AI Pioneers Injected Randomness to Model Human Creativity

Early AI pioneers who helped shape neural networks deliberately introduced randomness into algorithms to model human creativity, according to a new paper by Lindquist published in Past and Present and discussed with MIT News. The research traces modern AI hallucinations to this legacy, arguing that creativity and predictability remain in tension in systems intended to mimic the messy human mind. In the 1950s, computer scientists and psychologists argued that chance and irregularity were essential to invention. Bell Labs engineer John Pierce proposed that mathematics could introduce disorder into art, breaking cliches and enabling unexpected insight. Artificial neural networks became a tool for this effort, not because their answers were always correct, but because they could produce bizarre or unanticipated outputs. Psychologists hoped such systems would open a black box of the mind and expose productive mistakes rather than simply calculate logical rules. This view contrasted with symbolic AI, which treated intelligence as rule-based knowledge represented in facts and propositions. Although neural networks gained early traction, symbolic methods dominated for decades. Frank Rosenblatt, an influential psychologist behind early neural network models, died in 1971, and his loss may have weakened the neural-network camp. The approach persisted in cognitive science before gaining renewed strength in the 1980s and later becoming dominant as internet data and computing power expanded. Around 2012, neural networks began outperforming statistical machine-learning methods. Today, large language models such as ChatGPT and Claude are valued for reliable, correct answers, a reversal from their early promise. Researchers and companies now have practical, legal and commercial reasons to suppress hallucinations and strange outputs. Lindquist argues that this has sanded down the creative irregularity once prized by AI pioneers. The historical lesson is that reducing hallucinations may be possible, but it may come at the cost of inventive, unexpected thought. Efforts to make AI more predictable can unintentionally remove the very quality that made early neural networks seem lifelike.