To see what road we might have traveled, we must return to the scene of AI’s birth. From its origins in the mid-1950s—just a decade after ENIAC, the first digital computer, was built at the University of Pennsylvania—the AI research community made no secret that the kind of machine intelligence it sought to create was teleological: oriented toward attaining a specific goal, or telos.
Take the General Problem Solver, a software program developed in 1957 with support from the RAND Corporation. Its creators—Herbert A. Simon, Allen Newell, and J. C. Shaw—used a technique called “means-ends analysis” to create a so-called “universal” problem solver. In reality, the problems the software could tackle had to be highly formalized. It worked best when goals were clearly defined, the problem-solving environment was stable (meaning the rules governing the process were fixed from the start), and multiple iterations allowed for trying out a variety of means to achieve the desired ends.
Of course, this “rules-based” paradigm of AI research eventually lost out to a rival approach based on neural networks—the basis of all modern machine learning, including the large language models (LLMs) powering systems like ChatGPT. But even then, the nascent neural network approach was framed in problem-solving terms. One of the envisioned applications of the Perceptron, an early neural network designed for pattern recognition, was military: sifting through satellite imagery to detect enemy targets. Neural networks required a clearly defined target and trained models to achieve that task. Without a specific goal or a clear history of prior attempts at achieving it, they wouldn’t work.
I think it is not a coincidence that early AI tools closely mirrored the instrumental reason of clerical and administrative workers in the very institutions—government, corporate, and military—that spearheaded AI research. These were workers with limited time and attention, for whom mistakes carried significant costs. Automating their tasks through machines seemed both a logical next step and an efficient way to reduce errors and expenses. Some of this focus on goals can be traced to funding imperatives; early AI needed to prove its practical value, after all. But a deeper reason lies in AI’s intellectual inheritance from cybernetics—a discipline that shaped much of its early agenda but was sidelined as AI sought to establish itself as a distinct field.
The pioneers of cybernetics were fascinated by how feedback-powered technologies—ranging from guided missiles to thermostats—could exhibit goal-directed behavior without conscious intention. They drew analogies between these systems and the teleological aspects of human intelligence—such as lifting a glass or turning a door handle—that allow us to achieve goals through feedback control. In appropriating this cybernetic framework, AI carried the metaphor further. If a thermostat could “pursue” a target temperature, why couldn’t a digital computer “pursue” a goal?