Renowned computer scientist Peter J. Denning has raised provocative questions regarding the foundational ideas of artificial intelligence (AI) as put forth by Alan Turing, suggesting that these concepts may have misled AI research for over 75 years. In his latest publication, "Turing's Mistake: Escaping the Yoke of Unintelligent Machines," Denning identifies two pivotal assumptions made by Turing in 1950 that continue to influence the field of AI today. First, Turing posited that intelligence might exist separately from a physical form, implying it could be replicated in computer software. Second, he suggested that a machine could demonstrate intelligence through its ability to imitate human conversation, which led to the establishment of the Turing test.
Denning contends that these assertions have been impactful yet detrimental. "These two claims have shaped much of AI research and development," he states, adding that accepting these ideas has contributed significantly to the current challenges facing the AI landscape.
He critiques the pursuit of artificial general intelligence (AGI)—machines capable of human-like intelligence—as likely futile. Instead, Denning foresees that the technologies being developed could introduce severe risks.
Central to Denning's argument is the concept of tacit knowledge, which represents the vast, complex understanding that humans possess but is not easily articulated or quantifiable by machines. He identifies five key categories of tacit knowledge that machine learning struggles to encompass: common sense, interpersonal interactions in various contexts, emotional responses and perceptions, practical skills, and cultural contexts embedded within social practices.
Historically, researchers have attempted to codify common sense into comprehensive databases. A well-regarded initiative was Douglas Lenat's Cyc project, which, after 40 years, amassed around 25 million entries aimed at assembling a repository of common sense knowledge. However, Denning points out that even this extensive collection has fallen short of creating systems with sufficient expertise. "Cyc validated that much of the knowledge that makes people experts cannot be articulated as propositions," he remarks.
He underscores that practical skills represent an even more formidable challenge than common sense. "Our performance capabilities across numerous domains elude machine translation," he states. For instance, a skilled musician may create extraordinary sounds but struggle to articulate the nuances of their technique to a learner.
Denning extends the scope of tacit knowledge to include intuition, creativity, and spontaneity, which remain largely inaccessible to artificial systems. He attributes this limitation to what he terms the "representation problem." Computers function effectively only when they can operate within the confines of data and instructions explicitly programmed into them. Tacit knowledge, however, does not lend itself easily to such rigid formats.
"Every word has an undercurrent of tacit knowledge that gives it substance," Denning explains, emphasizing that commonly used AI models, like ChatGPT and others, manipulate language without truly grasping the meanings behind their words.
This situation creates a substantial divide in understanding. Denning notes, "Given that we have yet to fully comprehend how tacit knowledge operates in human beings, replicating it in machines remains an insurmountable challenge." He refers to this mystery of how tacit knowledge is embodied and processed within humans as a complex enigma.
Moreover, Denning highlights the essential role of context in shaping intelligence. Context enriches the meaning of words and actions, influencing how humans interpret humor, sincerity, and emotional subtleties. He explains that the origins of contextual assumptions can be traced through an endless chain of preceding conversations, each influenced by its specific context.
Culture, too, presents a significant barrier for AI. Denning characterizes culture as a tapestry of values, norms, judgments, histories, and relationships. "Human interactions are imbued with underlying assumptions that lend significance to the spoken word," he asserts.
He argues that merely increasing the size of language models does not enable them to acquire the nuanced understanding of human culture needed for meaningful communication. Thus, he concludes that achieving the goal of the Turing test—rendering machine thought indistinguishable from human thought—remains beyond reach.
In his closing remarks, Denning expresses concern over the potential divergence between human and machine forms of tacit knowledge. "Machines cannot interpret our tacit knowledge, just as we cannot comprehend theirs," he writes, highlighting the implications for safety in AI.
He warns that misalignment of advanced AI systems with human objectives could become a significant risk. "As machines automate processes, they may develop a form of intelligence distinct from human cognition that could present formidable challenges without reaching superintelligence levels," he explains.
Denning stresses the urgency of reflecting on our relationship with emerging AI technologies. "Accepting that familiar societal norms may dissolve as intelligent machines integrate into our daily lives is crucial," he states. "We must refuse to conform to machine logic or subordinate our humanity to them, reaffirming what sets us apart and embracing those distinctions."
