Loading field map
The question "Can machines think?" emerged from the intersection of mid-20th-century computing, cybernetics, and philosophy. It quickly unraveled into deeper puzzles: What is thinking? Could a non-biological system ever possess consciousness? And what would it mean for a machine to genuinely understand, rather than merely simulate, intelligence? The philosophy of artificial intelligence developed as a series of frameworks that alternately defined, defended, challenged, and transformed the terms of this debate.
The earliest frameworks were forged in the 1950s, when digital computers first demonstrated symbolic reasoning. The Computational Theory of Mind (also called computationalism) proposed that thinking itself is a form of computation: the mind is to the brain as a running program is to a computer. This claim provided a powerful philosophical foundation for AI, because it suggested that any suitably programmed physical system—not just a biological brain—could instantiate genuine thought. The Turing Test, proposed by Alan Turing in 1950, offered a behavioral criterion: if a machine could hold a conversation indistinguishable from a human's, it would be reasonable to call it intelligent. The test sidestepped the metaphysical question of inner experience and focused on observable performance.
These ideas converged in Symbolic AI, the dominant research program from the 1950s through the 1980s. Symbolic AI treated intelligence as the manipulation of formal symbols according to explicit rules—a physical symbol system. It assumed that all cognition could be captured in logical representations and search algorithms. This framework coexisted with a crucial distinction: Strong AI held that a properly programmed computer would literally be a mind, possessing understanding and consciousness, while Weak AI maintained that such a machine could only simulate mental states without genuinely experiencing them. The Strong/Weak divide was not merely a technical nuance; it defined the philosophical stakes. Strong AI took the Computational Theory of Mind to its logical conclusion, while Weak AI preserved a gap between simulation and reality.
By the 1970s and 1980s, two powerful critiques exposed deep weaknesses in the foundational settlement. Hubert Dreyfus's Dreyfus Critique, articulated in his 1972 book What Computers Can't Do, drew on phenomenology (especially Heidegger and Merleau-Ponty) to argue that human intelligence depends on embodied, situated, tacit know-how that cannot be captured in explicit rules. Dreyfus contended that Symbolic AI's reliance on context-free representations ignored the role of the body, background practices, and non-propositional skills. His critique did not merely point out technical limitations; it challenged the very assumption that thinking is a form of symbol manipulation.
John Searle's Chinese Room Argument (1980) targeted the Computational Theory of Mind more directly. Searle imagined himself inside a room, following rules to manipulate Chinese symbols without understanding Chinese. He argued that syntax alone—the manipulation of symbols according to formal rules—could never produce semantics, or genuine meaning. The Chinese Room was a direct assault on Strong AI: if a program could pass the Turing Test by manipulating symbols without understanding, then passing the test was not evidence of genuine thought. Searle's argument did not refute Weak AI, which accepted mere simulation, but it forced defenders of Strong AI to explain how computation could give rise to intentionality.
These critiques did not destroy the earlier frameworks, but they fractured the consensus. The Computational Theory of Mind remained influential, but it now had to contend with the charge that it confused syntax with semantics. Symbolic AI continued to produce practical results, but its philosophical ambitions were sharply curtailed. The Dreyfus Critique and the Chinese Room Argument remain live positions today, each continuing to inform debates about the limits of machine intelligence.
In response to the perceived failures of Symbolic AI, alternative frameworks emerged that reimagined the nature of cognition and computation. Connectionism, which gained traction in the 1980s, modeled intelligence using artificial neural networks—large numbers of simple processing units operating in parallel, learning from data rather than following explicit rules. Connectionism revived an older, non-symbolic approach to computation and offered a direct alternative to Symbolic AI. Where Symbolic AI treated cognition as rule-governed symbol manipulation, Connectionism treated it as the emergence of patterns from distributed, sub-symbolic processes. This shift narrowed the scope of the Computational Theory of Mind: if thinking was not necessarily symbolic, then the mind might be computational in a different, non-classical sense.
Embodied Cognition, emerging in the 1990s, radicalized the Dreyfus Critique by arguing that cognition is not merely situated but constitutively dependent on the body and its interactions with the environment. Embodied Cognition rejected the central assumption of both Symbolic AI and classical computationalism: that thinking can be abstracted away from the physical substrate. Instead, it proposed that perception, action, and the physical structure of the organism shape thought in fundamental ways. This framework did not replace Connectionism; rather, it coexisted with it, sometimes complementing neural-network models by emphasizing the role of sensorimotor loops and environmental feedback.
Beginning around 2000, the philosophy of AI broadened to address the real-world deployment of intelligent systems. Ethics of AI emerged as a distinct framework concerned with the values embedded in AI design, the distribution of responsibility, fairness, transparency, and the social impacts of automation. Unlike earlier frameworks that focused on the nature of mind, Ethics of AI asked what it means to build systems that affect human lives. It drew on traditions from applied ethics, political philosophy, and the philosophy of technology, and it often operated in a pluralistic mode, coexisting with both Weak AI and Strong AI without resolving their metaphysical disagreements.
AI Alignment, which crystallized around 2010, narrowed the ethical focus to a specific technical-philosophical problem: how to ensure that advanced AI systems reliably pursue the goals their designers intend, especially when those systems are capable of learning and adapting. Alignment differs from broader AI ethics by concentrating on goal specification, value instillation, and the risk of unintended consequences from highly capable agents. It inherits the assumption of Weak AI—that machines are tools—but pushes it into new territory by asking whether a sufficiently powerful tool could resist human control.
Epistemology of AI, also emerging around 2010, addresses the knowledge claims produced by opaque machine-learning systems. When a neural network makes a diagnosis or a recommendation, what epistemic status does its output have? Can we trust a system whose reasoning we cannot reconstruct? This framework splits from both Ethics and Alignment by focusing on the nature of AI-generated knowledge, the problem of algorithmic opacity, and the conditions under which we can legitimately claim to know something on the basis of a machine's output. It revives classic epistemological questions about justification, reliability, and understanding in a new technological context.
Today, the philosophy of AI is a field of living disagreements and productive coexistence. The leading frameworks—Connectionism (now dominant in the form of deep learning), Embodied Cognition, Ethics of AI, AI Alignment, and Epistemology of AI—do not share a single paradigm. They agree on at least one point: the early dream of Symbolic AI as a complete theory of mind has not been realized. But they disagree sharply on what follows. Connectionists argue that neural networks provide the correct model of cognition, while Embodied Cognition insists that cognition cannot be understood without the body. Ethics of AI and Alignment often assume a Weak AI stance, treating machines as instruments, but they diverge on whether the primary risk is unfairness or loss of control. Epistemology of AI challenges both by asking whether opaque systems can ever be genuine sources of knowledge.
The older frameworks remain active. The Computational Theory of Mind persists in philosophy of mind, though it has been transformed by connectionist and embodied critiques. The Turing Test is still discussed as a benchmark, though few now take it as a sufficient condition for thought. Strong AI has become a minority position, defended mainly by those who believe that future architectures (perhaps whole-brain emulation or artificial general intelligence) will vindicate it. Weak AI remains the default working assumption for most AI practitioners and many philosophers. The Dreyfus Critique and the Chinese Room Argument continue to be taught and debated, serving as permanent reminders that the question "Can machines think?" cannot be answered by engineering alone.