How many developer jobs has AI cost in the United States?

According to the Federal Reserve Board study (March 2026) by Crane and Soto, roughly 500,000 programmer jobs failed to materialize over three years compared to the trajectory expected before ChatGPT. This is not a net-loss count, it's a growth shortfall versus the no-AI scenario.

Did the AI effect on dev employment hit immediately?

No. The gap between observed and expected trajectories only widens significantly around mid-2024, roughly 18 months after ChatGPT shipped. That absorption delay is typical of tech transitions.

Is France seeing the same trend?

Yes. France's Apec reports a 18% drop in IT executive hires in 2024, the first contraction since 2009 outside of COVID. Developers specifically are down 20% in 2025, IT project managers down 23%, and juniors across all roles down 19%.

What is Patrick Artus's core argument?

Artus argues that productivity gains only translate into sustained growth if a sufficient share returns to labor. Since the 1980s, wages have grown at only 40% the pace of productivity. His worry: AI gains may further skew that distribution, especially in the US where capital returns are already much higher.

Are current developers seeing salary pressure?

Not according to Fed data. The adjustment runs through hiring volume, not through compensation for incumbents. Future entrants, and juniors in particular, are paying the price.

AI and dev jobs: what the Fed actually measured

5% per year. That was the pace at which US programmer hiring grew before November 2022. Well above the broader US labor market. Since ChatGPT shipped, the curve has flatlined.

This isn't a sentiment survey. It's a study published in March 2026 by the Federal Reserve Board, authored by Leland Crane and Paul Soto. The researchers cross-referenced the Bureau of Labor Statistics' Current Population Survey with the Department of Labor's O*NET database, then built an industry-level control variable to isolate occupation-driven effects from sectoral cycles. The verdict comes down to three words: coder-specific shock.

The same day these numbers hit the tech press, Patrick Artus published an op-ed in Le Monde under a programmatic headline. "AI's impact on growth will depend on the scale of redistributive policies and how income is shared." Two signals, two timeframes, one nerve: who pays for the transition.

What the Fed study actually measures

The effect didn't kick in right away. The gap between observed and expected employment trajectories only opens up meaningfully around mid-2024, roughly 18 months after ChatGPT launched. That absorption delay is worth flagging, because it's exactly what makes the effect hard to see while you're living through it. The curve doesn't crash, it slows. Like a train braking without ever slamming the brakes.

Across three years, the cumulative gap represents about 500,000 jobs "that would probably have existed" without large language models. Crane and Soto themselves stress how cautious you need to be with that number. Many would-be programmers shifted into adjacent fields: data engineering, machine learning, security. The 500,000 isn't a net job-loss count, it's a growth shortfall against the no-AI scenario.

A detail that changes the political reading: no clear wage drop. The adjustment runs through hiring volume, not compensation for incumbents. Which means working developers don't feel pressure on their paycheck. Future entrants pick up the tab.

France in the same train, without the same candor

France's Apec released its 2025 forecasts. The numbers track the US trajectory. IT executive hires: -18% in 2024, the first contraction since 2009 outside the COVID period.

Developer hires specifically: -20% in 2025 versus 2024. IT project managers: -23%. Juniors across all roles: -19%.

A late-2025 Stanford study corroborates the age pattern. Workers aged 22-25 in occupations highly exposed to AI saw their employment drop about 16% versus the expected trend after ChatGPT. Seniors stayed stable. The logic is consistent across Paris, San Francisco, and Stanford: the filter is closing at the entry point.

One thing stands out in French official communications. Apec attributes the drop to a "less favorable economic climate" and "limited medium-term visibility." AI doesn't show up in their primary list of causes.

That's perhaps statistically prudent, since Apec doesn't measure direct causation. But the temporal coincidence is getting harder to ignore. The first decline in French IT employment in twenty years, exactly when LLMs become a daily tool for dev teams, deserves at least an explicit hypothesis.

What Artus sees and what he misses

Patrick Artus doesn't write about labor markets. He writes about macro. His thesis fits in a simple equation: a productivity gain only translates into sustained growth if enough of it returns to labor. Otherwise demand can't keep up with supply, and the gain evaporates into capital concentration.

The figure Artus has been hammering for years: since the 1980s, wages have grown at 40% of the pace of productivity. The data is measured, not projected. Income sharing distorted long before AI. His worry is that AI gains will further accelerate that distortion, especially in the US model where capital profitability is already markedly higher (ROE 18% US vs 9% EU).

On that frame, Artus is right. The argument is solid, documented, and necessary to the debate. But his op-ed has a blind spot. It speaks in conditional future tense ("the impact will depend") about a mechanism that's already measurable in the present.

Crane and Soto have just put numbers on what macro had been treating as a hypothesis. The question isn't only how to share future gains. It's also how to reabsorb the 500,000 jobs that already didn't happen, and the French juniors knocking on a door that's narrowed.

The bridge between Artus's macro and the Fed's micro is exactly the political conversation that should be opening. Should there be a specific tax on AI productivity gains, as some economists propose? A sectoral redistribution requirement? A massive public investment in training the juniors no longer being hired?

None of these answers is trivial. And none gets triggered without first looking the numbers in the face.

That's maybe the real novelty of this moment. For three years, the AI-and-jobs debate ran on diverging projections. Today, we have hard data. The lag between observation and policy decision is now the limiting factor.