In 1972, a rather different idea of computing emerged from Europe. It did not concern speed, or hardware, or even efficiency in the usual sense. It concerned reasoning.
The language was called PROLOG, short for Programming in Logic. It was developed by Alain Colmerauer and Philippe Roussel in Marseille, with its theoretical spine supplied by Robert Kowalski in Edinburgh. Between them, they attempted something that had largely eluded computing to that point. They tried to make a machine prove things.
Up to then, most programming had been procedural. One instructed the machine step by step, rather like briefing a junior clerk. Do this, then that, and do it precisely in this order. PROLOG declined the entire premise. It asked the programmer not how to solve a problem, but what the problem was.
The distinction, though subtle on paper, was profound in practice.
In PROLOG, knowledge is written as a set of facts and rules. The machine is then asked questions, and it searches for answers by applying those rules logically. It does not follow a fixed path. It explores possibilities, retraces its steps when necessary, and continues until it either finds a solution or exhausts the space.
This mechanism, known as backtracking, gave PROLOG an unusual quality for its time. It could appear to reason.
A simple illustration makes the point. One might define relationships such as parenthood, and from these derive more distant relations without ever specifying the steps explicitly. The machine, presented with the right query, would infer the answer. Not because it had been told how, but because the structure of the rules compelled it.
It is here that PROLOG departed most sharply from its contemporaries, particularly LISP, which had dominated early artificial intelligence work. LISP excelled at symbolic manipulation, but it remained procedural at heart. PROLOG, by contrast, treated logic itself as the programming model.
For a time, this seemed rather promising.
The language found a natural home in areas where rules and relationships mattered more than arithmetic. Early work in natural language processing used PROLOG to parse sentences and represent meaning in structured form. Expert systems, then enjoying their brief commercial ascendancy, relied on similar rule-based reasoning to emulate specialist knowledge in medicine, engineering, and finance.
In Japan, during the ambitious Fifth Generation Computer Project of the 1980s, PROLOG was elevated to something approaching national strategy. It was intended to form the backbone of a new class of machines designed for knowledge processing rather than mere calculation. The ambition was formidable. The results, less so.
PROLOG’s strengths were also its constraints.
Logical inference is elegant, but it is rarely fast. As the number of rules grows, the search space expands, and the system can become sluggish, even unwieldy. The same combinatorial difficulties noted elsewhere in artificial intelligence made themselves felt here too. What works neatly in a small domain can become intractable in a large one.
There were other inconveniences. Programmers trained in conventional methods often found PROLOG disorienting. One does not instruct the machine. One describes a world and asks questions of it. This requires a different habit of mind, and not everyone cared to acquire it.
By the 1990s, attention had shifted. Statistical methods and, later, machine learning began to displace purely symbolic approaches. Systems that learned from data proved more adaptable than those that relied entirely on handcrafted rules. PROLOG, while never abandoned, retreated into more specialised roles.
Yet its influence has not disappeared.
Modern artificial intelligence, for all its reliance on data and probability, still grapples with questions of reasoning, structure, and explanation. In areas such as knowledge representation, automated theorem proving, and certain branches of legal and formal reasoning, the ideas embodied in PROLOG remain relevant. Quietly, persistently, they endure.
It would be an error to regard PROLOG as obsolete. It is, rather, part of an older lineage within artificial intelligence, one that treats intelligence not as pattern recognition, but as logic made operational.
That lineage never quite vanished. It simply fell out of fashion.
And fashions, as experience tends to show, have a habit of returning when least expected.