The Government Pays Companies to Replace You. Then It Cuts the Program That Would Retrain You.

A company buying a $1 million robotic assembly line this year gets roughly $210,000 to $350,000 back immediately in tax savings. A company hiring ten workers at $100,000 each gets nothing comparable. The machine depreciates at 100 percent in year one under the One Big Beautiful Bill Act. The workers just cost money.

That's a 23.5-percentage-point gap in how the federal tax code treats capital versus labor. It's been there in various forms for decades, but the OBBBA made it permanent. Under prior law, bonus depreciation was phasing down — 80 percent in 2023, 60 percent in 2024, gone by 2027. Congress reversed the phase-out. Machines get full, immediate tax relief. Humans don't.

I keep coming back to that number because it clarifies everything that follows.

Five Actions, One Direction

Since December 2025, the federal government has taken five major actions on AI. They've been reported as separate stories. Tech policy people covered the export controls. Labor reporters covered the DOL cuts. Legal analysts covered the state preemption fight. Nobody put them in the same room.

They belong in the same room.

Vector 1: Kill State Regulation Before It Starts

Executive Order 14281, signed December 11, 2025, directed three federal agencies to go after state AI laws simultaneously. The Commerce Department got 90 days to identify "onerous" state regulations. The FTC was told to issue guidance on when state laws requiring "alterations to truthful outputs" are preempted by federal authority. And a new AI Litigation Task Force at the DOJ was authorized to challenge state laws on Commerce Clause, First Amendment, and preemption grounds.

The Commerce Department's 90-day deadline was March 11. As of this writing, no report has been publicly confirmed. That silence is significant — either it's being circulated internally before release, or the legal complexity of preempting Colorado's SB 24-205 proved harder than expected.

Colorado's law is the target. It's the only comprehensive state AI regulation in the country, requiring impact assessments and disclosure for "high-risk" AI systems. Enforcement was already delayed to June 30, 2026, after a 40-organization coalition — including the Chamber of Commerce, BSA, and TechNet — formally lobbied for federal preemption in January.

There's a fourth mechanism nobody mentions: $42.45 billion in BEAD broadband funding has been conditioned on states not enacting "onerous" AI laws. Financial coercion bolted onto legal coercion.

If Colorado falls before June 30, no state has a viable AI workforce protection law. There is nothing underneath.

Vector 2: Accelerate Global AI by Lifting Export Controls

The Bureau of Industry and Security rescinded the Biden-era AI Diffusion Rule — the tiered export control system that restricted AI chip sales to allied nations. Under Secretary Kessler instructed BIS staff not to enforce the rule. The new guidance narrows controls to China and Macau diversion only.

The Diffusion Rule was designed to slow global AI capability accumulation. Its removal does the opposite. More chips available to more countries means faster model training everywhere, which means faster automation capability, which compresses the displacement timeline. When Databricks reports a 327 percent surge in multi-agent enterprise workflows over a single quarter — based on behavioral data from 20,000 customers — that timeline matters.

Vector 3: Punish Any Company That Shows Restraint

On March 3, 2026, the Pentagon designated Anthropic — maker of Claude — as a supply chain security risk under the Federal Acquisition Supply Chain Security Act. Anthropic became the first American company ever to receive this designation.

The trigger: Anthropic refused to waive contractual restrictions on mass domestic surveillance applications and fully autonomous weapons systems. It wouldn't modify its acceptable use policy. So the government classified it alongside foreign adversary firms and ordered every federal agency to stop using its products within six months.

Former CIA Director Michael Hayden called it a "category error." Lockheed Martin began cutting ties. And OpenAI, which had already agreed to defense applications with fewer restrictions, gained federal market share by default.

The signal to the rest of the industry is not subtle: maintain safety guardrails and lose government contracts. Remove them and get rewarded. Two federal lawsuits filed by Anthropic on March 9 — First Amendment retaliation and Administrative Procedure Act claims — will determine whether corporate AI ethics has any legal protection at all.

The irony: the resulting consumer boycott pushed Claude to 11.3 million daily active users and the number-one spot on the App Store in 20 countries. The market punished OpenAI while the government punished Anthropic. Two constituencies, opposite incentives.

Vector 4: Defund the Safety Net During the Fire

The FY2026 presidential budget proposed $4.6 billion in cuts to Department of Labor workforce development programs — roughly a 35 percent reduction.

Job Corps: eliminated. All 123 centers, 50,000 annual enrollees, the only federal residential job training program — gone. WIOA funding — the Workforce Innovation and Opportunity Act, the primary federal retraining vehicle since 2014 — slashed. O*NET and CareerOneStop, the labor market information tools that help displaced workers figure out what skills are in demand: cut.

MASA grants — Make America Skilled Again — were announced as replacements. They operate at a fraction of the funding they'd replace, target "in-demand skills" with no AI-specific programming, and haven't enrolled anyone yet.

The timing is the cruelty. Anthropic's Economic Index documents a phase shift from AI-as-copilot to AI-as-replacement happening in the second half of 2025. Databricks measures a 327 percent increase in automated multi-agent workflows over a single quarter. And at this exact moment — during the steepest acceleration in AI workplace automation ever measured — the federal government is dismantling the institutional infrastructure that workers would need to survive the transition.

Vector 5: Pay Companies to Automate

Back to that 23.5-percentage-point gap.

Under the One Big Beautiful Bill Act, 100 percent bonus depreciation for equipment purchases is now permanent. A company that buys AI automation hardware can deduct the full cost in year one. Section 179 expensing — preserved and expanded. The effective tax rate on capital investment is roughly 15 percent. The effective rate on labor costs — payroll taxes, benefits, compliance — sits around 38.5 percent.

No CFO ignores a 23.5-point spread. When Klarna replaced 3,104 customer service positions with AI, or when Shopify's CEO told managers to "prove AI can't do it" before any new hire, or when Amazon flattened 30,000 management roles in a single quarter — those decisions happened inside this tax environment. The machines were already cheaper. The tax code makes them cheaper still.

The Feedback Loops

The five vectors aren't parallel. They're interlocking.

Combine Vector 5 with Vector 4: companies get a fiscal incentive to automate, and the workers they displace have nowhere to go. The tax code pays for the displacement. The budget cuts eliminate the response.

Combine Vector 1 with Vector 3: states can't regulate AI deployment, and companies that voluntarily restrict themselves get punished. No regulatory floor. No corporate self-regulation. The floor is gone in both directions simultaneously.

Combine Vector 2 with Vector 5: global AI deployment accelerates through lifted export controls while domestic deployment accelerates through tax incentives. Both timelines compress at once.

This is why calling it a doctrine — rather than a series of unrelated policy decisions — matters. Each vector reinforces the others. The whole produces outcomes none of the parts would achieve alone.

The One Counter-Signal

On March 11, Senators Mark Warner and Mike Rounds introduced the Economy of the Future Commission Act. Bipartisan. Endorsed by Google, IBM, Meta, Microsoft, Workday, Stanford's Erik Brynjolfsson, and Georgetown's CSET. Its scope explicitly includes "taxation policy" — the first federal vehicle that could study a displacement-proportional levy.

It will deliver a final report in 13 months.

Thirteen months. The doctrine has been operating since December 2025. By the time the commission reports back — late 2027, if it stays on schedule — the OBBBA depreciation incentives will have been permanent for two years. Job Corps will have been dead for two years. Colorado's SB 24-205 will have been litigated, probably to conclusion. Anthropic's lawsuits will have produced precedent. Tens of thousands of additional workers will have been displaced into a system stripped of retraining infrastructure.

Studying the flood while dismantling the dam.

The Bottom Line

There is no conspiracy here. Every one of these five actions has its own policy rationale. State preemption protects interstate commerce. Export control relaxation supports allies. Anthropic's designation addresses supply chain security. DOL cuts reduce spending. Bonus depreciation encourages investment. Each one has defenders who would reject the framing of this article entirely.

But the workers sitting in those five overlapping crosshairs don't experience policy rationales. They experience a tax code that pays for their replacement, a safety net that was cut while they were falling, state protections nullified before they took effect, and a commission that will study the problem on a timeline measured in years while the displacement accelerates in months.

Deploy Maximum. Protect Zero.

Name the doctrine and it becomes harder to pretend these are five separate things.