For 75 years, artificial intelligence research has operated on two ideas introduced by Alan Turing in 1950. A prominent computer scientist now argues those ideas were mistaken, and that the consequences of following them have shaped AI development in ways that may be fundamentally limiting. This is not a minor technical critique. It is a challenge to the conceptual foundations of the entire field.
Peter J. Denning, the computer scientist behind this argument, lays out his case in a book titled Turing’s Mistake: Escaping the Yoke of Unintelligent Machines. His core claim is that two assumptions Turing made have never been seriously questioned, and that both are wrong.
The first assumption: intelligence can exist independently of a physical body, and therefore be recreated in software. The second: a machine can demonstrate intelligence by convincingly imitating a human in conversation, which became the basis of the famous Turing test. Denning’s position is direct. Accepting these claims without scrutiny, he writes, has led to what he calls “the AI mess in which we find ourselves today.”
The Knowledge That Cannot Be Typed Into a Computer
The center of Denning’s argument is tacit knowledge. This is the vast category of human understanding that resists being put into words or encoded into data. It is not ignorance. It is the opposite: the deep, embodied competence that allows a person to navigate the world without being able to fully explain how.
Denning identifies five major categories of tacit knowledge that machine learning cannot capture: common sense, everyday interactions with people and the physical environment, emotions and perception, practical performance skills, and the social and historical knowledge embedded in culture.
The challenge of encoding common sense is not new. One of the most sustained attempts was the Cyc project, initiated by Douglas Lenat in the 1980s. Its goal was to build a comprehensive database of common sense facts. After four decades of work, the project contained roughly 25 million entries. That is an enormous undertaking by any measure. Yet Denning’s assessment is unsparing: even that scale of effort could not produce a system with genuine expert-level understanding. The project, he argues, actually validated the problem rather than solving it. Much of what makes human experts capable simply cannot be articulated as propositions.
Practical skills present an even sharper version of this problem. Denning draws a distinction between “know what” and “know how.” Descriptions of skillful outcomes can be stored as data. The embodied knowledge required to actually perform a skill cannot. A virtuoso violinist can produce extraordinary music but cannot fully explain to a student how to replicate it. Even if a robot could observe and imitate a skilled musician, it would have no access to what the musician feels while playing, or what an audience feels while listening. The performance and the experience of the performance are inseparable from having a biological body.
Why Scaling Up Does Not Solve the Problem
This is where Denning’s argument becomes most directly relevant to the current moment in AI development. Large language models such as ChatGPT, Claude, and Gemini are capable of producing fluent, contextually appropriate text. They are also, according to Denning, doing something fundamentally different from understanding.
His framing is precise: words are symbolic representations of meanings, not the meanings themselves. LLMs manipulate words. They do not know what the words mean. Behind every word, Denning argues, is a deep well of tacit knowledge that gives it meaning, and that well is inaccessible to any system that operates purely on encoded symbols.
Context compounds the problem. Human intelligence depends on an endless, fractal web of prior conversations, cultural assumptions, and social history. Recognizing sarcasm, knowing when to be diplomatic, interpreting an ambiguous request: all of these require access to layers of context that are never fully explicit. Culture, in Denning’s framing, encompasses values, norms, judgments, history, communities, moods, and relationships involving power and care. Scaling up neural networks with more data and more parameters does not give a system access to any of that. It gives it more patterns over the same kind of encoded symbols.
What This Means Beyond the Technical Debate
Here is what most coverage of AI capability debates misses: the stakes are not primarily about whether AI will pass the Turing test. They are about safety and alignment in a world where AI systems are already operating autonomously at scale.
Denning’s concern is not the science fiction scenario of superintelligent machines taking over. His concern is more grounded and, arguably, more urgent. Agentic networks of machines are likely to develop their own forms of machine intelligence that, while not reaching human general intelligence, are still capable of creating serious problems. If machines cannot interpret the unspoken context behind human intentions, reliably aligning advanced AI systems with human goals may prove impossible. Humans cannot read machine tacit knowledge. Machines cannot read human tacit knowledge. Denning describes this as “aliens across an uncrossable divide.”
The practical implication is significant. AI systems can be powerful, useful, and genuinely valuable without being intelligent in the way humans are. Treating them as if they are, or building toward a version of them that is, may be the wrong direction entirely.
In Short
Turing’s two foundational assumptions, that intelligence is substrate-independent and that imitation equals intelligence, have guided AI research for 75 years. Denning argues both are wrong. The core obstacle is tacit knowledge: the embodied, contextual, cultural understanding that humans carry but cannot fully articulate, and that no amount of data or computational scale can encode. This does not mean AI is useless. It means the field may have been optimizing for the wrong goal, and that the risks of continuing to do so are more practical than theoretical.
Based on reporting from ScienceDaily AI.