Why early AI researchers deliberately made machines less predictable

A dependable calculator should not surprise its owner. A useful collaborator sometimes should. A new study by MIT historian Benjamin Lindquist examines how that difference shaped early thinking about artificial intelligence: some researchers deliberately sought irregularity in machines built to execute reliable calculations.[1]

Published in Past & Present on 19 October, The Irrational Computer draws on archival records, technical reports, internal memoranda and public discussion. Lindquist argues that influential British and American engineers came to treat controlled disorder as a possible ingredient of invention, rather than merely as noise to eliminate. The accessible journal abstract establishes that argument; the full archival study was not available for this article.[1]

Unexpected answers were once part of the attraction

In an MIT interview on 23 October, Lindquist discusses the influence of ideas about art and discovery. He describes Bell Labs engineer John Pierce’s interest in using mathematics to introduce irregularity into art, rather than simply imposing order. He also contrasts researchers who associated creativity with unexpected departures with those who considered it an extension of ordinary, well-executed reasoning.[2]

That contrast is more interesting than a simple story of machines gradually becoming more accurate. The target itself was disputed. A system designed to calculate a known answer and a system designed to suggest an unanticipated possibility could reasonably be judged by different standards.

Lindquist’s interview makes the reversal explicit: some early advocates valued neural networks partly because they could produce strange or unexpected answers. He contrasts that laboratory ambition with contemporary commercial settings, where unexpected outputs can have legal or monetary consequences. This is his historical interpretation, not a controlled measurement of today’s systems.[2]

Surprise and truth are separate dimensions

The useful modern distinction is between generating a possibility and endorsing it. An invented bridge design could be worth exploring before anyone knows whether it is structurally sound. An invented bridge inspection certificate would be unacceptable. Both are novel strings of information; the difference is the claim being made and the work required before anyone relies on it.

Randomness itself does not settle that distinction. A procedure could randomly choose between two correct proofs. A completely deterministic procedure could return the same incorrect calculation every time. These simple counterexamples show why a historical connection between randomness and creativity cannot, on its own, establish that factual errors are inevitable or necessary for useful invention.

Nor does a surprising answer automatically count as an original idea. A useful proposal has to survive selection: it must fit the problem, withstand criticism or prove fruitful when tested. Generating candidates and checking them are separate jobs. Combining them into one fluent conversational response can hide the point at which a suggestion has been promoted into an assertion.

The design question the history leaves open

One practical implication is to give exploratory and evidential tasks different interfaces. A brainstorming tool could make alternatives easy to inspect and compare. A tool reporting a measurement should preserve its source, method and uncertainty. That is an analytical proposal prompted by the history, not a finding that Lindquist has experimentally tested.

The historical study restores a question that enthusiasm for ever-higher accuracy can obscure: what kind of thinking is the machine supposed to assist? For some tasks, deviation is the beginning of useful work. For others, it is a defect. Better AI design does not require choosing surprise over reliability everywhere; it requires making clear when an answer is an invitation to investigate and when it is being offered as something established.

Report an error or suggest a correction