Q&A: How AI came to model the “messy” human mind

Benjamin Lindquist, an assistant professor in MIT's History Section with a shared appointment in the Schwarzman College of Computing and EECS, has published a paper in the journal Past and Present arguing that today's generative AI traces back to mid-20th century engineers who believed a thinking machine needed some dose of irrationality. The paper follows the origins of AI hallucinations to that early research, and Lindquist discussed the findings in an MIT News Q&A. His central claim is that creativity and unpredictability are entangled with error, so efforts to eliminate hallucination will carry costs.
The historical case rests on a counterintuitive idea about mathematics. In the early 1950s, John Pierce, a prominent electrical engineer at Bell Labs, argued that math's value to art lay in its capacity to remove order, not impose it. On that view, a calculated burst of the unexpected could push an artist or thinker away from clichés and toward insight. Many people in early computing shared the notion. Chance and randomness had long been linked to discovery in both science and art, often through the image of a leap of logic or a dream-like state that reframes a problem.
A common mechanism for building irregularity into otherwise deterministic machines was the artificial, or random, neural network. Psychologists built many of these early networks, and Lindquist notes they saw a model of the brain as a way to observe what happens inside it, including how brains make productive mistakes. The appeal was transparency: rather than a black box, the network would let researchers peer into the messy machinery of the mind.
Not everyone agreed. A rival camp held that creative thinking was simply ordinary thinking done better. The debate, as Lindquist frames it, was about what human inventiveness reduces to and how to rebuild it in a machine. That split maps onto a distinction developers still live with: whether a model's departures from expectation are the feature or the bug.
The Q&A draws a sharp contrast between the priorities of the pioneers and those of today's users. People rely on systems like ChatGPT or Claude because their answers are usually correct. Early advocates, by contrast, already had machines that produced reasonable, expected output; they valued neural networks precisely because they occasionally returned something strange or bizarre. Lindquist attributes part of the shift to setting. Early systems stayed in laboratories, where a wrong output had no consequences. Today, he says, legal and monetary pressures give companies like OpenAI and Anthropic strong reasons to avoid unexpected answers, and he describes a sanding down of strangeness: the capacity may remain at the core, but it is largely suppressed.
The paper also traces how neural networks moved from promising to dominant. For decades symbolic AI held sway, treating the world as rules and the researcher's job as encoding knowledge in facts and propositions to feed a computer. Neural networks gained traction in the 1950s but soon fell out of favor. Frank Rosenblatt, a psychologist and pivotal early figure, died accidentally on his 43rd birthday in 1971; he had been a charismatic presence who made large promises for the approach, and a turn toward symbolic AI was already under way when he died. Lindquist suggests the field might have unfolded differently had Rosenblatt lived.
Neural networks survived by migrating into cognitive science in the 1980s. Scaling them required abundant data and compute, which arrived with the internet. Around 2012 they began to outperform a statistical form of machine learning, and they kept growing. This is a familiar arc to anyone who has watched deep learning displace hand-engineered pipelines, and it is worth noting that the ingredients Lindquist identifies, data and compute, are the same ones that later became the standard explanation for the modern boom.
His conclusion is a caution rather than a prediction. It is possible to work toward quashing hallucinations, but doing so will have other consequences. The early engineers, in his reading, were onto something real: the surprising and creative conclusions a model can offer are connected to its errors. Building robustly toward one end of the spectrum tends to sacrifice something at the other, often unintentionally.
For developers, this suggests a practical framing for the hallucination problem. Temperature settings, sampling strategies, RLHF and guardrail layers all shift a model along a predictability spectrum rather than eliminate a discrete fault. A likely trade-off is that aggressive suppression of unexpected output also narrows the space of unusual, potentially useful suggestions, which matters most in brainstorming, drafting and other open-ended product surfaces.
Why it matters: Teams shipping generative features are usually asked to reduce wrong answers, and this historical account argues that the same mechanisms producing those errors also produce the surprising output users find valuable. The practical implication, as inference, is that products may need explicit modes or settings that trade accuracy for variety rather than a single global default. It also reframes evaluation: a benchmark that rewards only correctness may quietly measure how much strangeness a model has had sanded away.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Assistant Professor Benjamin Lindquist studies how engineers deliberately injected randomness into early AI, giving rise to deep-learning models that sometimes hallucinate.