The AI Oligarchy Isn’t Just Powerful—It’s Getting Away With Murder
Gary Marcus doesn’t write op-eds. He doesn’t tweet hot takes. He doesn’t even do podcast cameos unless he’s got a full plan in his back pocket. He’s the guy who stood in the middle of the AI hype circus in 2017, holding up a cracked mirror, and said, "This thing doesn’t understand what a cat is. And it’s going to burn your house down before it figures it out."
We thought he was a crank. Now, we’re all watching the smoke.
He’s not angry. He’s disappointed. And that’s what makes him dangerous.
Marcus isn’t some Luddite. He built AI systems. He co-founded companies that sold them. He knows the magic. He’s seen the code. But what he’s watching now isn’t innovation—it’s institutionalized negligence. And the people running it? They’ve convinced themselves they’re saving the world. Meanwhile, they’re turning truth into a commodity, democracy into a target, and public trust into a quarterly KPI.
This isn’t a tech problem. It’s a civilizational one. And Marcus has a blueprint for how to stop it.
We Didn’t Build AI to Manipulate Elections
Let’s be clear: the LLMs we’re deploying weren’t designed to lie. They were designed to predict the next word.
But prediction, at scale, becomes persuasion. And persuasion, in the hands of companies with no accountability, becomes weaponization.
We’ve seen it. Fake campaign ads in Ohio, generated in minutes, voiced by cloned voices of local mayors. Deepfake videos of candidates admitting to crimes they never committed. AI-generated voter suppression bots that whisper doubt into the ears of elderly Black women in Georgia, telling them their polling place moved—again.
And who’s responsible? No one. Not legally. Not ethically. Not even publicly.
Marcus calls this "accidental misinformation." That’s a polite word for systemic collapse. These aren’t glitches. They’re features. LLMs hallucinate because they have no grounding in reality. They don’t know what’s true. They only know what’s statistically likely.
And the companies? They’re betting that you won’t notice. Or that you won’t care.
The Netherlands’ child welfare algorithm wrongly flagged 26,000 families for fraud. Australia’s robodebt scheme used the same logic—automated, opaque, unchallengeable. People lost homes. People died.
And we called it "efficiency."
The Moral Descent Wasn’t an Accident
You think AI got dangerous because it got too smart?
No.
It got dangerous because the people who built it stopped caring.
Marcus traces the arc back to Dartmouth, 1956. The original goal: machines that could mimic human thought. Not to replace us. Not to profit from us. To understand us.
Fast-forward to 2025. Meta’s AI team isn’t building for insight. It’s building for retention. OpenAI’s board isn’t debating ethics—it’s debating whether to release GPT-5 before the election cycle.
The shift? It wasn’t technical. It was moral.
"The desire to do good decreases with time," Marcus writes, "in the eternal quest for growth." (p. 84)
That’s not a quote. It’s a eulogy.
The companies that once bragged about "organizing the world’s information" now brag about "maximizing engagement." The engineers who once wanted to cure disease now want to optimize click-through rates. The venture capitalists who funded AI for healthcare now fund AI for deepfake porn.
And we let them.
Because we were distracted. Because we were dazzled. Because we were told this was inevitable.
Marcus says: it’s not.
Here’s the Plan Nobody’s Talking About
Marcus doesn’t want to ban AI.
He wants to make it accountable.
And he’s not asking for permission.
His plan is surgical. Precise. And terrifying to the people who profit from chaos.
First: a federal AI agency. Not a task force. Not a committee. A cabinet-level department—with subpoena power, audit rights, and the authority to shut down models that violate public safety standards.
Think FDA for algorithms. Think FAA for autonomous systems. We regulate planes because they can crash. We regulate drugs because they can kill. We regulate AI because it can destabilize democracies.
Second: minipublics.
No, not lobbyists. Not consultants. Not former regulators. Real people.
Marcus proposes embedding citizen panels—randomly selected, trained, compensated—directly onto the executive boards of the top five AI firms. These aren’t advisory. They have voting rights. One vote. One voice. One seat.
Imagine: a retired teacher from Nebraska, a nurse from Detroit, a high school student from Texas—all with equal power to block a model release because it risks voter suppression.
That’s not dystopia. That’s democracy.
Third: legal liability.
If your algorithm falsely accuses someone of welfare fraud and they lose their home? You pay. If your model generates a fake news story that triggers a stock plunge? You pay. If your system copies a thousand books without permission and trains on them? You pay.
No more "we didn’t know" defenses. No more "it’s not our fault" escapes. Developers, CEOs, boards—everyone bears responsibility.
Fourth: tiered oversight.
Not one agency. Three layers.
- Level 1: Internal audits by the company, with public reports.
- Level 2: Independent third-party auditors, funded by a public trust, not the companies.
- Level 3: The federal AI agency, with emergency shutdown authority.
Like nuclear plants. Like pharmaceutical trials. Like aviation.
We don’t let airlines self-certify safety. Why do we let AI firms self-certify truth?
The Real Enemy Isn’t the Tech—It’s the Erosion of the State
Here’s the part they don’t want you to hear.
The problem isn’t just Big Tech.
It’s the hollowing out of the state.
Marcus doesn’t just blame Elon Musk. He blames the voters who cheered when Musk dismantled the Department of Education’s digital infrastructure. He blames the legislators who defunded the FTC’s AI unit. He blames the governors who let state bureaucracies collapse so they could tweet about "freedom."
The "DOGE" purge in the U.S. wasn’t just a purge of bureaucrats. It was a purge of institutional memory. Of technical competence. Of the very capacity to regulate.
And now, the AI firms are moving in.
They’re not just lobbying. They’re replacing.
They’re offering to "build" the AI oversight systems themselves. They’re offering to "train" the regulators. They’re offering to "host" the data.
It’s not collaboration. It’s colonization.
Marcus calls it "technocolonisation." And he’s right.
The state isn’t failing because it’s weak. It’s failing because it’s been abandoned.
The Tragedy of the AI Commons
There’s a line in Marcus’s book that haunts me.
"Everybody goes for the trout while they are plentiful, and suddenly there are no trout left at all."
That’s the tragedy.
We’re not running out of data. We’re running out of truth.
We’re not running out of models. We’re running out of trust.
We’re not running out of innovation. We’re running out of meaning.
Every time we let an AI generate a fake news story, we erode the foundation of public discourse.
Every time we let a model scrape copyrighted books without permission, we kill the incentive for authors to write.
Every time we let a system make life-altering decisions without transparency, we teach citizens that institutions don’t care.
And when that trust is gone? No algorithm can rebuild it.
Marcus isn’t a pessimist.
He’s a realist.
He knows we can fix this.
But only if we stop pretending this is a technical problem.
It’s a political one.
And politics? That’s something we still control.
We just have to want to.
