Artificial intelligence has advanced at a remarkable pace, from large language models that can generate essays and computer code to agentic systems capable of coordinating tasks across digital environments. Yet a new argument from computer scientist Peter J. Denning suggests that AI may be pursuing an impossible objective: machines that appear increasingly capable but never truly understand the human world.
In his new book, Turing’s Mistake: Escaping the Yoke of Unintelligent Machines, Denning argues that modern AI has inherited two flawed assumptions from Alan Turing’s influential 1950 paper, “Computing Machinery and Intelligence.” Turing proposed replacing the difficult question “Can machines think?” with the more operational “imitation game,” now widely known as the Turing test.
Denning’s challenge is that the test blurred a crucial distinction: the difference between imitating intelligence and possessing understanding. According to Denning, Turing’s ideas encouraged the belief that intelligence could be separated from the human body and recreated in software, and that a machine’s ability to imitate human conversation could be treated as evidence of intelligence.
This, Denning argues, has shaped three-quarters of a century of AI research. His central claim is stark: artificial general intelligence, or AGI, may not be achievable because the most important parts of human intelligence cannot be encoded into machines.
At the centre of Denning’s critique is the concept of tacit knowledge. This is the vast domain of human understanding that is difficult, and perhaps impossible, to formalise. It includes common sense, practical know-how, intuition, emotional recognition, cultural awareness, and the subtle contextual judgements people make every day. Humans know how to interpret a raised eyebrow, a sarcastic remark, an awkward silence, or the difference between a joke and an insult. Much of this knowledge is not stored as explicit rules. It is embodied, social, historical, and learned through participation in human life.
Denning argues that machine learning systems cannot capture five major categories of tacit knowledge: common sense, everyday interactions with people and environments, feelings and perception, performance skills, and the social and historical knowledge embedded in culture. This matters because the current AI boom relies heavily on systems that process patterns in language and data. Large language models can generate fluent and convincing text, but Denning argues that they manipulate symbols rather than grasp meanings. In his view, words are not the same as the human experiences and assumptions that give words significance.
The history of AI contains repeated attempts to encode common sense into machines. One of the most ambitious was Douglas Lenat’s Cyc project, which began in the 1980s and sought to build a large database of everyday knowledge. Denning notes that after decades of effort, even millions of formal entries could not amount to the background understanding needed to make expert systems truly expert. This is not simply a question of scale. Adding more rules or training data may not solve the problem if the missing ingredient is not information but lived experience.
