A World of Paperclips: The Systematic Erasure of Economic Truth

We built instruments to measure a world that no longer exists. Now we're flying blind while arguing about the altitude.

There's a particular kind of institutional blindness that only reveals itself after catastrophic decisions have already been made. The instruments looked fine. The readings were consistent. Everyone agreed on what the data said. The problem was that nobody asked whether the instruments were still measuring what they were built to measure.

That's where we are with AI and the economy right now.

The conversation happening in op-eds, Fed press conferences, and investor research notes goes something like this: AI is driving extraordinary growth, capital expenditure is surging, productivity is improving, and yet the job market is softening—layoffs in tech, slower hiring across white-collar sectors, a peculiar anxiety among workers who haven't lost their jobs yet but feel like they should have. The mainstream interpretation oscillates between two familiar narratives: either AI is delivering on its promise and creative destruction is underway, or the investment bubble is about to burst and we'll all feel foolish for believing. Both camps are fighting over conclusions without questioning the premises.

The premises are the interesting part.

What "Growth" Actually Means When the Numbers Are Broken

GDP, the number everyone uses to assert that things are fine or not fine, is a measure of spending. Specifically, it measures the market value of goods and services produced. This is a reasonable proxy for economic activity when the composition of that activity is relatively stable—when the ratio of investment to consumption, of productive spending to speculative accumulation, stays within historical ranges.

It is not a reasonable proxy when a single industry begins absorbing capital at unprecedented velocity for outputs whose economic value is almost entirely speculative and whose productivity effects, if they materialize at all, will arrive asymmetrically, unevenly, and probably later than anyone's current model assumes.

AI infrastructure investment—data centers, chips, energy capacity, networking—is now running at a scale that meaningfully moves aggregate GDP numbers. Nvidia's revenue alone is restructuring how we read national accounts. When a hyperscaler spends $50 billion building data centers, that spending shows up as investment, which flows into GDP. The implicit assumption embedded in GDP accounting is that investment spending represents the creation of productive capacity—that the expenditure is justified by expected future returns that are, at minimum, plausible.

This assumption is doing an enormous amount of quiet work right now.

Because if a significant fraction of AI infrastructure investment is malinvestment—capital deployed in pursuit of a capability curve that plateaus earlier than projected, or for use cases that never achieve the revenue density required to justify the build-out—then current GDP figures are partly measuring the economic equivalent of digging holes and filling them in. Not because the holes are stupid on their face. Because the expectation embedded in the measurement is that they'll produce something. When that expectation fails, you don't just lose the future productivity gain. You've already counted it once, as current growth.

This is how investment bubbles interact with GDP in ways that create accounting flattery during inflation and genuine pain during deflation. The 2000s housing bubble showed this clearly. Construction spending, mortgage origination, derivative complexity—all of it registered as economic activity, as growth, as evidence of a healthy system. Until it registered as something else entirely.

The uncomfortable question isn't whether AI investment might be partly speculative. Of course it is—all investment is partly speculative. The question is whether our measurement systems are capable of distinguishing between speculative investment that eventually grounds in productive capacity and speculative investment that doesn't. They aren't. They never have been. We've just collectively agreed to defer that question until after the fact.

The Job Market Signal Is Also Broken, But Differently

If AI is making growth look artificially strong, it's making the labor market look artificially weak—and the mechanism is almost the opposite.

Here's the dynamic that's underappreciated: when companies announce that AI tools are increasing worker productivity, they often do so in the context of justifying headcount freezes or reductions. The CEO says "our people are doing more with less thanks to AI" and the financial press transcribes it faithfully. What gets analyzed far less carefully is whether the productivity claim is real, whether the headcount reduction is causally connected to AI adoption, or whether "AI is allowing us to do more with less" is simply the current vintage of the eternal corporate rationalization for workforce reduction.

Companies have always needed narratives for layoffs that don't sound like "we over-hired, the economy slowed, and we need to reduce costs." The available narratives cycle through history with remarkable consistency: restructuring, right-sizing, operational excellence, automation. AI is the new automation, with the added advantage that it sounds genuinely transformative rather than merely operational. It's a better story. It carries more explanatory weight. It lets management sound like they're steering toward something rather than retreating from something.

The behavioral incentive here is sharp: AI provides a socially prestigious explanation for cost-cutting behavior that would otherwise require uncomfortable acknowledgment of prior overcapacity. The tech sector massively over-hired between 2019 and 2022. The correction was inevitable regardless of what happened with AI. But "we are strategically realigning our workforce around AI capabilities" is a substantially more attractive narrative than "we hired 40% too many people during a period of zero-interest-rate-fueled revenue optimism."

So the job market signal gets distorted in two directions simultaneously. Actual employment effects of AI—which are real, uneven, and probably understated in official statistics because they manifest as hiring slowdowns rather than visible layoffs—are being attributed to a cleaner AI adoption narrative than the underlying reality supports. And AI adoption itself is being credited with productivity gains that are, in many cases, not yet measurable in output, only in headcount.

The result is a labor market that looks weaker than it structurally is (because the tech over-hiring correction inflates visible distress) while simultaneously looking more AI-transformed than it actually is (because the AI narrative absorbs the explanation for what is actually cyclical adjustment).

The Capital Allocation Problem Nobody Is Pricing

There is a version of the AI investment story that works. Compute infrastructure gets built, model capability scales, applications proliferate, revenue follows, returns materialize. Some meaningful fraction of the current capex boom reflects rational anticipation of this path. The hyperscalers aren't stupid. The chip companies aren't stupid. A lot of people have thought carefully about these bets.

But there's a subtle systems problem embedded in how the capital is allocated that doesn't require anyone to be stupid to produce a bad outcome.

The capital flowing into AI infrastructure is concentrated. Four or five companies are spending the majority of it. Their incentives are not primarily to generate the highest risk-adjusted return on that capital. Their incentives are to not be left behind. This is strategically rational at the individual firm level and potentially irrational at the system level—the classic coordination problem where everyone defecting from restraint produces outcomes nobody would have chosen collectively.

When the primary driver of investment is competitive anxiety rather than positive expected value calculation, you get overbuilding. Not because anyone made a bad decision by their own internal logic. Because the game theory of platform competition creates systematic pressure toward excess that only resolves when someone flinches—or when the market provides a sufficiently dramatic signal that the music has stopped.

The additional complexity is that AI infrastructure has unusual economic properties. Data centers depreciate. Energy contracts extend. Talent is expensive and sticky. The optionality of "wait and see" is largely unavailable to companies for whom AI capability is a survival-level competitive concern. So even a hyperscaler who privately doubts the near-term revenue math of a particular infrastructure build will often proceed anyway, because the cost of being wrong about AI matters less than the cost of being caught without capacity when it turns out to matter.

This creates a system where rational individual behavior generates collectively irrational outcomes, where the investment boom is partially self-validating (more investment → more capability → more pressure on competitors to invest), and where the bust, if it comes, arrives suddenly rather than gradually—because the feedback loops that would normally slow investment (disappointing returns, rising cost of capital, softening demand signals) are masked by competitive dynamics and accounting conventions that defer recognition of underperformance.

Why a Bust Might Hurt Less Than Expected (For the Wrong Reasons)

Here's the counterintuitive proposition buried in all of this: if there is an AI investment correction, it might cause less macroeconomic pain than historical investment busts—not because AI is uniquely resilient, but because a meaningful portion of current "AI growth" was never real in the sense of representing sustainable economic activity in the first place.

When a bubble deflates, you lose two things: the speculative premium on future growth expectations, and the real economic activity that was supported by bubble-era spending. The first produces asset price pain. The second produces employment pain, supply chain contraction, credit stress. The 2008 housing correction hurt so badly partly because the speculative premium was enormous, but also because the real economic activity—construction employment, mortgage origination, real estate services—was genuinely large and embedded in communities that had organized around it.

AI's real economic activity is narrower and more concentrated. The workers building data centers are real, but there aren't that many of them relative to the capital being deployed. The chip manufacturing employment is largely offshore. The software engineers whose salaries are inflated by AI demand are a small and economically unusual population—high-earning, geographically concentrated, and for the most part not the marginal economic actors whose distress transmits most severely through the broader consumer economy.

This means an AI investment correction would primarily show up in: equity valuations (very visible, very emotionally salient), corporate capex budgets (which would revert, causing real but concentrated pain), and tech employment (which is already experiencing its correction through the hiring slowdown).

What it would not do, at least not in the same way housing did, is destroy the balance sheets of ordinary households, collapse a credit system with tentacles in every community bank, or eliminate employment categories that take a generation to reconstitute.

The irony is that the very features of AI investment that make the current GDP readings artificially flattering—its concentration in a small number of firms, its capital-intensity relative to employment intensity, its distance from the retail credit system—are the same features that would limit contagion in a correction scenario.

We might be overestimating AI's contribution to durable growth while simultaneously overestimating how badly its failure would hurt.

The Measurement System Is the Message

What's actually happening here is a collision between measurement infrastructure built for one kind of economy and economic activity that has taken on forms those instruments weren't designed to capture.

GDP was refined in an era when the relationship between investment and productive output was relatively legible. You built a factory; you made things; you sold things. The feedback loop between investment and output was tight enough that investment spending was a reasonable proxy for capacity building.

That relationship has always been more complicated than the textbook. But AI represents a genuine break in the signal. The investment is in computational capacity whose productivity effects are conditional on use cases, business model innovation, and organizational change that may or may not materialize. The employment effects are both direct (the sector itself) and indirect (what AI tools do to work across every other sector), and the indirect effects are nearly impossible to measure cleanly because they manifest as changes in productivity rather than changes in headcount.

So you end up with a situation where the most sophisticated economy in human history is being navigated using instruments that are simultaneously overclaiming growth (because investment in uncertain futures gets counted as present output) and overclaiming disruption (because AI provides a convenient narrative cover for labor market changes that have multiple causes).

And the people making decisions—the Fed officials, the investors, the corporate planners, the policy architects—are not stupid. Most of them know the numbers are noisy. But they're bounded by the institutional necessity of making decisions with the instruments available, which creates pressure to treat imperfect signals as clean ones, to coordinate around consensus readings even when the underlying uncertainty is much higher than the consensus acknowledges.

The real risk isn't that AI will fail or succeed. The real risk is that we won't be able to tell which one happened until the evidence has been accumulating for years—and that in the meantime, significant decisions about monetary policy, capital allocation, and workforce strategy will be made on the basis of numbers that are telling two different lies simultaneously.

History is full of systems that looked coherent from inside the measurement framework and only revealed their distortions after the fact. We have no particular reason to believe we're immune to that problem now. We have some reason to believe AI, specifically, has made it worse.

That's not a prediction. It's a description of the epistemological situation we're actually in—which is considerably more uncomfortable than either the AI optimists or the AI skeptics tend to acknowledge, because it doesn't resolve cleanly into either narrative.

The instruments are broken in different directions at once. Everything looks like something. Very little can be trusted to mean what it appears to mean.

Welcome to flying by feel.