Questions and Answers: The path by which AI learned to model the messy human mind

Meta Description: MIT historian Benjamin Lindquist says early AI engineers added randomness to mimic creativity, tying today's AI hallucinations to old tradeoffs.

Tags: AI hallucinations, neural networks, generative AI, history of computing, randomness

Featured image: A friendly abstract illustration showing branching neural pathways, small sparks of randomness, and a clean line that gradually straightens, in soft blue and orange.

For readers trying to understand why generative AI sometimes produces strange answers, AI hallucinations may have deeper roots than recent model flaws. In a new paper in Past and Present, Benjamin Lindquist, an assistant professor in the MIT History Section, writes that generative AI models are rooted in mid-20th-century engineering ideas that treated some amount of irrationality as essential for a machine that could think like a human. Lindquist, who holds a shared position between MIT Schwarzman College of Computing and the Department of Electrical Engineering and Computer Science and is a principal investigator in the Laboratory for Information and Decision Systems, describes how early AI pioneers intentionally injected randomness into algorithms designed to solve problems deterministically. MIT News spoke with Lindquist about these ideas. His account traces AI hallucinations back to that research and frames them as an inevitable tradeoff between creativity and predictability.

Why early engineers wanted surprise

According to Lindquist, scientists and artists have long treated chance as central to discovery and to generating original ideas. For them, creativity involved surprising logical jumps, or a dream-like state that lets people see things from new angles.

In the early 1950s, John Pierce, a prominent electrical engineer at Bell Labs, promoted a popular idea that math could be valuable in art. Many people working in early computing shared that view. Although mathematical art is often associated with neat, precise patterns, Pierce reached the opposite conclusion. For him, math was useful because it could take order away from art. Used deliberately, math could inject a sudden irregular element, break familiar patterns, and create conditions where insight might appear.

One popular method for adding irregularity to predictable computers was artificial neural networks, including random versions. Psychologists built many of the first neural networks. They hoped a brain model would let researchers see internal mental activity, including useful errors. Rather than treating the mind as a black box, researchers hoped these systems could open that box and reveal the messy processes behind thought.

A second camp argued that creative thinking was simply normal thinking done better. Lindquist describes these as two simple heuristics for understanding intelligence and creativity, and says they do and do not mesh. The debate centered on what human inventiveness boils down to and how to reconstitute it in machine form.

From lab experiments to symbolic AI and back

Neural networks that used randomness to solve problems began gaining attention in the 1950s, but they fell out of favor early. Frank Rosenblatt, a psychologist and key early figure, died prematurely in an accident on his 43rd birthday in 1971. Lindquist calls him charismatic and forceful, and says he made many grand promises about what neural network systems could do. By the time he died, symbolic AI was already on the rise, and Lindquist suggests events might have gone differently had he lived.

Symbolic AI, a specific kind of AI, was the dominant approach for a long time. It treats the world as a set of rules. The researcher's job is to rebuild knowledge of the world using facts and propositions, then place that structure into a computer.

But neural networks never disappeared. By the 1980s, they had shifted more into cognitive science. To become more powerful and useful, they needed large amounts of data and computational power. Lindquist connects their later rise to the advent of the internet, when data became widely available and computing capacity expanded. Around 2012, neural networks began to outperform and outpace a statistical form of machine learning, and they kept growing.

Why today's AI systems suppress strangeness

Today, people often think of AI systems such as ChatGPT or Claude, and they value them because they usually give correct answers. Lindquist found a reversal in the early history: many initial supporters liked systems precisely because they sometimes gave answers that deviated from expectations. They already had machines that could provide reasonable, expected responses. They saw neural networks as valuable when those networks produced unexpected, even strange or bizarre, answers.

The difference may reflect how these systems operate. Early neural networks were confined to laboratories, so strange or wrong outputs had few real-world consequences. Now, Lindquist says, there are legal and monetary reasons why OpenAI or Anthropic does not want a chatbot that gives unexpected or creative answers. In his telling, modern systems have undergone a gradual smoothing of strangeness. That capacity may still exist at their core, but it is now largely suppressed.

What the history warns developers

The lesson Lindquist draws is cautionary. The ideas and experiences of early engineers suggest developers can work toward quashing hallucinations, shaping AI so it avoids unexpected or unwanted answers. But, he argues, there will be other consequences. Those historical figures were onto something: there is a connection between the surprising and creative conclusions AI can offer and the errors or hallucinations it produces.

For readers, that means the question is not only how to make AI less wrong, but also what may be lost when it becomes less surprising. Building systems to behave one way or another involves tradeoffs. Lindquist suggests it may be impossible to be completely at either end of the spectrum; when work is pushed robustly toward one goal, something else can be sacrificed unintentionally.