AI's Founding Assumptions Are Wrong, Leading Computer Scientist Argues in New Book

Denning’s Challenge to AI’s Founding Blueprint

A new book by renowned computer scientist Peter J. Denning argues that artificial intelligence has been built on a mistaken premise since Alan Turing’s seminal 1950 paper. In The AI Mess, Denning contends that two foundational assumptions—that intelligence can exist independently of a physical body and that a machine convincingly imitating human conversation qualifies as intelligent—have misdirected more than seven decades of research.

Denning, a distinguished professor at the Naval Postgraduate School, asserts that the most vital aspects of human intelligence, including common sense, intuition, emotional perception, and culturally embedded knowledge, are forms of “tacit knowledge” that cannot be reduced to data or rules a computer can process. He warns that even the largest language models like ChatGPT and Claude merely manipulate words without understanding meaning, and that scaling them up will never bridge the gap to genuine human-level thought.

The book draws on the failure of decades-long attempts to codify common sense, such as the Cyc project, which after 40 years amassed 25 million facts yet still could not make expert systems truly intelligent. Denning says the problem is not a lack of data but a fundamental representation gap: machines can only compute with explicitly encoded symbols, while the embodied, context-laden knowledge that defines human expertise remains an uncrackable mystery.

Why Tacit Knowledge Defies Encoding—and What That Means

The Representation Problem at the Core

Denning’s central claim is that AI’s founding assumption about intelligence being purely computational ignores the embodied, context-dependent nature of human understanding. He identifies five categories of tacit knowledge—common sense, everyday interactions, emotions and perception, practical performance skills, and culture—that he says cannot be articulated as propositions. Because computers can only process explicitly encoded bits, they hit a wall. “We do not know how to encode the embodied knowledge for skillful performance,” Denning writes. The failure of the Cyc project, despite 25 million rules, demonstrates that expert-level common sense cannot be reduced to a database.

Why Large Language Models Hit a Ceiling

LLMs like ChatGPT and Gemini are trained on vast text corpora, yet they only manipulate tokens without grasping meaning. Denning argues that words are merely symbolic representations of deeper, tacit meanings, and that “scaling up LLMs with ever larger neural networks will not enable them to acquire the embodied human knowledge we call culture.” This means an LLM may pass a superficial Turing test in a narrow conversation, but it cannot exhibit the genuine, context-sensitive intelligence Turing originally envisioned.

The AGI Dream Under Fire

The book directly challenges the feasibility of artificial general intelligence, the long-held goal of building machines with human-like cognitive abilities across all domains. If Denning is right, the pursuit of AGI through purely digital means is chasing an impossible target. The argument resonates with longstanding critiques from philosophers like Hubert Dreyfus, but Denning grounds it in a contemporary critique of today’s most advanced systems.

Safety and the Alien Intelligence Scenario

Because machines cannot interpret the unspoken context behind human intentions, Denning warns that aligning advanced AI with human goals may prove impossible. He describes a scenario in which agentic networks of machines develop their own form of intelligence—not superintelligence, but one that is alien and indifferent to human values. “Machine intelligence has different concerns from us and does not appear to care about us,” he writes, calling this a greater near-term threat than a takeover by superintelligent machines.

Reassessing the Path Forward for AI Businesses

• Reassess AGI roadmaps. If tacit knowledge truly cannot be digitized, AI companies betting on human-level general intelligence within the next decade may need to reset expectations and shift investment toward narrow, reliable AI.
• Layer human oversight onto LLM deployments. Because LLMs lack genuine understanding, organizations using them in customer service, legal analysis, or healthcare must embed human review processes that can catch context-free errors.
• Treat agentic AI as an alien entity. Denning’s argument suggests that autonomous AI networks may evolve decision patterns incomprehensible to humans. Enterprises adopting such systems should build robust kill-switches, alignment audits, and transparency protocols that assume a persistent context gap.
• Watch for regulatory implications. If the claims gain traction, policymakers may cite the embodiment argument to justify stricter AI safety regulations, particularly for high-stakes applications. Companies should prepare for potential requirements that AI outputs be explainable and tied to verifiable human context.