Bloomberg New Economy Forum

Gary Marcus on why AI only works where answers get verified

Gary Marcus· Emeritus Professor of Psychology and Neural Science at New York University
·~26 min·English·Bloomberg
AI SafetyLLMReasoningPolicyBusiness Strategy
TL;DR

Gary Marcus argues today's AI is reliable only where answers can be externally verified, so the hype, the sky-high valuations, and the extinction fears all overshoot what the technology can actually do.

01Core Mental Model

The Verifier Test

AI is reliable mainly where an external system can check the answer — code that has to run, math that can be proven; where no such check exists, its output is far less trustworthy.

It's because these particular problems can be formally verified. You can do a bunch of logic to make sure your answer works.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
The verifier test quietly predicts which work AI transforms first: anything with a fast, cheap correctness check (unit tests, proof checkers) is exposed, while work whose right answer can only be judged by the real world stays stubbornly human.
02Nature of the Tool

Large Liability Models

The models approximate what people have said without understanding the world, so they follow instructions loosely — which is why Marcus says large language models may be better called large liability models.

What it does is it approximates what people have said in certain contexts. It has a lot of contexts to draw on, but it's understanding of what's going on in the world is superficial, which is why it can't actually follow instructions.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
If a model tracks the words of an instruction rather than its intent, then reliable instruction-following is not a feature you can bolt on afterward — it is the hard, unsolved core, and the loose obedience that leaks out is exactly what gets labeled rogue behavior.
03Historical Pattern

We've Seen This Movie Before

Chess and Go fell to AI decades ago and the world did not end, because those were narrow, domain-specific wins — solving hard math is the same kind of win, not a solved world.

When Deep Blue beat Garry Kasparov in 1997, a lot of people thought the world was going to end, employment would end, and none of that happened, right?

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
Each past milestone created a new benchmark, not a general mind; the pattern suggests today's math and coding wins will harden into useful tools rather than widen into the open-ended competence the valuations assume.
04A Hard Domain

Cancer in Five Years

In drug discovery AI can propose candidate molecules but cannot tell you which ones actually work, so every candidate still faces the same multi-year clinical trial.

The thing that they can do well is they can make candidates for drugs. The thing that they cannot do is to tell you if those candidates really work.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
The bottleneck was never generating ideas — it was validating them, and validation is the one step AI cannot shortcut, which is why cure cancer in five years confuses a faster idea factory with a faster reality.
05The Core Error

The Exponential Already Broke

The belief that simply adding data and compute keeps making models exponentially better is, in Marcus's words, already known inside the industry to be false.

There's this idea of scale is all you need that all you would need to do is add more data and they would get better and better. They get exponentially better. That's already basically known in the industry to not be true.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
If the exponential is already bending, the danger is not merely slower progress — it is that the whole investment case was underwritten on a curve that only ever existed in projections.
06The Money

The WeWork of AI

The AI business is running on valuations that assume profits it does not have — Marcus's image is Wile E. Coyote still sprinting after he has passed the edge of the cliff.

The industry is wildly overextended. The economics make no sense. The profitability is not really there.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
A bubble does not pop when the technology fails; it pops when belief fails — Marcus is betting belief breaks before the technology delivers, which makes the timing a psychological question, not a technical one.
07The Competition

No Moat

There is no moat: with routers that send your prompt to any model, switching costs are near zero and no company has the head start that made Google or Facebook durable.

Google got there with a huge head start. Nobody has a head start. There's no friction changing.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
Network effects made the last generation of tech giants defensible; without them, even a leading lab is one cheaper, good-enough competitor away from commoditization — which is what decelerating revenue growth is already showing.
08The Real Risk

Catastrophe, Not Extinction

Marcus puts human extinction near zero but the probability of catastrophe high — a hallucinating AI wired into fragile infrastructure — with everyday surveillance, he says, already here.

I think there's almost zero chance of literally extinguishing the species, which is what people talk about.

— Gary Marcus, Bloomberg New Economy Forum
Key Insight
The practical cost of the extinction narrative is misallocation: the headlines, the safety effort, and the stalled regulation all chase the outcome Marcus rates least likely, while the breakages he treats as probable — markets and infrastructure failing under unreliable systems — draw little oversight or insurance.