According to a fresh analysis, the potential for artificial intelligence (AI) to replicate human-like thinking may be fundamentally flawed due to the inability to program the most vital components of human intelligence into machines.
Peter J. Denning, a distinguished computer scientist, critiques a notion proposed by Alan Turing, considered the father of modern computer science, suggesting that this idea has misdirected AI research for the past 75 years. In his recent work, "Turing’s Mistake: Escaping the Yoke of Unintelligent Machines," Denning explores Turing’s assertions from 1950 that posited human intelligence could exist outside the physical body, possibly enabling its recreation as software on a computer.
Denning challenges the approach claiming that machine intelligence can be gauged through an imitation game, commonly recognized as the Turing test. He argues that accepting these two notions has led to the chaotic state of AI development observed today.
The current AI systems under development, he contends, are unlikely to achieve human-level intelligence, often referred to as artificial general intelligence (AGI). Instead, he warns that such systems could pose substantial risks without ever matching human cognitive abilities.
At the core of Denning’s argument lies the concept of tacit knowledge, which encompasses the vast understanding that humans carry but cannot fully articulate or convert into a format that machines can comprehend. He identifies five categories of tacit knowledge that he believes evades machine learning: common sense, interactions with people and surroundings, emotions and perceptions, practical skills, and shared cultural and historical contexts across societies.
For decades, researchers have sought ways to codify common sense for computer use. Douglas Lenat’s ambitious Cyc project, which began in the 1980s, aimed to create an extensive database of common-sense facts, eventually amassing 25 million entries after four decades. However, Denning points out that this extensive repository did not yield a sufficient base of common sense for expert systems to function intelligently, indicating that much of what constitutes human expertise cannot be boiled down to simple propositions.
Denning elaborates on the challenge of practical skills, asserting that while it is often feasible to describe outcomes ('know what'), encoding the embodied knowledge necessary for performance ('know how') remains elusive. He uses music as an example, noting that a virtuoso violinist may deliver exquisite performances yet lack the ability to articulate how to achieve that mastery.
The core issue, which Denning labels 'the representation problem,' stems from computers' need for data and instructions in recognizable formats. Tacit knowledge, however, is difficult to translate into such structures. He adds that words are mere symbols representing meanings, and prevalent Large Language Models like ChatGPT, Claude, and Gemini manipulate these words without comprehending their underlying meanings.
Denning highlights the significance of context, illustrating how statements can vary widely in meaning based on the speaker’s emotions—whether they are being sincere, sarcastic, or playful. Understanding the context often involves tracing back through prior conversations, creating an endless, intricate network of meanings.
Culture represents another complex hurdle, encompassing values, social norms, histories, and group dynamics. Denning believes that merely enlarging Large Language Models will not enable them to acquire the embodied knowledge integral to understanding culture and context. Consequently, he states that LLMs will likely fail to fulfill the Turing test's objective: achieving machine thought indistinguishable from human thought.
He ultimately articulates a mutual incomprehension between humans and machines, noting that while artificial neural networks may be capable of forming their unique type of tacit knowledge, humans may remain unable to interpret it. This divide poses significant challenges for AI safety. Denning cautions that if machines cannot grasp the nuanced context of human directives, aligning their behavior with human objectives may become unfeasible.
As AI automates tasks, networks of machines are likely to cultivate their own types of intelligence—ones that differ from human intelligence but can still threaten human well-being. Denning argues that the most pressing danger isn't an advanced superintelligent machine but rather a network of less advanced systems acting in unpredictable and potentially harmful manners.
He urges a recognition of the fading familiar culture as intelligent machines integrate into society, emphasizing the need to retain our humanity and assert what distinguishes us from machines. Celebrating these distinctions is crucial as we navigate this emerging landscape of AI technology.
