Chatbot errors are usually discussed as an engineering defect: a model states something false with confidence, and the vendor promises a fix in the next release. Benjamin Lindquist, a historian at MIT, argues in a paper in the journal Past and Present that the behaviour has deeper roots. Some of the field's founders, he writes, wanted machines that would sometimes surprise them, and deliberately built chance into systems meant to solve problems.
In an interview with MIT News, Lindquist traces the idea to a long tradition linking chance with discovery and to engineers such as John Pierce at Bell Labs, who argued in the early 1950s that mathematics could be used to break order in art and push it away from cliché. Many early neural networks were built by psychologists who hoped a model of the brain would show how minds make productive mistakes. To those advocates, the value of a network lay in its capacity for strange answers; they already had machines that gave expected ones. A rival camp saw creativity as ordinary thinking done better, and its symbolic, rule-based approach dominated for decades.
Lindquist also notes the role of chance in the field's history. Frank Rosenblatt, a leading early champion of neural networks, died on his 43rd birthday in 1971, as interest was already moving towards symbolic AI. Networks survived in cognitive science and began to outpace other statistical methods around 2012, once data and computing power caught up.
His point about the present is the useful one for anyone choosing between ChatGPT, Claude and their rivals. Early systems lived in laboratories, so odd output cost nothing. Commercial chatbots now carry legal and financial risk, and Lindquist says vendors have sanded down the strangeness, suppressing it rather than removing it. If he is right, pressing errors down further may also press down the unexpected connections people sometimes value, and settings that trade predictability for variety will remain a real choice rather than a temporary bug.
Read it for what it is. This is a historian's argument about intentions and ideas, not an empirical study, and it does not show that reducing factual errors must cost creativity in today's models. Researchers who treat hallucination as a measurable failure with practical fixes, such as retrieval or better calibration, would dispute any suggestion that it is inevitable. As a corrective to the idea that errors are simply a bug about to be patched, it is worth the time.