The Subsidy Problem: Why Companies Are Firing the People AI Needs to Work

THE SUBSIDY PROBLEM

Let's start with what Anthropic actually found.

In their March 2026 labor market study, Anthropic's researchers introduced a new metric called 'observed exposure', the gap between what AI could theoretically do and what it's actually doing in professional settings. The results are striking. For computer and math workers, AI can theoretically handle 94% of their tasks. In practice, it covers 33%. For office and administrative roles: 90% theoretical, a fraction in use.

That gap is not a bug. It's the space where human judgment, institutional knowledge, regulatory understanding, and contextual intelligence live. It's the reason a doctor hasn't been replaced by an LLM that can theoretically authorize prescription refills. It's why a financial analyst still has a job even though Claude can process balance sheets. The gap exists because real work happens in context, and context is the one thing AI cannot yet generate for itself.

Now look at what Meta did with that information.

THE HUMAN SUBSIDY

Meta is cutting approximately 15,000 jobs, 20% of its workforce. Reuters reported the layoffs in mid-March. Meta's stock rose 3% on the news.

But here's the critical distinction that Uri Gal, professor at the University of Sydney, identified: these workers are not being replaced by AI. They are being eliminated to fund AI. Meta is simultaneously committing $135 billion to AI infrastructure, data centers, researchers, compute. The humans being fired are subsidizing a technology bet.

This is not automation. This is sacrifice.

Block did the same thing a month earlier. Jack Dorsey cut 40% of Block's workforce, citing 'intelligence tools.' Stock up over 20%. But Block didn't demonstrate that AI was doing the work those people did. It demonstrated willingness to bet that AI would eventually do that work, and the market rewarded the bet.

The pattern is now systemic. Since January 2026, over 59,000 tech workers have been laid off. Digital Journal reports that roughly 20% of all tech layoffs this year are explicitly linked to AI restructuring. But as SingularityHub's analysis puts it: 'The timing and framing of the layoffs attributed to AI warrant closer examination.' Corporate restructuring, pandemic-era over-hiring, and investor pressure to demonstrate AI commitment are all operating at the same time as genuine automation.

The market doesn't distinguish between 'AI replaced these workers' and 'we fired these workers to buy AI.' Both get the stock bump. Both get the analyst upgrade. Both get the narrative of innovation.

THE KNOWLEDGE PARADOX

Here's where it gets structurally dangerous.

Anthropic's research shows that AI's theoretical capability far exceeds its actual deployment. The gap between the blue bar (what AI could do) and the red bar (what it's doing) represents the accumulated knowledge, judgment, and contextual understanding of the humans currently doing those jobs.

To close that gap, AI needs training data. It needs workflow integration. It needs domain expertise to validate outputs. It needs the institutional knowledge of the people who know why things are done the way they're done, not just what gets done.

Those are exactly the people being fired.

This is the subsidy problem: companies are destroying the human capital that their AI investment depends on in order to fund that investment. It's like demolishing a building to pay for the architect who will design its replacement, then realizing the new architect needs the original blueprints that were inside the building.

The Anthropic researchers themselves note the most AI-exposed workers are 'older, female, more educated, and higher-paid.' They hold graduate degrees at four times the rate of the least exposed workers. They earn 47% more on average. These are not entry-level workers performing routine tasks. These are the people with the deepest institutional knowledge, the ones who know which 33% of theoretical tasks actually get used, and why the other 61% remain theoretical.

When you fire them, you don't just lose their output. You lose the map.

THE HIRING FREEZE IS WORSE THAN THE LAYOFFS

The Anthropic study found something that matters more than the layoffs: a 14% drop in job-finding rates for workers aged 22-25 in AI-exposed occupations since ChatGPT launched. Young workers aren't being fired. They're never getting hired.

This is the quiet version of the subsidy problem. The experienced workers being laid off at least had time to develop judgment, build institutional knowledge, create the workflows that AI is now learning from. The young workers who never get hired never develop any of it.

The pipeline is being cut at both ends. Senior knowledge holders are being eliminated to fund AI. Junior workers who would become the next generation of knowledge holders are never entering the system. The institutional knowledge base is shrinking from the top and not being replenished from the bottom.

In five years, when the gap between AI's theoretical and actual capability needs human expertise to close, who will have it?

THE MARKET IS NOT THE ECONOMY

Every company in this story saw its stock price rise after announcing layoffs. Block up over 20%. Meta up 3%. The market is rewarding what it interprets as AI-forward strategy.

But Goldman Sachs estimates that if AI were used for everything it's currently capable of, roughly 2.5% of US employment would be at risk. Not 20%. Not 40%. Two and a half percent. The Federal Reserve's own analysis shows no systematic increase in unemployment for highly exposed workers since late 2022.

The market is pricing in a future that the data doesn't yet support, and companies are destroying present value to chase that pricing signal. This is not rational resource allocation. This is a coordination failure where every individual company's incentive (cut humans, buy AI, get stock bump) creates a collective vulnerability (destroy the knowledge base AI needs to work).

Paul Kedrosky's phrase from The Atlantic keeps echoing: 'all of these interlocking points of fragility.' Add this one to the list. The human knowledge infrastructure that AI depends on is being treated as a cost to be cut rather than an asset to be preserved.

WHAT THIS MEANS FOR BRANDS

If you're a brand or organization navigating this, three things are true simultaneously:

1. AI will displace some work. The Anthropic data is clear on this. Denying it is denial.

2. The displacement timeline is longer than the market is pricing. The gap between theoretical and actual capability is enormous and will take years to close, and it can only close with human expertise guiding it.

3. The companies that will win are the ones that understand the subsidy problem and refuse to participate in it. Preserve your institutional knowledge. Invest in the humans who can close the gap between what AI can theoretically do and what it actually does in your specific context. That judgment, that contextual intelligence, is the last durable competitive advantage.

The market will reward you for firing people today. The market will not rescue you when your AI can't function because the people who knew how your business actually works are gone.

This is not sentimentality. It's infrastructure.