When early AI researchers treated randomness as a source of invention

Some early AI researchers wanted machines to surprise them. In a new Past & Present paper, MIT historian Benjamin Lindquist traces efforts to introduce randomness into systems otherwise associated with calculation and logical rules. Unpredictable output could be evidence of invention, in this view, rather than simply a defect to eliminate.[1][2]

Lindquist describes an intellectual setting shaped by modernist art, psychology and theories of discovery. In an MIT News interview, he points to Bell Labs engineer John Pierce, who saw mathematical irregularity as a way to disrupt artistic convention. Early neural-network researchers also hoped that modelling productive mistakes could reveal something about human thinking.[2]

The competition was between two images of the mind

One approach to AI sought to represent knowledge as facts and logical operations. Another drew inspiration from networks of simple units that could learn and generate patterns not specified in advance. Lindquist’s account treats the disagreement as partly a dispute over creativity itself: is an original idea the result of reasoning especially well, or of departing from established reasoning?[1][2]

The distinction changed what a successful demonstration looked like. A system designed to reproduce orderly reasoning could be valued for consistency. A machine intended to model inventive association might instead be interesting because it produced an unexpected connection. Surprise had a different role depending on the theory of mind that the researcher was trying to build.[1][2]

The interview also follows changes in the standing of neural networks: early enthusiasm, a shift towards symbolic approaches and their continued development in cognitive science before later gains from larger datasets and computing resources. Lindquist’s argument is that ideas about irregularity and invention persisted through those technical changes, rather than appearing for the first time with today’s generative models.[2]

A historical argument about what counts as intelligence

Lindquist draws a connection between that history and present tensions over unpredictable AI output. It is an interpretation of intellectual origins, not an experiment showing that a particular modern system must make factual errors to be creative. The historical question is why researchers valued departures from an expected answer; the practical question is when a user has asked for one.[1][2]

A novelist seeking an unusual metaphor and a traveller checking a departure time have different standards for a useful answer. The first may welcome an unexpected association; the second needs a verifiable fact. Reading the history with those purposes in mind avoids treating every surprising output as the same kind of achievement.[1][2]

The account makes a familiar word—intelligence—less settled. Early designers chose which human abilities to imitate and what evidence would count as success. Contemporary labels such as assistant, agent and collaborator make similar choices about expected behaviour. Lindquist’s history asks us to notice those assumptions before treating one ideal of machine intelligence as inevitable.[1][2]

The lesson is that the history of AI includes a deliberate search for productive unpredictability. We should preserve room for discovery where it genuinely helps while refusing to confuse invented evidence with originality. A useful machine may be capable of surprise, but it still needs to know when the job calls for a fact.[1][2]

Report an error or suggest a correction