Turing’s Mistake: Prominent Computer Scientist Challenges 75 Years of AI Foundations
The trajectory of artificial intelligence research, long guided by the visionary principles of Alan Turing, may have been directed toward a fundamental misunderstanding of the nature of intelligence itself. Peter J. Denning, a distinguished professor of computer science at the Naval Postgraduate School and a former president of the Association for Computing Machinery (ACM), has issued a provocative critique of the industry’s foundational dogmas. In his latest work, Turing’s Mistake: Escaping the Yoke of Unintelligent Machines, Denning argues that the pursuit of human-like cognition in software is hampered by two core assumptions made by Turing in 1950 that have led the field into a theoretical and practical "mess."
Denning’s thesis centers on the rejection of the idea that intelligence can be detached from a physical, biological body and the belief that imitation of human conversation is a sufficient metric for true intelligence. As the global tech industry pours trillions of dollars into the development of Artificial General Intelligence (AGI), Denning’s critique serves as a sobering reminder that the current path may not only be flawed but could potentially introduce systemic risks that the industry is ill-equipped to manage.
The Historical Context: Turing’s 1950 Legacy
To understand Denning’s critique, one must look back to the origins of the field. In October 1950, Alan Turing published his seminal paper, "Computing Machinery and Intelligence," in the journal Mind. In this paper, Turing famously bypassed the difficult philosophical question "Can machines think?" by replacing it with a practical test: "Can machines do what we (as thinking entities) can do?"
This inquiry birthed the "Imitation Game," later known as the Turing Test. Turing’s core premise was built on two pillars: first, that intelligence is a function of information processing that can exist independently of a physical body (disembodied intelligence), and second, that if a machine can simulate human-like responses in a text-based conversation, it must be considered intelligent.
For 75 years, these pillars have served as the North Star for AI development. From the early logic-based systems of the 1960s to the modern transformer architectures of today’s Large Language Models (LLMs), the goal has remained consistent: to build a machine that mimics human cognitive output. However, Denning argues that this focus on output—the "imitation"—ignores the essential "being" and "knowing" that defines human intelligence.
The Tacit Knowledge Problem: The Limits of Encoding
At the heart of Denning’s argument is the concept of "tacit knowledge," a term popularized by philosopher Michael Polanyi, who famously stated, "We know more than we can tell." Denning identifies a fundamental barrier in the transition from human experience to machine code. He asserts that a vast majority of human understanding is non-propositional; it cannot be reduced to a set of rules, facts, or data points.
Denning categorizes five major areas where tacit knowledge remains inaccessible to machines:
- Common Sense: The intuitive understanding of how the world works, which humans acquire through physical existence.
- Everyday Interaction: The seamless navigation of social environments and physical spaces.
- Emotions and Perception: The internal states that color every human thought and decision.
- Practical Performance Skills: The "know-how" of physical mastery, such as playing an instrument or performing surgery.
- Social and Historical Context: The deep-seated cultural narratives and historical weight that give meaning to language.
While machines excel at "know-what"—the storage and retrieval of facts—they are fundamentally incapable of "know-how." Denning uses the example of a virtuoso violinist. While a computer can analyze the frequencies of the notes and a robot could be programmed to move a bow across strings, the machine cannot feel the music, nor can it explain the intuitive physical adjustments a master makes in response to the acoustics of a room or the mood of an audience.
The Failure of Symbolic AI and the Cyc Project
Denning’s critique is supported by the historical outcomes of "Good Old-Fashioned AI" (GOFAI). In the 1980s, researchers believed that if they could simply encode enough facts into a computer, common sense would emerge. The most ambitious of these efforts was Douglas Lenat’s Cyc project.
Launched in 1984, Cyc aimed to create a comprehensive ontology and knowledge base of human common sense. For forty years, researchers manually entered over 25 million assertions, such as "trees are usually outdoors" or "you cannot be in two places at once." Despite this massive effort, Cyc failed to produce a system that could truly understand the world. Denning points to this as empirical evidence that expertise is not merely a collection of propositions. The "representation problem"—the inability to translate the richness of embodied experience into symbolic logic—remains the graveyard of many AI ambitions.
The Illusion of Meaning in Large Language Models
The rise of generative AI and models like ChatGPT, Claude, and Gemini has reignited the belief that the Turing Test is within reach. These models are capable of producing prose that is indistinguishable from human writing. However, Denning argues that this is merely a more sophisticated version of Turing’s "imitation."
Denning explains that LLMs operate purely on the level of syntax and probability, not semantics. They manipulate symbols (words) based on statistical patterns found in massive datasets, but they possess no "well of meaning" behind those symbols. When a human says "I am tired," the words are a symbolic representation of a physical, biological state. When an LLM generates the same phrase, it is simply predicting that "tired" is a statistically likely follower to "I am" in a specific context.
"Words are but symbolic representations of meanings, not the meanings themselves," Denning writes. This distinction is critical for AI safety and reliability. Because the machine does not understand the consequences or the intent of the words it produces, it is prone to "hallucinations"—confidently stating falsehoods—and lacks the ethical grounding that comes from social and cultural embodiment.
The Fractal Nature of Context and Culture
A significant portion of Denning’s work focuses on the role of context. Human intelligence is not a static property; it is a dynamic process that occurs within a specific environment. Context allows for the interpretation of sarcasm, irony, and subtle social cues. Denning describes context as "fractal," meaning every conversation rests on previous conversations, which in turn rest on historical and cultural foundations.
Machines, lacking a history of social interaction and a biological stake in the world, cannot access this fractal depth. Culture—encompassing values, power dynamics, and shared moods—is something humans live within. For an AI, "culture" is just more data to be processed, leading to a superficial mimicry that falls apart when faced with novel or high-stakes social situations.
Implications for AI Safety: The "Alien" Intelligence
Perhaps the most urgent aspect of Denning’s argument concerns the future of AI safety. The current discourse on AI risk often focuses on "superintelligence"—the fear of a machine becoming so smart that it takes over the world. Denning argues this is another symptom of Turing’s mistake. The real danger is not a superintelligent machine, but an "alien" intelligence.
Denning posits that as we build agentic networks of machines, they will develop their own forms of "machine intelligence" that are entirely disconnected from human values and tacit understanding. "Machines cannot read our tacit knowledge and we cannot read theirs," he notes. This creates an "uncrossable divide."
If a machine cannot grasp the unspoken context of human intentions, it cannot be truly aligned with human goals. This "alignment problem" becomes exacerbated when we grant machines autonomy in critical infrastructure, such as power grids, financial markets, or weapon systems. We are creating "unintelligent machines" that possess immense power but zero wisdom, operating on a logic that is fundamentally alien to the human experience.
A Call to Reassert Humanity
The conclusion of Denning’s critique is a call for a paradigm shift in how society views and interacts with technology. He suggests that the "AI automation singularity" can be avoided if we stop trying to make machines our peers and instead recognize them as the tools they are.
Denning encourages a refusal to submit to the "yoke" of low-intelligence machines. This involves:
- Acknowledging the Divide: Accepting that machine intelligence and human intelligence are fundamentally different and perhaps incompatible.
- Valuing Embodiment: Prioritizing human roles in fields that require tacit knowledge, such as healthcare, education, and creative arts.
- Design for Safety: Building systems that are subservient to human judgment rather than trying to replicate it.
As the 75th anniversary of Turing’s paper approaches, Denning’s Turing’s Mistake offers a timely intervention. By deconstructing the foundational myths of the AI field, he provides a framework for a future where technology serves humanity without attempting to replace the inimitable qualities that make us human. The "AI mess," as Denning calls it, is not an inevitability but a result of a specific philosophical path—one that we may now need to abandon in favor of a more grounded, embodied understanding of intelligence.
