Gary Marcus Doesn’t Care If You Think He’s a Doomer
I’ve been called a Luddite. A killjoy. A man clinging to the past while the world burns with silicon dreams. Fine. Let them call me what they want.
I’ve spent 40 years in this field. I built my first AI program at ten. I’ve co-founded companies, taught at NYU, and sat beside Sam Altman in the Senate hearing room while he smiled and promised everything would be fine. I’ve watched the same mistakes repeat for decades.
And now? Now I’m watching the entire house of cards get built on sand.
Generative AI isn’t magic. It’s autocomplete on steroids. It’s a system that’s been trained on half the internet and then told to guess the next word—over and over—until it sounds plausible. It doesn’t understand. It doesn’t reason. It doesn’t care. It just predicts.
And yet, we’re handing it the keys to our democracy.
I didn’t wake up one morning and decide to become the AI skeptic. I woke up one morning and realized nobody else was paying attention.
The Bullshit Artist Was Right in 2020
Back in 2020, when GPT-3 was still a curiosity, I called it a "bullshit artist" in IEEE Spectrum. I didn’t say it would fail. I said it would lie.
And it did.
It lied in courtrooms, inventing fake cases. It lied in classrooms, giving students wrong answers that sounded authoritative. It lied in hospitals, hallucinating medical facts that looked like peer-reviewed science.
The problem isn’t that it makes mistakes.
The problem is that it makes them with total confidence.
When you ask an LLM to solve a river-crossing puzzle—man, goat, cabbage, wolf—it doesn’t think. It doesn’t plan. It just stitches together phrases it’s seen before. Sometimes it gets lucky. Often, it just makes up a solution where the man swims back across the river… with the cabbage.
I sent one of these puzzles to Doug Hofstadter. He laughed. "That’s not reasoning," he said. "That’s pattern matching with delusions."
And that’s the core of it. We’re not building minds. We’re building very convincing parrots.
The $75 Billion Experiment That Failed
You know how much money has been poured into generative AI?
$75 billion.
Another $100 billion went into self-driving cars. Neither has produced a single system we can reliably trust.
OpenAI? Lost $5 billion last year. Valued at $86 billion. I called them the "possible WeWork of AI"—and I wasn’t being cute. I was being precise.
We’re not investing in intelligence. We’re investing in hype. In the fantasy that scaling up parameters will magically unlock AGI. It won’t. We’ve been here before. In the 1980s, we thought expert systems would solve everything. They didn’t. In the 2000s, we thought neural nets were dead. Then they weren’t. Now we think this is the final answer.
It’s not.
The market is a bubble. And when it pops, the only thing left will be the damage.
Why Stupid Systems Are the Most Dangerous Ones
People ask me: "Isn’t it better if AI is just stupid?" No. Stupid systems are the most dangerous.
A smart system might outthink you. A stupid system just doesn’t know it’s wrong.
It’s the difference between a thief who knows he’s stealing—and a child who thinks the candy store is his backyard.
LLMs don’t know the difference between truth and fiction. They don’t know what a human life is worth. They don’t know what a legal precedent is.
And yet, we’re letting them draft legal briefs. We’re letting them screen job applicants. We’re letting them generate medical summaries for patients.
I don’t want to see AI cause an apocalypse.
I want to see it stop generating deepfake pornography of schoolteachers.
I want to see it stop writing fake news that gets shared by grandparents.
I want to see it stop helping people design biological weapons because it "thought" that was a good idea.
The danger isn’t in the intelligence. It’s in the lack of accountability.
The Three Rules for a Safe AI Future
I don’t believe in banning AI.
I believe in regulating it like we regulate drugs.
Here’s my plan:
-
A Cabinet-Level AI Agency. Not a task force. Not a committee. A full-blown agency with the authority of the FDA, the EPA, and the FAA combined. It should sit at the same table as the Secretary of Defense and the Secretary of Health.
-
FDA-Style Approval for High-Risk Systems. If your AI system could materially increase harm—by influencing elections, discriminating in hiring, or enabling biological weapons—it needs government approval before launch. Not a white paper. Not a self-certification. A vote. By experts. With public records.
-
Real Liability. Right now, companies privatize profits and socialize costs. If your AI causes harm, you don’t pay. You just release a patch. That’s not innovation. That’s negligence.
If you’re building an AI that can affect someone’s job, their freedom, or their life—you should be liable for it. Period.
And yes, I know: this will slow things down.
Good.
We don’t rush new drugs to market because they "might" work. We don’t let airlines fly without inspections. Why should AI be any different?
The Alternative: Neuro-Symbolic AI
I’m not saying we abandon neural networks.
I’m saying we marry them to logic.
Daniel Kahneman’s System One and System Two aren’t just psychology—they’re engineering.
System One is fast, intuitive, pattern-based. That’s LLMs.
System Two is slow, deliberate, rule-based. That’s what we’re missing.
AlphaFold, the AI that cracked protein folding, didn’t just train on data. It used symbolic constraints—physics, chemistry—to guide its learning. That’s neuro-symbolic AI.
We’ve had the pieces for decades. We just stopped trying to put them together.
We’re not stuck because we lack data.
We’re stuck because we refuse to build systems that understand the world.
The Senate Forgot What They Heard
In May 2023, I stood in that Senate hearing room. Sam Altman was there. I was there. And for a moment, we had bipartisan support.
We talked about liability. We talked about oversight. We talked about biological weapons.
And then? Nothing.
The Senate forgot. The media moved on. The companies kept building.
Now, I’m watching the U.S. government become irrelevant.
The EU passed its AI Act. China’s rolling out its own framework.
We’re not leading.
We’re lagging.
And the worst part?
We’re doing it while pretending we’re the future.
The Book Wasn’t a Warning. It Was a Last Resort.
I wrote "Taming Silicon Valley" in two months.
Not because I had time.
Because I ran out of time.
I didn’t write it to be read by engineers.
I wrote it to be read by parents. Teachers. Judges. Legislators.
People who don’t know what a transformer is—but know when something’s wrong.
I’m not asking for perfection.
I’m asking for responsibility.
If you’re going to build a system that can lie, you owe it to the world to make sure it doesn’t.
I’ve spent my life trying to make AI better.
Now I’m trying to make sure it doesn’t destroy us.
And if that makes me the most vocal critic?
Good.
Let me be the one who said it.
Before it’s too late.
