a news feed, whereas a human can often “sense” a shift in the cultural zeitgeist or “read between the lines” of a leader’s body language in a press conference.

The J-Curve and the AI Productivity Paradox

The Productivity J-Curve – a concept originating from research by Professors Erik Brynjolfsson, Daniel Rock and Chad Syverson, could help explain this disconnect.
The “J-curve” describes a phenomenon where a new technology initially causes a dip in productivity (the curve’s bottom) before eventually leading to a massive surge (the upward stroke of the J). The reported gap could be that executives have already hit the upward stroke, while workers are still stuck in the “dip.” For executives, the J-curve is shallow or non-existent. Because their work is high-level and symbolic (the “hard” logic AI is good at), they can use AI as a “plug-and-play” tool. Executives may then see immediate gains from AI adoption. An executive can use AI to draft a speech or analyze a strategy memo in seconds. This requires almost no “intangible investment” (reorganizing workflows or deep training). They are harvesting the benefits of the technology immediately without the messy middle.

For staffers, by contrast, the J-curve is deep and painful. Their work is grounded in execution and “common sense” (the “easy” things AI is bad at, per Moravec). To make AI productive at the worker level, the company must invest in intangible capital as well as ensure incentives for rewriting standard operating procedures, cleaning data, and training staff. Workers then spend their time fixing AI mistakes (the Moravec Tax) and in trying to integrate a tool that doesn’t yet fit their complex, real-world tasks. This unmeasured investment looks like a productivity loss on paper because they are working harder just to maintain their previous output.

A “J-Curve” Depth Gauge for the AI Productivity Paradox

Frozen mircochip inside an iceberg. AI winter, artificiall intelligence pessimism concept. Vector illustration.

getty

We can employ Moravec’s Paradox to predict how deep a role’s J-curve will be:

• Tasks With Low Moravec Sensitivity (Executive Roles): Work is mostly abstract reasoning. For roles that involve synthesizing information to make high-level decisions, the J-curve is shallow. These roles live in the “hard for humans, easy for AI” zone of Moravec’s paradox.
• The Moderate J-Curve (“Technical Expert” Roles): Roles that are currently in a “Jagged Frontier” of AI adoption. Developers could spend more time in debugging AI-generated code that lacks “common sense”, or dealing with AI workslop.
• High Moravec Sensitivity (Worker Roles): Work requires perception, physical coordination, or high-stakes accuracy. This results in a deep, long J-curve where things get worse (more work, less time saved) before they get better.

The vast gulf in the returns from AI is not a sign that the AI productivity paradox is impossible to solve; it is a sign that organizational redesign hasn’t happened yet. Until companies move past the flashy demo phase and solve the Moravec problems at the frontline (improving AI’s reliability and contextual awareness), workers will remain trapped in the bottom of the J-curve, doing the invisible labor that makes the executives look productive.