Gary Marcus vs. Nature: Why Talk of AGI's Arrival Is Premature

Gary Marcus and his colleagues Walter Quattrociocchi and Valerio Capraro have criticized a recent article in the prestigious journal Nature that claimed artificial general intelligence (AGI) had been achieved. The authors of the response argue that the impressive performance of large language models (LLMs) on various benchmarks, and even on complex math problems, is not evidence of genuine intelligence. In their view, those who claim AGI exists make a fundamental error: they confuse the ability to perform narrowly specialized tasks with truly general intelligence. The piece is a call for greater care in the use of terminology and for a deeper, more considered analysis of what we mean by "intelligence."

Increasingly, people are claiming that artificial general intelligence already exists. Perhaps the most recent and high-profile such claim appears in an article published in Nature. These claims are often fueled by impressive achievements of large language models (LLMs), whose results show strong performance on a range of test datasets, fluent work across diverse domains and, in some cases, even correct solutions to open math problems. These developments are often taken as irrefutable proof that humanity has reached the threshold of artificial general intelligence.

However, as Marcus and his co-authors rightly note, such interpretations rest on a fundamental conflation: they equate results on individual, often well-studied and standardized tasks with intelligence itself. Performing individual tasks, even with impressive results, cannot be considered sufficient proof of general intelligence. In their paper, the authors show that recent claims of AGI rest on a conceptual error: confusing statistical approximation, which is growing ever more sophisticated, with intelligence itself. They also argue that recent claims (for example, those published by Chen et al., 2026) of an alleged success in building AGI depend on redefining what the term "AI" has historically meant.

The authors' central argument is that today's LLMs, despite their remarkable abilities, are essentially very complex statistical machines. They are trained on enormous datasets and learn to predict the next word or sequence of words based on probabilistic patterns. This lets them generate text, answer questions and even solve problems that require a certain level of logic or knowledge. But according to Marcus and his colleagues, this does not mean the model "understands" the task in the human sense, or that it can transfer knowledge and skills to entirely new, unforeseen situations, a key aspect of general intelligence. They argue that it looks more like a highly advanced form of imitation or approximation than like real thinking.

The consequences of this confusion could be significant. Premature claims of AGI can lead to excessive optimism, misallocated resources and, more importantly, to underestimating the real problems and challenges of building genuine artificial intelligence. They can also make us less critical about the capabilities and limitations of existing AI systems, on the assumption that these systems have a level of understanding they do not actually have. Moreover, they can slow research aimed at building more reliable, interpretable and truly intelligent systems.

In conclusion, Gary Marcus and his co-authors urge the scientific community and the general public to take a more sober and critical approach to evaluating progress in artificial intelligence. They stress the importance of distinguishing the impressive statistical abilities of LLMs from genuine general intelligence, which implies the ability to reason, learn, adapt and understand across a wide range of contexts. Until we reach that level, claims that AGI has arrived should be regarded as premature and based on a misreading of the data.