Efforts to make AI more reliable may also narrow the range of answers it can produce. That tension has deep roots: early computer researchers explored randomness because they thought unexpected results could help machines imitate human creativity.
Benjamin Lindquist, an assistant professor in the MIT History Section, traces this history in a paper in Past and Present. His account links today’s debate over AI hallucinations to an older question: can a machine be creative without sometimes producing results that are strange or wrong?
Why early researchers welcomed the unexpected
Some scientists and artists saw chance as part of discovery. They connected creativity with leaps of logic and shifts in perspective, rather than only with orderly reasoning. Early computing researchers carried similar ideas into their thinking about machines.
In the early 1950s, electrical engineer John Pierce of Bell Labs argued that mathematics could introduce irregularity into art. Rather than making artistic work more orderly, he believed mathematical methods could help disrupt clichés and create unexpected combinations that might lead to insight.
Artificial neural networks offered one way to introduce that irregularity into computers. Many early networks were built by psychologists interested in modeling the brain. They hoped such models could help researchers see how minds work, including how mistakes might sometimes contribute to productive thought.
This approach differed from another view of creativity: that it was ordinary thinking performed better. The disagreement went to the heart of early AI research. What makes people inventive, and how might a machine reproduce that ability?
From laboratory experiments to everyday tools
Early supporters valued neural networks partly because they could return answers that departed from expectations. Computers already offered ways to produce reasonable, predictable results; a system that occasionally gave a strange response seemed to offer a different kind of value.
Today, people often judge AI systems by whether their answers are correct. Lindquist points to a change in context as one reason for that shift. Early systems were used in laboratories, where an odd or incorrect answer had fewer real-world consequences. Chatbots are now used in settings where unexpected answers can carry legal and monetary implications.
As a result, contemporary systems have been shaped to suppress some of their strangeness. That can make them more useful for tasks where people need dependable answers, while also limiting the unpredictable responses that early researchers associated with creative potential.
How neural networks returned to prominence
For a long period, symbolic AI was the leading approach. It treated the world as a system of rules and aimed to encode knowledge as facts and propositions that a computer could use.
Neural networks began gaining attention in the 1950s, but they later lost ground. Frank Rosenblatt, a psychologist and influential early figure in the field, died accidentally on his 43rd birthday in 1971, as symbolic AI was already gaining favor. Lindquist suggests that Rosenblatt’s continued involvement might have changed the course of events, though that possibility cannot be known.
Neural networks persisted and shifted toward cognitive science by the 1980s. Their development depended on access to large amounts of data and computing power. The internet made more data available, while computing capacity grew. Around 2012, neural networks began to outperform and outpace a statistical form of machine learning, and their influence continued to expand.
The tradeoff behind reducing hallucinations
Lindquist’s historical account offers a caution for current AI development. Reducing unwanted or unexpected answers may improve reliability, but it can also affect the system’s capacity to produce surprising conclusions. The history suggests these qualities may be connected.
That does not mean strange or mistaken answers are inherently valuable. Rather, it highlights a design tension: pushing a system strongly toward one behavior can have effects elsewhere. If developers prioritize predictability, they may also suppress some of the irregular output that early researchers hoped could support creativity.
The history therefore reframes hallucinations as more than a technical flaw to eliminate. It asks what kinds of output society wants from AI, and what qualities might be lost when systems are tuned to avoid uncertainty. The challenge is to understand the tradeoffs while building tools suited to the consequences of their use.