The Fed Meets Tuesday to Set Interest Rates. Its Models Can't See 10 Million Displaced Workers.

On Tuesday, March 18, the Federal Open Market Committee will announce its interest rate decision. Twelve people around a table in the Eccles Building will weigh inflation data, employment numbers, and GDP growth to determine the price of money for 330 million Americans.

They will do this using analytical models that cannot detect the single largest structural change in the American labor market since deindustrialization.

This is not a metaphor. The Federal Reserve's core frameworks — the Phillips curve, NAIRU, the Beveridge curve, the Taylor Rule — contain no variable for AI-driven structural displacement. They can see that unemployment rose to 4.4 percent in February. They cannot see why. And the why is everything.

The 0.47 Correlation Nobody Can Prove

In August 2025, the St. Louis Fed published a remarkable finding: occupations with higher AI exposure experienced larger unemployment increases between 2022 and 2025, with a 0.47 correlation coefficient. Computer and mathematical occupations — the most AI-exposed — showed the steepest unemployment rises.

When they used actual AI adoption data instead of theoretical exposure, the correlation was even stronger: 0.57.

The authors' caveat was equally remarkable: "The findings presented here represent correlation, not causation." The Fed's own researchers can see the signal. They cannot confirm it's AI rather than "economic uncertainty, postpandemic monetary policy tightening and chance timing."

This is the blind spot in three sentences. The Fed can measure unemployment. It can measure AI adoption. It can see the two move together. But its models have no mechanism to connect them — and without that connection, the rate-setting committee has no basis to treat the displacement as structural rather than cyclical.

The difference matters more than almost anything else in economic policy. If unemployment is cyclical — people temporarily out of work because demand fell — rate cuts fix it. Cheaper borrowing stimulates spending, businesses expand, workers get rehired. This is the playbook, and it's worked for 80 years.

If unemployment is structural — jobs permanently eliminated by technology — rate cuts don't fix it. Worse: cheaper capital makes AI investment cheaper, which accelerates the displacement that caused the unemployment in the first place. The medicine becomes the disease.

Five Instruments, All Broken

The Fed doesn't use just one model. It has a toolkit. Every instrument in it is distorted by AI displacement in a specific, identifiable way.

The unemployment rate (U-3). The headline number asks a binary question: are you employed or actively seeking work? Klarna eliminated 3,104 positions over two years through hiring freezes and natural attrition. Zero WARN Act triggers. Zero unemployment claims filed. Zero visibility in BLS data. The workers who would have filled those positions were never counted as displaced — they're "not in labor force" or "employed elsewhere." Stanford researchers found a 13 percent relative employment decline for 22-to-25-year-olds in AI-exposed occupations. But those workers never had the jobs. You can't be counted as displaced from a position that was never posted.

The Phillips curve. For decades, this was the master equation: low unemployment means tight labor market means rising wages means rising inflation. The Fed raises rates to cool it, or cuts to stimulate it. The relationship has been flattening since the 2000s. Paul Kedrosky's analysis documents the progression: steep in the 1960s-70s, moderate in the 1980s-90s, essentially flat now. The reason the standard explanations miss: when employers can credibly threaten to automate, wages don't rise even when unemployment is low. The curve flattens because "labor market tightness" no longer means what it meant.

NAIRU. The natural rate of unemployment — the rate consistent with stable inflation. The Fed estimates it using models that link actual unemployment, the natural rate, and inflation to their historical values. The Richmond Fed acknowledged the problem in August 2025: these models attribute most unemployment changes to cyclical factors, not structural ones. If AI permanently eliminates categories of employment, NAIRU shifts upward — but the models detect this shift only after years of data. During those years, policy is calibrated to an outdated target.

The Beveridge curve. Job openings versus unemployment — the Fed's gauge of labor market tightness. The St. Louis Fed discovered in November 2025 that the post-2010 surge in job vacancies is "almost entirely due to poaching vacancies" — firms trying to hire workers who already have jobs, not the unemployed. Ghost job listings, AI-generated postings, and skills-mismatch vacancies inflate the vacancy-to-unemployment ratio, making the labor market look tighter than it is.

GDP. An NYU researcher named Xupeng Chen formalized what may be the most important concept for understanding this moment. In a March 10, 2026 paper, he introduced "Ghost GDP": when AI-generated output substitutes for labor-generated output, nominal GDP circulates through fewer wage-earning agents. GDP looks fine. Corporate revenues look fine. But the consumption-relevant income — what workers actually have to spend — declines. The Fed sees "resilient GDP" and may hold rates high during what is, for workers, a severe contraction.

The Chairman Knows

Jerome Powell is not unaware of this. In June 2025, testifying before the Senate Banking Committee, he was asked about Anthropic CEO Dario Amodei's prediction that AI could erase half of all entry-level white-collar jobs.

Powell's response: "The Fed doesn't have the tools to really address the social issues and labor market issues that will arise from this."

He also said AI experts told the Fed that what's happening now is "nothing compared to what we can do in two years."

This is the chairman of the Federal Reserve acknowledging, under oath, that his institution cannot handle what's coming. It is an institutional admission of irrelevance on the defining economic question of the decade.

And then he set interest rates anyway.

Tuesday's Impossible Decision

Here is what the FOMC faces on March 18:

Oil is at $119 a barrel. The Strait of Hormuz crisis has disrupted 20 percent of global supply. Inflation is accelerating from energy costs. Every model says: hold rates, or raise them.

Meanwhile, February showed negative 92,000 jobs. Unemployment hit 4.4 percent. Long-term unemployment rose 27 percent year-over-year. Q1 tech layoffs hit 45,363 with one-fifth explicitly AI-attributed. The BLS benchmark revision revealed 2025 created 3.2 times fewer jobs than initially reported. Every labor indicator says: cut rates.

The two signals point in opposite directions. This is the classic stagflation trap — the scenario central bankers fear most. But it's worse than the 1970s version, because the unemployment component has a structural element that rate policy cannot reach.

If the Fed holds rates to fight oil inflation, it deepens both cyclical unemployment (demand contraction from high energy costs) and structural unemployment (expensive capital doesn't slow AI displacement — a $15,000-a-year AI agent is still 10 times cheaper than a $160,000-a-year knowledge worker even at higher electricity prices).

If the Fed cuts rates to address unemployment, it eases cyclical job losses but accelerates structural displacement. Cheaper capital means cheaper AI investment. The "displacement spiral" that Chen models — each firm's rational decision to substitute AI for labor reduces aggregate labor income, which reduces aggregate demand, which pressures other firms to automate — runs faster when money is cheap.

There is no rate that solves both problems, because one is cyclical and the other is structural, and the Fed's models cannot distinguish between them in the aggregate unemployment number.

The 3.2× Miscount

It's worth dwelling on the BLS benchmark revision, because it illustrates how deeply the measurement problem runs.

Throughout 2025, the Bureau of Labor Statistics reported the economy was creating hundreds of thousands of jobs per month. Politicians cited the numbers. The Fed used them. Markets moved on them. Then the annual benchmark revision revealed the economy had actually created 181,000 jobs — not 584,000. The overcount was 3.2 times.

The primary culprit: the birth-death model, which estimates new business formation. The model assumes that when existing businesses grow, new businesses are also forming and hiring. But what if existing businesses are growing without hiring — because AI lets them handle more revenue with fewer workers? The Klarna pattern: revenue up 108 percent, headcount down 52 percent. The birth-death model counted phantom employees at phantom firms.

The Fed made rate decisions all year based on a labor market 3.2 times stronger than reality. And nobody knew until the revision.

What Would Actually Help

The Fed could, in theory, build better instruments. An AI-adjusted NAIRU that incorporates structural displacement. Distributional employment indicators that track entry-level versus experienced, knowledge work versus physical, augmented versus eliminated. A Ghost GDP metric that separates labor-income-supported growth from AI-productivity-driven growth.

But these require something that doesn't exist: reliable, real-time data on AI displacement at the occupation and firm level. This is exactly what the Warner-Hawley AI-Related Job Impacts Clarity Act would provide — quarterly company reporting on AI's employment effects to the Department of Labor. It's been framed as a labor policy bill. It's actually a monetary policy necessity. Without it, the Fed is setting the price of money for 330 million people using models that cannot see the defining economic transformation of their lifetimes.

Warner-Hawley is stuck in committee.

The Bottom Line

On Tuesday, twelve people will decide whether money should be more or less expensive. Their decision will shape the economic environment for every American worker — including the millions being displaced by AI in ways that don't register in any number the Fed tracks. Jerome Powell knows the Fed doesn't have the tools. He said so, on the record, to the United States Senate. Then he went back to using those tools, because they're the only ones he has. The result is monetary policy being made for an economy that no longer exists — an economy where unemployment means what it used to mean, where GDP growth means prosperity is broadly shared, and where a rate cut can bring back a job that a language model eliminated. None of those things are true anymore. The models haven't caught up. The rates reflect a world that's already gone.