RESEARCH · RESEARCH · #2487
MIT historian argues AI hallucinations stem from mid-20th-century randomization in early neural nets
Benjamin Lindquist (MIT) argues in a paper in Past and Present that early AI researchers intentionally injected randomness into algorithms and neural networks to mimic human creativity; he traces today’s generative-AI hallucinations to those mid-20th-century design choices and warns of inherent tradeoffs between creativity and predictability in systems that model human thought. In an interview with MIT News he contrasts early enthusiasm for occasional strangeness with modern efforts (e.g., by OpenAI and Anthropic) to suppress such behavior for safety and commercial reasons.
KEY POINTS
- Benjamin Lindquist (MIT) argues in a paper in Past and Present that early AI researchers intentionally injected randomness into algorithms and neural networks to mimic human creativity; he traces today’s generative-AI hallucinations to those mid-20th-century design choices and warns of inherent tradeoffs between creativity and predictability in systems that model human thought.
- In an interview with MIT News he contrasts early enthusiasm for occasional strangeness with modern efforts (e.g., by OpenAI and Anthropic) to suppress such behavior for safety and commercial reasons.
- Understanding these historical design choices clarifies why hallucinations are a structural feature of generative models and highlights a core tradeoff that affects current model development and governance.
WHY IT MATTERS
Understanding these historical design choices clarifies why hallucinations are a structural feature of generative models and highlights a core tradeoff that affects current model development and governance.