The Philosophical Foundations of a 75-Year Detour
The genesis of the current AI trajectory can be traced back to the mid-20th century. In 1950, Alan Turing proposed what he called the "Imitation Game," now universally known as the Turing Test. His hypothesis was twofold: first, that intelligence is a function of information processing that can be replicated in software, independent of biological hardware; and second, that if a machine could pass as a human in a text-based exchange, it should be considered "intelligent."
Denning argues that these two claims have become the "yoke" under which researchers have labored, leading to the pursuit of Artificial General Intelligence (AGI) that he believes is fundamentally unattainable. By treating intelligence as a purely symbolic or mathematical exercise, Denning suggests that the industry has ignored the essential nature of human cognition, which is deeply rooted in physical experience and social context. He argues that our collective "acquiescence" to these claims has created the "AI mess" of the 2020s, characterized by systems that are proficient at pattern recognition but entirely devoid of understanding.
The Tacit Knowledge Problem: Why Machines Cannot "Know"
At the center of Denning’s critique is the concept of "tacit knowledge," a term popularized by philosopher Michael Polanyi, who famously noted that "we know more than we can tell." Denning identifies five specific categories of human understanding that he believes are fundamentally resistant to digitization: common sense, everyday social interactions, emotional perception, practical performance skills, and the deep-seated historical knowledge embedded in human culture.
While modern Large Language Models (LLMs) like GPT-4 or Claude 3.5 can simulate the appearance of this knowledge by predicting the next token in a sequence, Denning asserts they lack the "embodied" component required for true expertise. He distinguishes between "know-what" (propositional knowledge) and "know-how" (performative knowledge). A computer can store the sheet music and a technical description of a violin concerto, but it cannot possess the embodied "know-how" of a virtuoso who feels the tension of the strings and the resonance of the wood.
A Chronology of the Search for Common Sense
The attempt to bridge this gap is not new. Denning highlights the historical efforts to encode common sense into machines, most notably the Cyc project. Launched in 1984 by Douglas Lenat, Cyc was intended to be the definitive "encyclopedia" of common sense.
- 1984: Project Cyc begins with the goal of codifying millions of "rules of thumb" that humans use to navigate the world.
- 1990s: Researchers realize that common sense is not a static list of facts but a dynamic system of inferences.
- 2010s: Despite containing over 25 million assertions and 40 years of manual labor, Cyc fails to produce a machine that can navigate a simple social interaction with the nuance of a child.
- 2020s: The industry shifts toward "Connectionism" and LLMs, hoping that massive scale will allow machines to "absorb" common sense through statistical correlation.
Denning argues that the failure of Cyc was not a lack of data, but a validation of the "representation problem." Much of what makes a human an expert cannot be articulated in propositions. Therefore, it cannot be encoded into the bits and bytes that constitute a machine’s reality.
The Representation Problem and the Illusion of Meaning
Denning’s analysis delves into the technical limitations of digital computation. Computers, by definition, operate on physical representations of symbols. For a machine to process a concept, that concept must be translated into a form the hardware can manipulate. However, Denning argues that words are merely symbolic placeholders for meanings that exist in a "deep well" of tacit human experience.
"Words are but symbolic representations of meanings, not the meanings themselves," Denning writes. This distinction is vital for understanding the current state of AI safety and alignment. When an LLM generates a response, it is not drawing from a reservoir of understanding; it is navigating a high-dimensional mathematical space of word frequencies. This creates a fundamental divide: humans host tacit knowledge in their bodies and histories, while machines host statistical weights in their circuits. Because scientists cannot yet explain how humans host this knowledge, they have no blueprint for translating it into a machine-readable format.
Context, Culture, and the Fractal Nature of Intelligence
A significant portion of Denning’s work focuses on the role of context. Human intelligence relies on an endless, "fractal" chain of contexts—previous conversations, cultural norms, and historical precedents—that allow for the interpretation of sarcasm, humor, and sincerity.
Denning argues that culture is not just a dataset to be scraped; it is a lived experience involving power dynamics, care, and social judgments. He contends that scaling up neural networks will never allow a machine to acquire culture because culture is embodied. Without a biological body to feel the weight of social consequences or the warmth of human connection, a machine remains an "alien" entity. This lack of context is why AI systems frequently "hallucinate" or provide technically correct but socially disastrous advice.
Implications for AI Safety and Human Alignment
The most urgent aspect of Denning’s critique concerns AI safety. Many researchers, including figures like Nick Bostrom and Eliezer Yudkowsky, have warned of "superintelligent" machines taking over the world. Denning, however, identifies a more immediate and insidious threat: the "low-intelligence" machine that is given agency without understanding.
Because machines and humans exist on opposite sides of an "uncrossable divide," aligning their goals is not a matter of better programming, but a fundamental impossibility. If a machine cannot interpret the unspoken context of human intent, it cannot be trusted to act on our behalf in complex scenarios. Denning warns that "agentic networks" of machines—autonomous systems that can make decisions and interact with one another—are likely to develop their own form of "machine intelligence" that is entirely alien to human values.
This threat is not one of a "Terminator-style" uprising, but rather a slow erosion of human agency. As we delegate more decisions to systems that do not "care" about us or understand the "why" behind our needs, we risk being governed by a "yoke" of unintelligent logic.
Reasserting Humanity in the Age of Automation
Denning’s conclusion is both a warning and a call to action. He suggests that the "AI automation singularity" is not an inevitable milestone of progress, but a choice that society can still decline. To live safely with these machines, he argues, humans must first accept that our familiar culture is changing and then actively reassert the qualities that distinguish us from software.
Industry reactions to Denning’s thesis are likely to be polarized. Proponents of AGI, such as leaders at OpenAI and Google DeepMind, maintain that emergent properties in larger models will eventually overcome the tacit knowledge barrier. However, Denning’s perspective aligns with a growing school of "AI skeptics" and "human-centric" researchers who argue that the current path leads toward a "stochastic" future where meaning is replaced by probability.
The broader impact of Denning’s argument suggests a need for a paradigm shift in AI regulation and development. Rather than chasing the ghost of human-like AGI, Denning advocates for a focus on "intelligence" that serves human needs without pretending to be human. By recognizing the limits of what can be encoded, society can begin to build systems that are truly "tools" rather than "masters," finally escaping the 75-year-old yoke of Turing’s foundational mistake.
Related posts:
