The foundational pillars of modern artificial intelligence are currently under intense scrutiny as Peter J. Denning, a Distinguished Professor at the Naval Postgraduate School and a pioneer in computer science, challenges the core assumptions that have guided the field since its inception. In his latest work, Denning posits that the trajectory of AI research has been fundamentally misdirected for three-quarters of a century, rooted in a 1950 thesis by Alan Turing that may have fundamentally misunderstood the nature of human intelligence. Denning’s critique arrives at a pivotal moment when the global tech industry is investing hundreds of billions of dollars into Large Language Models (LLMs) and the pursuit of Artificial General Intelligence (AGI), suggesting that the "AI mess" currently facing society is a direct result of these early conceptual errors.
The Dual Assumptions of 1950
To understand Denning’s critique, one must return to Alan Turing’s seminal 1950 paper, "Computing Machinery and Intelligence." In this work, Turing proposed what he called the "Imitation Game," now universally known as the Turing Test. Denning identifies two specific assumptions within this paper that he believes have led researchers astray. The first is the "decoupling" of intelligence from physical form—the idea that the human mind is essentially software that can be run on different hardware, such as a silicon-based computer. The second is the "behavioral" definition of intelligence, which suggests that if a machine can mimic human conversation well enough to deceive a human judge, it must be considered intelligent.
Denning argues that these assumptions created a "yoke" that has constrained AI development within a narrow, representational framework. By focusing on imitation rather than actual understanding, and by ignoring the necessity of a biological body, the field has produced systems that are sophisticated at statistical prediction but devoid of the "tacit knowledge" that defines human existence.
The Paradox of Tacit Knowledge
At the center of Denning’s argument is a concept popularized by philosopher Michael Polanyi: "We know more than we can tell." This is known as the tacit knowledge problem. Human intelligence is not merely a collection of facts and rules; it is a deeply embodied experience. Denning categorizes the forms of knowledge that machines cannot capture into five distinct areas: common sense, everyday social interactions, emotions and perception, practical performance skills, and the deep historical and social knowledge embedded in culture.
While modern AI can process trillions of tokens of text, it lacks the "know-how" of a human practitioner. Denning uses the example of a virtuoso violinist. A computer can be programmed with the "know-what"—the notes, the tempo, and the mathematical frequencies of a masterpiece. However, it cannot replicate the "know-how"—the subtle, unteachable physical adjustments and emotional resonance that a human musician brings to a performance. This knowledge is not stored in bits; it is stored in the nervous system, the muscles, and the lived history of the individual.
A Chronology of the Quest for Machine Intelligence
The history of AI is a recurring cycle of high expectations followed by "AI Winters" when those expectations fail to materialize. Denning’s critique provides a framework for understanding why these cycles occur.
- 1950: Alan Turing publishes his paper, setting the stage for functionalism and the imitation-based approach to intelligence.
- 1956: The Dartmouth Workshop marks the formal beginning of AI as a field. Early pioneers like John McCarthy and Marvin Minsky believe that every aspect of learning or intelligence can be so precisely described that a machine can be made to simulate it.
- 1980s: The rise of "Expert Systems" attempts to code human expertise into "if-then" rules. This era saw the launch of the Cyc project by Douglas Lenat, an ambitious attempt to codify every piece of human common sense.
- 2010s: The "Deep Learning" revolution shifts the focus from symbolic logic to neural networks. Machines begin to excel at pattern recognition in images and text.
- 2020s: The emergence of Generative AI and LLMs. Systems like GPT-4 demonstrate an unprecedented ability to mimic human prose, leading many to believe that Turing’s goal of AGI is within reach.
Denning points to the Cyc project as a cautionary tale. Despite 40 years of work and 25 million entries of common-sense facts, the system never achieved the fluid understanding of a five-year-old child. This failure, Denning argues, validates the idea that expertise cannot be reduced to propositions.
The Representation Problem and the Limits of Scaling
The current AI boom is built on the premise that "scaling" is the solution. The industry believes that by adding more parameters, more data, and more computing power, machines will eventually develop emergent properties of true reasoning. Denning disputes this, citing the "representation problem."
Computers operate through symbols and calculations. However, human language is not just a set of symbols; it is a shorthand for a vast "deep well" of tacit meaning. When an LLM generates a sentence, it is calculating the most probable next word based on statistical patterns in its training data. It does not "know" what a "mother," a "home," or "grief" feels like. It has no biological feedback loops, no hormones, and no mortality. Because the machine lacks the human context that gives words their weight, it remains an "alien" intelligence—one that can simulate the surface of our communication without ever touching the substance.
The Fractal Nature of Context and Culture
Intelligence is not a standalone property; it is highly dependent on context. Denning explains that human context is "fractal" and endless. Every conversation rests on previous conversations, which in turn rest on cultural norms, historical events, and personal relationships. Humans navigate this complexity effortlessly because they are embedded in the culture.
AI systems, conversely, are "decontextualized." They are trained on a static snapshot of the internet—a "scraping" of human output that is disconnected from the living, breathing environment where that output was created. Denning argues that scaling up neural networks will never bridge this gap because culture is not a data set; it is a lived practice. As a result, LLMs will never truly pass the Turing Test in a way that signifies actual thought; they will only become more convincing "stochastic parrots."
Implications for AI Safety and Alignment
One of the most pressing concerns in contemporary tech policy is the "Alignment Problem"—the challenge of ensuring that AI goals match human values. Denning’s analysis suggests that the alignment problem may be unsolvable within the current paradigm. If machines and humans are "aliens across an uncrossable divide," we cannot expect a machine to understand the unspoken intentions and ethical nuances that guide human behavior.
Denning warns of a specific type of risk: the "automation singularity." This is not a scenario where a superintelligent AI takes over the world, but rather a world where "agentic networks of machines" with low-level, alien intelligence begin to manage critical infrastructure, legal systems, and social interactions. Because these machines do not share our tacit knowledge or our concern for human well-being, their "problem-solving" could lead to outcomes that are logically consistent but humanly disastrous.
Analysis of Broader Impacts
The implications of Denning’s thesis are far-reaching, affecting everything from economic policy to educational standards. If Denning is correct, the massive capital investment in AGI may be a speculative bubble destined to burst when the limits of scaling are finally reached.
From a labor perspective, Denning’s focus on "practical performance skills" suggests that human workers who rely on embodied knowledge—surgeons, artisans, plumbers, and high-level negotiators—may be more resilient to automation than previously thought. Conversely, professions that have been reduced to "symbolic representation" and "know-what" (such as basic data analysis or standardized legal drafting) are at the highest risk of being subsumed by what Denning calls "unintelligent machines."
The response from the AI community has been polarized. Some researchers, particularly those in the "connectionist" camp, argue that the brain itself is a biological computer and that tacit knowledge is simply a form of pattern recognition we haven’t yet mastered. However, a growing number of "embodied cognition" advocates agree with Denning, suggesting that true AI will require robots that interact with the physical world in the same way biological organisms do.
Reasserting the Human Element
Denning concludes by urging a radical shift in how society relates to technology. He calls for a refusal to submit to the "yoke" of machines that are fundamentally less intelligent than we are. This involves re-centering human judgment in decision-making processes and recognizing that our unique ability to navigate the uncodifiable aspects of life—intuition, imagination, and empathy—is our greatest strength.
The "AI mess" we find ourselves in today is, in Denning’s view, a crisis of identity. By accepting Turing’s premise that we are essentially machines, we have allowed ourselves to be treated like machines. Denning’s work serves as a manifesto for the "reassertion of humanity," a call to celebrate the differences between silicon and soul, and a warning that if we do not escape the yoke of unintelligent machines, we may lose the very qualities that make us human.
Related posts:
