I recently overheard a conversation in a River Falls, Wisconsin, coffee shop that has stayed with me.
A person was venting to a friend about being forced to use AI—likely Microsoft Copilot—within a work spreadsheet. After several rounds of frustrating trial and error, she eventually gave up and finished the task manually. In the end, the “shortcut” had cost her more time than if she’d just done the work herself from the start.
To many, this looks like failure, but I see it as a green flag. It is a sign that the system is working exactly as it should: by keeping a human in the loop, much like a pilot in a cockpit. Being frustrated is proof, to me, they’re effectively using this amazing tool in the real world.
But if we want to phase out human-in-the-loop systems for AGI, there is one critical problem we must solve first—at least, if we care about society. Beyond the technical hurdles, the real challenge is taxes.
Governments are funded by people, not software. Our schools, roads, and healthcare systems rely on the taxation of human income. When a worker is removed from the equation, that tax base vanishes with them. At scale, the “efficiency” of AI creates a massive shortfall in public revenue. The results are predictable: crumbling infrastructure, reduced services, and an even heavier tax burden on the few workers who remain.
I know what the technologists and free-market absolutists will say. I’ve heard the arguments, but they crumble under scrutiny:
“Corporate taxes will cover the gap.”
Argument: If companies replace humans with AI, their profits will skyrocket. We can just tax those profits to fund society.
Reality: Payroll taxes are unavoidable “flow” taxes. Corporate profit taxes are a game of hide-and-seek. Corporations are experts at shifting IP to tax havens, engaging in stock buybacks, and “reinvesting” to show zero profit on paper. You cannot fund a government on the goodwill of creative accountants.
“It’s just a tool, like a tractor.”
Argument: We didn’t tax tractors when they replaced farmhands; we shouldn’t tax AI for replacing lawyers.
Reality: Tractors magnified physical labor; they didn’t simulate the mind. More importantly, the industrial revolution took decades. The AI displacement is happening in quarters. “Retraining” is a myth when the goalpost moves faster than the semester.
“You can’t tax software; it has no location.”
Argument: If you tax AI “agents,” companies will just host the models in a deregulated jurisdiction and beam the work in.
Reality: Correct, you can’t tax the code. But you can tax the value flow, the revenue generated in-country, or the massive energy consumption of the data centers. We don’t need to give AI personhood to tax the electricity it burns to replace a taxpayer.
“This is just Luddite fear-mongering.”
Reality: It’s not fear; it’s math. If 30% of the workforce is displaced, and the remaining 70% have to pay for the social safety net (or UBI) required to keep the displaced alive, the math breaks.
Anyway…
This is why we should be cautious about celebrating the wholesale replacement of people with AI agents. I am an optimist when it comes to LLMs and the possibilities they unlock (especially in SWE), but enthusiasm shouldn’t blind us to economic reality.
Productivity gains that bypass human labor also bypass the tax system that sustains our society. AI should be used to offload drudgery—to free humans for meaningful work—not to quietly dismantle the systems that support us all.
Some investors know this. They aren’t trying to improve society; they are trying to exit it. As my AI assistant gemini-3-flash-preview put it so blindly today:
“It’s a bold strategy: replace the humans who pay for the roads, then take the billions you made and flee to an island where you don’t need roads.”
The evolution shouldn’t be human to AI, but human to human-in-the-loop. It’s the necessary compromise: we guarantee the quality of the output, capital gets its efficiency, and—crucially—the tax base that keeps civilization running remains intact.