The Man Whose Job Is to Tell You It Won't Work

The hearing room on Capitol Hill, May 16, 2023. Sam Altman, CEO of OpenAI - the hottest company in the world - sits to Gary Marcus's left. Christina Montgomery of IBM sits to his right. Under oath, Marcus says the quiet part out loud: the technology being deployed is fundamentally broken, the companies building it cannot be trusted to regulate themselves, and we need an agency with actual teeth - something like the FDA for AI.

Altman nods along. The senators ask about risks. Marcus describes a "perfect storm of corporate irresponsibility, widespread deployment, lack of adequate regulation, and inherent unreliability." The man next to him runs the company being described.

This is the Gary Marcus experience. Not hostile. Not angry. Just persistently, methodically, documentably correct - and willing to say so in any room, on any stage, before any audience.

He has been AI's most consistent skeptic for over 25 years. Not the conspiracy-theory kind. The kind who read the papers, built the startups, trained the students, and noticed that the emperor's clothes were cut from very thin cloth. While the industry celebrated each new benchmark, Marcus was pointing at what the benchmarks missed. While investors poured hundreds of billions into large language models, he was asking whether language models that hallucinate with supreme confidence were really what anyone wanted deployed in hospitals and courtrooms.

The Kid Who Coded Before He Could Drive

Baltimore, 1978. Gary Marcus is eight years old and running a paper-based computer simulation. By ten he is fascinated by artificial intelligence. By sixteen he has written his first AI program. The curiosity isn't academic - it's urgent. He wants to understand how minds work, whether built of neurons or silicon.

He skips the last two years of high school to enroll at Hampshire College. Graduates in three years. Walks into MIT's doctoral program at 19. At 23, he has a PhD from the Brain and Cognitive Sciences department, supervised by Steven Pinker - one of the most celebrated cognitive scientists alive. His dissertation is on inflectional morphology, the patterns children use when they say "breaked" instead of "broke." The errors aren't random. They're windows into how the mind structures language.

That insight - that structured rules and messy pattern-matching coexist in human cognition - never left him. It became the lens through which he would spend the next three decades evaluating artificial intelligence. And finding it wanting.

He joined NYU as a professor in 1993. Spent 23 years studying cognitive development. Watched neural networks go from a backwater curiosity to the dominant paradigm in AI. Watched the field claim - repeatedly - that this time, the problem was basically solved. Watched each claim quietly implode. Kept notes.

A Quarter Century of Correct

Six Books. One Argument.

Each book is a different angle on the same fundamental question: what does understanding actually require?

The Debates That Defined a Career

Marcus has spent decades in public argument with people who run the institutions they disagree about. His debates are not theoretical.

What He's Actually Saying (And Why It Matters)

Strip away the debates and the books and the newsletter and what you get is one argument, made consistently since the late 1990s: language is not the same as understanding, and pattern-matching is not the same as reasoning.

Large language models are extraordinary at predicting what text should come next. They have been trained on essentially all the text that exists. They can produce fluent prose, working code, plausible medical advice, and confident nonsense - often in the same paragraph. Marcus calls this "the Eliza effect": the human tendency to attribute understanding to systems that merely mirror our language back at us.

His specific technical objections are: LLMs cannot maintain adequate world models; they cannot extrapolate reliably outside their training distribution; they hallucinate not occasionally but structurally; they have no way to know what they don't know; and the scaling hypothesis - the idea that more data plus more compute will eventually produce genuine intelligence - has no theoretical foundation. It is an empirical observation extrapolated into a cosmological claim.

What he wants: an FDA-like regulatory agency with authority to require pre-deployment testing, mandate transparency, and recall AI products that cause harm. Universal basic income to absorb AI-driven job displacement. International coordination to prevent a race to the bottom. He has proposed this in books, in op-eds, in Senate testimony, and in 103,000-subscriber newsletter posts. The proposals remain largely unimplemented. He keeps writing.

The Cognitive Scientist Who Learned Guitar at 40

Here is the thing about Gary Marcus that the AI criticism sometimes obscures: he is genuinely curious about human learning, and he uses himself as the experimental subject.

On the eve of his fortieth birthday, he had never played a musical instrument. He described himself as "coordinatively challenged." He decided this was scientifically interesting - he would learn guitar, document everything, and turn it into a book that was simultaneously a memoir and a study of adult learning, critical periods, and the science of skill acquisition.

He visited Day Jams, a rock camp for children ages 8-15. He performed in a band called Rush Hour. He practiced scales while his daughter watched. The resulting book, Guitar Zero, became a New York Times Bestseller - partly because it was genuinely funny, partly because it demolished the received wisdom that adults cannot learn new complex skills, and partly because reading about someone else struggling to play a G chord is oddly comforting.

The guitar episode is not a detour from his work. It is the same work. The same question: what does it actually take to learn something? What are the constraints? What can be changed? The guitar, the PhD dissertation on children's language errors, the critique of AI - they are the same project, wearing different shoes.

Six Predictions for 2026

At the start of 2026, Marcus published his predictions. He has been making annual predictions for years. The track record is publicly available. Here are the 2026 calls:

The People Around the Argument

What's Happening Now

Go Deeper