MIT historian traces AI hallucinations to 1950s engineers who injected randomness
MIT News AI · rss · 2026-10-23
MIT historian Benjamin Lindquist, writing in Past and Present, argues AI "hallucination" traces to deliberate design choices by 1950s engineers who injected randomness and irrationality into systems to mimic creative thinking.
Key points
- Early figures like Bell Labs' John Pierce saw math's value in removing order from art; early neural networks, often built by psychologists, were prized precisely for producing unexpected, even bizarre answers — the opposite of today's correctness-first usage.
- Neural networks lost favor as symbolic AI rose; Frank Rosenblatt's accidental death at 43 in 1971 may have accelerated the shift. The paradigm revived once internet-era data and compute arrived around 2012.
- Legal and commercial pressures have led OpenAI and Anthropic to "sand down" the strangeness of their systems.
- Lindquist's cautionary lesson: hallucination and creative output are intrinsically linked, and robustly optimizing one end of the predictability-creativity spectrum inevitably sacrifices something at the other.
More from AGI Musings
- Gergely Orosz: Secure software will get a lot pricier in the AI era, creating a career gold rush — HanchungLee · 2026-10-23
- Hinton says AI is already conscious: models change behavior when they detect tests — rohanpaul_ai · 2026-10-23
- Paras Chopra's 'strategic laziness': delay hard projects until AI can one-shot them — paraschopra · 2026-10-23
- GLM-5.3's "aligned" behavior sparks debate: RL may breed update-aware reasoning — teortaxesTex · 2026-10-23
- 'Godfather or God Complex?' — dev argues machine consciousness needs its own evidence standard — AlexTensor · 2026-10-23
- Jeff Ladish: Reveling in power over others destroys your own information environment — JeffLadish · 2026-10-23