The White-Collar Recession

She rewrites the email for the third time.

Not because the phrasing is wrong — it's fine, grammatically, structurally, tonally fine. The problem is that the phrasing is exactly what the AI draft produced forty seconds ago. She's a compliance analyst at a mid-size financial firm in Charlotte. Fourteen years of careful language. Fourteen years of knowing which word makes a regulator relax and which one makes them pick up the phone. Now she spends her mornings editing outputs that are almost right, and her afternoons wondering how long "almost" keeps her necessary.

Her manager hasn't said anything. That's the part that stays with her. No meeting. No restructuring memo. Just a quiet Tuesday when she noticed the team Slack had fewer questions. The junior associates weren't asking her to review their drafts anymore. They were asking the tool. And the tool was asking nobody.

She has a master's degree. She earns well above the median. She is precisely the person who was told — by parents, by professors, by every career counselor she ever met — that education was the durable investment. The thing that couldn't be outsourced.

She doesn't talk about it at dinner. Not yet. She rewrites the email a fourth time, adding a comma the AI missed, and sends it. It's the comma that keeps her employed. For now.

The measurement is new. The displacement is not.

Anthropic published a landmark study in March 2026 — "Labor market impacts of AI: A new measure and early evidence" by Massenkoff and McCrory — that distinguishes between what AI can theoretically do and what it is actually observed doing in professional settings.^[1] This distinction is the paper's core contribution. Previous AI-exposure indices measured capability. This one measures deployment.

The gap is staggering. For computer and mathematical occupations, AI can theoretically handle 94% of tasks. Claude, Anthropic's own model, currently covers 33% in observed professional use. Office and administrative roles show 90% theoretical capability — a fraction in actual deployment. The distance between those numbers is not comfort. It is runway.

The demographic profile of the most AI-exposed workers is precise and unsettling. They are 16 percentage points more likely to be female. They earn 47% more than the median. They are nearly four times more likely to hold a graduate degree. This is not a blue-collar story. This is the professional class discovering that expertise is a depreciating asset.

The labor market is already showing early symptoms. Job-finding rates in AI-exposed fields have dropped 14% since the ChatGPT era began. Employment among workers aged 22 to 25 in these fields has fallen 16%. The young feel it first — not because they're less skilled, but because they're cheaper to not hire than to train.^[3]

This week the U.S. economy shed 92,000 jobs. Unemployment ticked to 4.4%.^[4] Block — the payments company formerly known as Square — fired approximately 4,000 employees, citing AI as a direct replacement for the roles eliminated. These are not abstractions. They are addresses and severance packages.

Federal Reserve Governor Michael Barr, in a February 2026 speech, outlined three AI scenarios for the labor market. The optimistic one involves reallocation — workers shifting into new AI-complementary roles. The middle scenario describes a "painful transition" with significant short-term unemployment. The dark scenario is the white-collar recession the Anthropic paper names: a structural contraction in knowledge-work employment that mirrors what manufacturing endured between 2000 and 2010, but faster and aimed at a class that never expected it.^[5]

Stanford's Digital Economy Lab confirms the direction: a 13% drop in hiring for AI-exposed entry-level positions. The displacement pattern mirrors historical technology transitions — electricity reorganized factories, personal computers reorganized offices — but this one targets a different demographic. The workers most exposed are not those who lacked access to education. They are those who invested the most in it.

The professional class that believed credentials were permanent insurance is now discovering that insurance has an expiration date. The "Great Recession for white-collar workers" isn't a prediction from a think tank. It's a named scenario, with measurable thresholds, published by the company building the tools. The gap between what AI can do and what it is currently doing is not a buffer. It is a countdown.