Generative AI traces back to mid‑20th‑century engineers who argued that a machine capable of human‑like thought must contain some irrationality. Benjamin Lindquist, an MIT historian, explains in a new paper that these pioneers deliberately injected randomness into otherwise deterministic algorithms to mimic creative thinking, linking this practice to the origin of AI hallucinations.\ \ Why did they do this? Many scientists and artists have long claimed that chance plays a crucial role in discovery and creativity. Bell Labs engineer John Pierce famously suggested that mathematics is valuable in art because it can remove order and inject unexpected bursts, breaking clichés and prompting insight. To achieve such “irregularity,” researchers used artificial (random) neural networks. Early neural nets were built by psychologists who believed that modeling the brain would let scientists observe productive mistakes, turning the mind’s “black box” into a transparent system.\ \ At the same time, another camp argued that creative thinking is simply better ordinary thinking. The debate boiled down to two simplistic heuristics—randomness equals creativity versus normal thinking done better—and how to reconstitute human inventiveness in machines.\ \ Compared with the past, today’s AI (e.g., ChatGPT, Claude) is valued for giving mostly correct answers. Lindquist notes that early advocates prized neural networks precisely because they sometimes produced surprising, even bizarre, outputs. Modern systems are constrained by legal and commercial pressures, leading to a “sanding down” of strangeness that was once a feature of AI.\ \ Historically, symbolic AI dominated for decades, treating the world as a set of rules to be encoded as facts and propositions. Neural networks, with their stochastic problem‑solving, gained attention in the 1950s but fell out of favor after the charismatic pioneer Frank Rosenblatt died in 1971, coinciding with the rise of symbolic AI. Neural nets persisted in cognitive science, and only with the internet’s data explosion and increased compute power did they become powerful again. Around 2012 they began to outperform statistical machine‑learning methods and have since become the prevailing paradigm.\ \ The lesson Lindquist draws is cautionary: suppressing hallucinations can improve reliability, but it may also diminish the system’s capacity for creative insight. The link between surprising, creative conclusions and errors suggests an inherent trade‑off; pushing a system toward one extreme inevitably sacrifices something else. Future AI development must balance robustness with the potential for novelty.\ \ Review: Randomness is the seed of AI creativity, and eliminating hallucinations entirely could blunt its breakthrough potential. Striking a balance between reliability and innovation will be key to the next generation of intelligent systems.