We’ve Mispriced Thinking

AI could end the oldest trade-off in economics. So far, it hasn't.

Adam Smith’s pin factory is probably the most famous passage in economics. He studied a workshop where ten workers, each performing one small step, produced thousands of pins a day. While one worker alone might make just a handful of pins, by dividing the work into narrow, repetitive tasks the output multiplied. Then, hundreds of pages later, Smith names the cost: the worker who spends his life on a few simple operations becomes, in his words, “as stupid and ignorant as it is possible for a human creature to become.” He’s left incapable of rational conversation, of generous sentiment, of forming just judgment about the ordinary duties of private life.¹

This was prosperity, purchased at the cost of the laborer’s humanity. Smith saw no way around this. He pitied the worker, but recognized this as the sacrifice that was required.

More recently, what we call knowledge work has operated with the same trade-off. 60% of the average knowledge worker’s day goes to a narrow set of mind-numbing tasks: searching for information, chasing updates, switching between tools that do not talk to each other.² The simple operations of the knowledge work era. The cognitive pin factory.

Now, AI is different. Rather than simply divide the work as most previous technologies did, it could take over the mindless labor altogether. The time that went to the cognitive equivalent of making pins could go to the faculties Smith said the factory destroys: understanding, invention, the capacity for just judgment. For the first time since Smith, we have a technology that could deliver prosperity without requiring the human sacrifice. Smith would have been amazed.

But so far, AI is not breaking the trade-off. Workers who have offloaded routine tasks to AI are reporting a new kind of cognitive exhaustion, driven not by drudge work but by the cognitive load of evaluating output they did not produce and are now accountable for.³ One form of overload, replaced by another.

But the overload is not the real risk. The real risk is what happens when AI goes beyond the routine tasks into the hard human work that Smith called out: judgment, understanding, invention. The temptation to hand that over too is enormous. And at twenty dollars a month, there is nothing in the price to stop you.

Users pay twenty dollars a month. A single professional user recently ran up a $150,000 monthly bill. An engineer’s weekly token usage cost tens of thousands. And even at those prices, the AI providers are losing money.⁴ The cost of producing AI’s cognitive labor is rising, but the price most people see is artificially low, held there by investor subsidy.

Smith called this the gap between natural price and market price. At the natural price, there would be friction, a reason to weigh what to delegate and what to keep. But the subsidy distorts that signal. At twenty dollars a month, you do not pause to consider what is worth handing over, and AI arrived cheaply, everywhere, adopted without anyone deciding what they wanted it to do for them.

And at the economy-wide level, AI has not yet moved the needle.⁵ Most enterprise projects stall before reaching production.⁶ What would move the needle is something AI cannot produce on its own: ingenuity, the judgment to know what is worth building. The more capable the human, the more the technology tends to deliver. The more the capacity erodes, the less any of it is worth. But nothing in the current system tracks whether the human capacity that both the market and the person depend on is growing or eroding.

The AI companies that build the models have a disciplined way of measuring performance. In their parlance these are called evals: evaluations of performance against a goal. In AI development, the eval determines the behavior. The choice of what to measure is a decision about what matters. There are evals for the model’s capability, for coding speed, for benchmark performance.⁷ A few researchers have started developing evals for the human side: whether AI preserves agency, deepens understanding, supports flourishing.⁸ But the organizations actually deploying AI, the ones running the leaderboards and the performance reviews, have no consistent way of measuring it. They are measuring machine throughput: how much AI was consumed, with no visibility into what it produced, what it cost, or what it did to the capacity of the person consuming it.⁹

We need an eval for whether the promise is being delivered. Not how much was produced, but whether AI helped the person get to an insight faster, and whether that insight changed something that mattered. Call it time to insight.

Smith accepted the sacrifice because he could not imagine prosperity without it. We can. We have the technology. We do not yet have the eval.

---

Footnotes:

¹ Smith, The Wealth of Nations, Book V, Ch. 1. ² Asana Anatomy of Work Index, 10k+ workers. ³ Bedard et al., “When Using AI Leads to Brain Fry,” HBR, March 2026, 1,488 workers. ⁴ Epoch AI: training costs growing 2.4x/year. Fortune: OpenAI financials. Roose, NYT: single Claude Code user exceeded $150k/month. ⁵ NBER/Atlanta Fed survey, 750+ CFOs. Fortune, 6,000 CEO study. ⁶ MIT NANDA, “The GenAI Divide”: 95% of organizations seeing zero return despite $30-40B in enterprise investment. ⁷ GitHub Copilot adopted by 90% of Fortune 100; coding productivity studies show double-digit improvements. ⁸ MIT Media Lab AHA program, “Benchmarks for Human Flourishing with AI” workshop, October 2025. Six dimensions including Comprehension & Agency. ⁹ Roose, NYT, March 2026. Hard Fork, March 20, 2026: token leaderboards and performance reviews based on AI consumption at Meta, OpenAI, Shopify.