GOFAI

• 1. Introduction

  ›

  GOFAI, short for good old-fashioned artificial intelligence, is a term the philosopher John Haugeland introduced in 1985, in his book Artificial Intelligence: The Very Idea. Haugeland used it to name classical symbolic AI, the approach set against neural networks, situated robotics, narrow symbolic AI and neuro-symbolic AI. He coined the phrase to press two questions. Could this kind of artificial intelligence ever produce a genuinely intelligent machine? And is it actually the method the human brain uses to think? Those two questions reach far beyond computer science, into philosophy and psychology. Together they set the terms for a long argument about what reasoning really is.

• 2. The First Yes

  ›

  In 1963, the AI pioneer Herbert A. Simon speculated that the answer to both of Haugeland's questions was yes. Simon pointed to the performance of two programs he had co-written, the Logic Theorist and the General Problem Solver, as his evidence. He backed this with his own psychological research into how humans solve problems.

  Across the 1950s and 60s, AI research of this kind reshaped intellectual history far beyond computer labs. It helped spark the cognitive revolution and led to the founding of cognitive science as an academic field. It also became the central example behind the philosophical theories of computationalism, functionalism and cognitivism in ethics. The same research anchored the psychological theories of cognitivism and cognitive psychology. That legacy would eventually collide with a much narrower definition used inside AI labs themselves.

• 3. From Philosophy To Practice

  ›

  Within AI development today, engineers use GOFAI more narrowly, for programs built from deliberate, explicit instructions written for one specific task. This use of the term sets it apart from approaches built on machine learning instead of hand-written rules.

  AlphaGo and Apple's initial design for Siri are cited as examples of GOFAI applications in this practical sense. That contrast between hand-written rules and machine learning sits inside a much older argument about where human reasoning itself comes from.

• 4. A Rationalist Inheritance

  ›

  Haugeland places GOFAI inside the rationalist tradition of Western philosophy, which treats abstract reason as the highest human faculty. That faculty, in this tradition, is what separates people from animals. This idea appears in Plato and Aristotle, in Shakespeare, Hobbes, Hume and Locke, and it stood at the center of the Enlightenment. It resurfaces in the logical positivists of the 1930s and in the computationalists and cognitivists of the 1960s.

  Shakespeare captured that faith directly: 'What a piece of work is a man, How noble in reason, how infinite in faculty... In apprehension how like a god, The beauty of the world, The paragon of animals.'

  Symbolic AI in the 1960s succeeded at simulating high-level reasoning: logical deduction, algebra, geometry, spatial reasoning and means-ends analysis. It expressed all of this in precise English sentences, much like the ones people used themselves. Many observers, philosophers, psychologists and the AI researchers among them, became convinced they had captured the essential features of intelligence itself. This was not simple hubris. Rationalism itself entailed that conclusion, since denying it would call much of the Western philosophical tradition into question. That confidence would not go unquestioned for long.

• 5. The Continental Objection

  ›

  Continental philosophy, running through Nietzsche, Husserl and Heidegger, rejected rationalism outright. These thinkers argued that high-level reasoning is limited and prone to error. They believed most of what people can do comes from intuition, culture and an instinctive feel for a situation.

  Hubert Dreyfus and Haugeland were the philosophers most familiar with this continental tradition, and they became among the first critics of GOFAI. They questioned whether it was truly sufficient for genuine intelligence. Their skepticism would eventually force a much closer look at what Haugeland's claims for GOFAI actually were.

• 6. Defining The Undefinable

  ›

  Even Haugeland's supporters and critics, across philosophy, psychology and AI research, have struggled to pin GOFAI down precisely. Drew McDermott, for instance, called Haugeland's description of GOFAI 'incoherent' and argued that GOFAI itself is a 'myth'.

  Haugeland coined the term to test what he called the claims essential to all GOFAI theories, which he reduced to two points. First, our capacity to deal with things intelligently comes from our capacity to think about them reasonably, including sub-conscious thinking. Second, that capacity for reasonable thought amounts to a faculty for internal, automatic symbol manipulation.

  This claim closely tracks the physical symbol system hypothesis Herbert A. Simon and Allen Newell proposed in 1963: 'A physical symbol system has the necessary and sufficient means for general intelligent action.' It also resembles what Hubert Dreyfus called his 'psychological assumption': 'The mind can be viewed as a device operating on bits of information according to formal rules.'

  For Haugeland, the 'symbols' inside GOFAI are discrete physical things assigned a definite meaning, things like <cat> and <mat>, not raw signals or unidentified numbers. That definition excludes the zeros and ones of digital machinery. It also leaves out older techniques like cybernetics, perceptrons, dynamic programming and control theory, plus newer ones like neural networks and support vector machines.

  Haugeland's two questions ask whether GOFAI alone is sufficient for general intelligence. For that reason his definition also excludes neuro-symbolic AI and narrow symbolic AI systems built to solve only one specific problem. That narrow, sufficiency-focused definition is exactly what a later generation of AI scientists would push back against.

• 7. The Reply

  ›

  Russell and Norvig, writing about the critique from Dreyfus and Haugeland, noted that 'the technology they criticized came to be called Good Old-Fashioned AI (GOFAI).' They added that, in their words, GOFAI 'corresponds to the simplest logical agent design.' Capturing every contingency of appropriate behavior in necessary and sufficient logical rules is genuinely difficult, they wrote, a problem they named 'the qualification problem.'

  After the 1980s, symbolic AI work grew more robust, adding probabilistic reasoning, non-monotonic reasoning and machine learning to handle open-ended domains. Most AI researchers today expect deep learning, or more likely a synthesis of neural and symbolic approaches, to be needed for general intelligence. That synthesis already has a name: neuro-symbolic AI.