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

Today's AI hallucinations have roots in a deliberate choice by mid-20th-century engineers, according to Benjamin Lindquist, an MIT assistant professor of history with a shared position in electrical engineering and computer science. In a paper in the journal Past and Present, titled "The Irrational Computer," and an October 23 MIT News interview, he argues that these pioneers added randomness to otherwise deterministic algorithms because they considered some irrationality necessary for machines to think creatively.

Lindquist cites John Pierce, a Bell Labs electrical engineer who argued in the early 1950s that math could strip order from art and introduce useful surprise. Many early neural networks were built by psychologists hoping to observe the brain's productive mistakes, while a rival camp saw creativity as ordinary thinking done better. Those early advocates, he says, prized neural networks for strange answers, the opposite of what users want from ChatGPT or Claude today. He attributes the shift partly to legal and financial pressures on companies like OpenAI and Anthropic, describing a "sanding down" of strangeness that remains largely suppressed.

Lindquist also recounts how neural networks lost ground to rule-based symbolic AI, a shift already underway when their charismatic champion Frank Rosenblatt died in 1971. They persisted in cognitive science by the 1980s, then surged as internet-era data and computing power arrived, pulling ahead of statistical machine learning around 2012.

His warning for developers is that surprising insights and errors are connected, so efforts to eliminate hallucinations are likely to carry unintended costs, including to creativity.