Promises Ahead of Their Time: Three AI Predictions and Their Fate

At a press conference organised by the US Navy in 1958, the researcher Frank Rosenblatt presented the "Perceptron", and the New York Times described it as the "embryo of an electronic computer" that the Navy expected would be able to walk, talk, see, write, reproduce itself and be conscious of its existence. The device actually built, the Mark I Perceptron at the Cornell Aeronautical Laboratory, was completed in 1960 and could learn and recognise letters. Between the description and the device lies a distance worth measuring.

The Perceptron: an idea that came back

In 1969 Marvin Minsky and Seymour Papert showed that a single-layer network cannot learn the exclusive-or (XOR) function. The source notes that the book was widely misread, and that its reputation contributed to a decline of interest in, and funding for, neural networks. These networks then returned to the mainstream in the 1980s, which some see as confirming Rosenblatt's expectations. The idea endured; its time came later. For the two winters the field went through, see the article "The History of AI: Waves of Enthusiasm and Two Winters of Disappointment".

The chess prediction: ten years became forty

Herbert Simon and Allen Newell set out their estimates in a paper published in Operations Research in January–February 1958, after an address Simon gave in 1957. It said that "within ten years a digital computer will be the world's chess champion, unless the rules bar it from competition", and that it would discover and prove "an important new mathematical theorem". It added that the range of problems computers can handle would become "coextensive with the range to which the human mind has been applied". The deadline was missed: IBM's Deep Blue beat Garry Kasparov 3.5 to 2.5 in May 1997, some forty years later, evaluating 200 million chess positions a second. Simon was right about the direction and wrong about the time. The gap remains between excelling at a narrow game and the broad flexibility of a human being, the subject of "Narrow and General AI: Why the Yardstick Misleads Us".

The radiology prediction: timing and path both missed

In 2016 Geoffrey Hinton said: "People should stop training radiologists now", judging it obvious that AI would outperform human radiologists within five years. Radiology Business reported in May 2025, summarising a New York Times report, that the Mayo Clinic in Rochester had grown its radiology staff by 55% since 2016, to about 400, and that a report by the American College of Radiology projects the specialty's physician supply to grow by 26% over thirty years. This does not mean the systems have not improved. It is more likely that the move from good performance on one task to dispensing with a profession passes through decisions, regulation and accountability, the logic of tasks explained in "A Job Is a Bundle of Tasks: Which Occupations Fade and Which Change?".

How should we read promises?

An editorial lesson from the three episodes:

• Distinguish direction from timing: the first may prove right while the second is wrong by decades.
• Ask for measurement, not description: what does the device do today, and by which yardstick?
• Keep your decision adjustable, and do not build it on a single date given by a speaker.

Caution here is not pessimism. Exaggerated promises harm the researcher when expectations are dashed as much as they harm the listener who believes them. What tomorrow brings is with God alone, and it falls to us to take the means: to measure and to review.