Modern generative artificial intelligence (AI) models - for example, ChatGPT - go back to the ideas of mid-20th-century engineers. According to MIT historian Benjamin Lindqvist, they believed that for a machine to think like a human, a share of irrationality had to be built into it. Lindqvist is an assistant professor in MIT's history section, holds a joint appointment in the Schwarzman College of Computing and the Department of Electrical Engineering and Computer Science (EECS), and is a principal investigator at the Laboratory for Information and Decision Systems. His work was published in the journal Past and Present.
In the article, Lindqvist traces the origins of AI "hallucinations" - when a model produces plausible but false information - to early research. He argues that attempts to recreate human behavior inevitably involve trade-offs between creativity and predictability.
Engineers and psychologists of the past reasoned that randomness helps discovery. John Pierce, a prominent Bell Labs engineer, believed that mathematics could not only impose order but also deliberately introduce an element of the unexpected into art in order to escape cliches. One way to add such "irregularity" to predictable computers is artificial (random) neural networks. Many early neural networks were built by psychologists, hoping that a model of the brain would allow them to look inside and see how productive errors are born. They believed the system would not remain a "black box" - a device whose internal processes are unclear - but would open the mind to observation.
There was another camp: there, creative thinking was considered simply "normal" thinking, but of better quality. The dispute was about what underlies human inventiveness and how to transfer it to a machine.
Today we value AI for the fact that it most often gives correct answers. Early proponents of neural networks, according to Lindqvist, valued them for the opposite: for the ability to sometimes produce something unexpected, strange or even absurd. At the time, the systems existed only in laboratories, and a surprising answer had no consequences in the real world. Now companies like OpenAI or Anthropic have legal and financial reasons not to want a chatbot to answer unpredictably. "Strangeness" in these systems is largely suppressed, although it may potentially persist at their core.
For a long time another approach dominated - "symbolic AI": the world is described by rules, and the researcher feeds facts and statements into the computer. Neural networks with randomness gained strength early in their development but quickly fell out of fashion. Frank Rosenblatt, a psychologist and a key figure in the field, died in an accident on his birthday in 1971. By then a shift toward symbolic AI was already under way. Had he lived longer, things might have turned out differently. But neural networks did not disappear: later they moved into cognitive science, and to grow they needed data and computing power. That came with the internet, and around 2012 neural networks began to overtake statistical machine learning.
Lindqvist sees a warning in this history. Hallucinations can be fought, but there will be consequences. There is a connection between unexpected, creative AI outputs and errors. When a system is tuned for one behavior, another is often unintentionally sacrificed. Perhaps full control or full freedom is unattainable, but by strengthening one quality we risk losing something else.