Gary Marcus is an unsettling figure within the triumphalist narrative of contemporary artificial intelligence. A psychologist and cognitive scientist trained at MIT, and Professor Emeritus of Psychology and Neural Science at New York University, he first rose to prominence for his work on how children acquire language — those “goed” and “breaked” errors that reveal deep mental rules — and later for his books on the mind and brain, such as The Algebraic Mind, Kluge, or Rebooting AI.
Over the last decade, his center of gravity has shifted: from academic researcher to the “critical conscience” of AI. He combines entrepreneurial experience (having founded Geometric Intelligence, acquired by Uber, and Robust.AI) with intense public activity in the media and through his newsletter, Marcus on AI. There, he has specialized in debunking hype, documenting the failures of current systems, and defending — against the grain — the idea that we need a different technical and political direction for AI. He writes practically every day.
A Brilliant Technology, but Inherently Unreliable
For Gary Marcus, the paradox of generative AI is that it can perform spectacular feats while remaining a structurally unreliable technology. Large language models continue to hallucinate, invent sources, confuse basic concepts, and lack an internal notion of “truth” or “reality” to lean on. In his newsletter, he insists that none of this is a passing bug, but rather a direct consequence of their training: they are optimized to predict the next word, not to state facts. Therefore, he warns, using them without mediation in critical contexts (health, justice, infrastructure) is like delegating delicate decisions to someone who is highly eloquent but lacks both responsibility and memory.
Pure Deep Learning Has Hit a Conceptual Wall
Marcus does not hesitate to acknowledge the historic merit of deep learning, but he believes we have confused “extraordinary” with “sufficient.” In his view, the basic architecture of LLMs — deep networks trained solely on data — cannot, no matter how much you scale them, capture certain properties of human thought: causal reasoning, systematic understanding, and deliberate planning.
His thesis is that the public has been sold the idea that “more data and more GPUs” would be enough to reach a kind of general intelligence, when in reality we are hitting structural limits: models that mimic text patterns without understanding the world they speak of. This wall is not one of computation, but of representation: without explicit structures of rules, variables, and relationships, models will remain impressive imitators rather than robust thinkers.
The Alternative: Neurosymbolic AI, Hybrid by Design
As an alternative to the “neural networks for everything” approach, Marcus has spent decades advocating for a hybrid path: neurosymbolic AI. The idea is simple yet profound: combine the ability of neural networks to handle ambiguity, perception, and language with symbolic components capable of manipulating rules, graphs, programs, and explicit logical structures.
Economic Hype vs. a Much More Modest Reality
In his recent writings, Marcus settles the score with the economic hype surrounding generative AI. His diagnosis is that an immediate revolution in productivity and automation was promised, yet the data does not support the scale announced. He cites the Remote Labor Index study, according to which models can only take over 2.5% of human tasks — and even then, he considers that figure inflated because it excludes all physical labor.
Get fernandocomet’s stories in your inbox
Join Medium for free to get updates from this writer.
While there are useful use cases — summarizing, drafting, assisting with coding — these are a far cry from “replacing most jobs” or achieving a transformative AGI. From his perspective, many companies are discovering that the actual return on their investment in generative AI is lower than what the commercial narrative led them to believe.
Disinformation, Democracy, and the Erosion of Human Agency
Marcus frames generative AI as an accelerator of processes that were already damaging the public sphere, especially disinformation and polarization. The ability to produce synthetic text, images, and video at near-zero cost multiplies manipulation campaigns, floods the media ecosystem, and makes it extremely difficult to distinguish the authentic from the fabricated.
In this “information fog” he argues, the natural reaction of many people is cynicism: “nothing is reliable” — an attitude that weakens democratic participation as much as, or even more than, fake news itself. Furthermore, he is concerned about the silent integration of models into critical infrastructure — defense, cybersecurity, weapons systems — where an opaque “world model” could end up conditioning strategic decisions without clear supervision or accountability: a “black box moment” in which we surrender human agency to algorithms that calculate but are incapable of caring.
Work, Employment, and the Uncomfortable Reality of Partial Automation
Unlike the more alarmist or triumphalist headlines, Marcus adopts a curiously sober stance on employment and automation. He does not deny that profound changes are coming, but he doubts the narratives predicting a mass substitution of jobs within a few years. His reading of current data suggests that many roles will be altered — fragmented, augmented, or made more precarious — but not necessarily eliminated, and that a large portion of the economy will continue to rely on physical, relational, and situated-judgment tasks that current models cannot handle.
He is more concerned about the skills gap: those who know how to integrate AI into their work with critical judgment will have a clear advantage over those who do not, widening inequalities in an already strained labor market. In other words, the problem is not just “how many jobs are lost,” but who gets to keep the best pieces of those that are transformed.
Recognizing Real Value: Increments, Not Miracles
Being critical does not mean being a denialist, and Marcus is careful to recognize where he sees genuine value. He speaks of generative AI as a useful tool for incremental productivity: writing first drafts, exploring ideas, translating, reviewing code, or experimenting with design or text variations that a human then reviews.
He emphasizes that when used this way — as an augmentor rather than a replacement — it can improve workflows without the need to buy into the “general intelligence” narrative. He also points to hybrid cases where the combination of statistical models with symbolic structures or domain knowledge is already generating real breakthroughs in science and technology, from protein folding to mathematics.
His criticism, at its core, is directed at the inflation of promises: there is business and there are advancements, but that does not justify presenting generative AI as the sole and central engine of the next economic revolution.
The Self-Assumed Role of the “Necessary Party Pooper”
Finally, Marcus has been highlighting the gap between industry promises and the actual behavior of these systems, and he now sees many of his warnings confirmed by issues regarding reliability, disinformation, and economic return.
If more attention had been paid to critical voices, he argues, perhaps we wouldn’t be trying to manage a bubble of expectations and an information ecosystem saturated with generative content. From the outside, one can debate whether his tone is excessive, but his function in the ecosystem is clear: to remind us that behind the technical dazzle, there are questions of architecture, power, and society that cannot be solved with a simple model update.
