MIT Researcher Traces AI Hallucinations to 1950s Randomness Injection, Warns of Creativity Tradeoffs
Updated
Updated · MIT News · Oct 23
MIT Researcher Traces AI Hallucinations to 1950s Randomness Injection, Warns of Creativity Tradeoffs
1 articles · Updated · MIT News · Oct 23
Summary
Benjamin Lindquist argues in a new Past and Present paper that AI hallucinations stem partly from early engineers’ deliberate use of randomness to make machines think more like humans.
In the 1950s, researchers such as Bell Labs engineer John Pierce and early neural-network advocates treated unexpected outputs as a feature, believing creative insight required irregularity, productive mistakes and leaps beyond strict logic.
That view contrasts with today’s chatbot era, when companies suppress strangeness because wrong answers carry legal and financial risks even if the underlying systems still retain that capacity.
Lindquist says the history points to a persistent tradeoff: reducing hallucinations can also curb the surprising, creative behavior that randomness-based AI was originally built to produce.
If AI hallucinations are a deliberate creative feature from the 1950s, can developers ever eliminate them without destroying machine intelligence?
With AI-generated fake citations infiltrating government laws, are we sacrificing global institutional trust just to make algorithms mimic human irrationality?