In 1948 Alan Turing wrote a report titled “Intelligent Machinery” for the National Physical Laboratory, where he was then working. His director, Sir Charles Darwin, dismissed it as a “schoolboy essay” and it went unpublished for two decades. It is now recognized as one of the earliest and most prescient discussions of machine intelligence - written before the term “artificial intelligence” existed.
The report sketched ideas that would take the field years to rediscover. Turing proposed what he called “unorganised machines”: networks of simple, neuron-like binary elements wired together largely at random, with no built-in program. Such a machine, he argued, could be trained - by a process he likened to the education of a child, applying reward and punishment - to perform useful tasks, and with enough elements could be made to compute anything a universal machine could. This is recognizably a description of trainable neural networks and of learning by reinforcement, a decade before Rosenblatt’s perceptron. Turing also discussed search and what he called “genetical” search, an early gesture toward evolutionary computation.
“Intelligent Machinery” shows Turing thinking about intelligence not only as the symbol-manipulating logic of his universal machine but as something that could grow from a randomly connected substrate through learning. Both threads - symbolic computation and trainable networks - run through the rest of AI’s history, and Turing had already glimpsed the second one in 1948.