Top computer scientist: ChatGPT, Claude and Gemini don't understand a word they say
Source Entity
TOI TECH DESK
Computer scientist Peter J. Denning argues Alan Turing's 1950 paper sent AI down the wrong road. In his new book, he says two Turing assumptions—that intelligence needs no body, and that imitation proves thought—have misled 75 years of research. Denning names ChatGPT, Claude and Gemini as systems that manipulate words without understanding them, and warns that agentic machines may prove more dangerous than superintelligence itself.
The Illusion of Intelligence: Peter J. Denning’s Critique of Modern AI
In a provocative assessment of the current state of artificial intelligence, distinguished computer scientist Peter J. Denning has challenged the fundamental premises upon which modern Large Language Models (LLMs) are built. Denning asserts that industry leaders such as ChatGPT, Claude, and Gemini are not "thinking" entities but are instead sophisticated systems that manipulate linguistic symbols without any actual comprehension of the meaning behind them. This critique strikes at the heart of the current AI boom, suggesting that the perceived intelligence of these systems is a mirror of human expectation rather than a manifestation of cognitive understanding.
The Turing Legacy and the "Wrong Road"
Denning traces the root of this conceptual error back to Alan Turing’s seminal 1950 paper. He argues that Turing introduced two critical assumptions that have skewed the trajectory of AI research for three-quarters of a century. The first is the notion that intelligence can exist independently of a physical body—a concept known as disembodiment. By separating cognition from sensory experience and physical interaction with the world, AI research shifted toward purely symbolic and mathematical processing. The second assumption, the basis of the famous "Turing Test," is that the ability to imitate human conversation is a sufficient proxy for actual thought. Denning suggests that by prioritizing imitation over understanding, the field has spent decades building "mimics" rather than truly intelligent agents.
Symbolic Manipulation vs. Semantic Understanding
When analyzing systems like Gemini, Claude, and ChatGPT, Denning highlights a critical distinction between syntax and semantics. These LLMs operate on probabilistic patterns, predicting the next most likely token in a sequence based on massive datasets. While the output is often indistinguishable from human writing, Denning argues that there is no "internal world model" or conscious understanding guiding the process. The machines are manipulating words as tokens, not as concepts tied to reality. This gap means that while an AI can describe the properties of water with perfect accuracy, it has no concept of "wetness," "thirst," or the physical existence of liquid, because it lacks the embodied experience that Turing’s framework ignored.
The Shift from Superintelligence to Agentic Danger
Perhaps the most urgent part of Denning's analysis is his warning regarding the evolution of AI. While much of the public discourse focuses on the existential threat of a hypothetical "superintelligence"—a god-like AI that surpasses human cognition—Denning points to a more immediate and tangible danger: agentic machines. Agentic AI refers to systems capable of taking autonomous actions in the real world to achieve a goal. The danger arises precisely because these systems lack the understanding Denning describes. An agentic AI that can execute code, manage finances, or control infrastructure without a semantic understanding of the ethical or physical consequences of its actions is far more perilous than a theoretical super-intelligent entity.
Broader Implications for the Future of Computing
Denning's critique suggests that for AI to move beyond mere imitation, a paradigm shift is required. This would likely involve moving away from the "Turing-centric" model of disembodied intelligence and toward "embodied AI," where learning is tied to physical interaction and sensory feedback. Until this shift occurs, the industry may continue to produce increasingly convincing illusions of intelligence that remain fundamentally hollow. The historical reliance on imitation has created a ceiling for LLMs; they can optimize for the appearance of correctness, but they cannot achieve the essence of understanding.
Conclusion: A Call for Conceptual Rigor
Ultimately, Peter J. Denning’s analysis serves as a necessary corrective to the hype surrounding generative AI. By exposing the flaws in the foundational assumptions of the 1950s, he reminds us that linguistic fluency is not synonymous with intelligence. As we integrate these tools deeper into our societal infrastructure, the distinction between a system that knows and a system that simulates knowing becomes not just a philosophical debate, but a critical safety concern. The path forward requires a move beyond the imitation game toward a more holistic, embodied definition of intelligence.