MIT历史学者林奎斯特追溯AI“幻觉”起源：早期工程师为何刻意注入随机性

一场关于“犯错”的旧争论

MIT历史系助理教授本杰明·林奎斯特在《Past and Present》发表的新论文中提出，今天被视为AI缺陷的“幻觉”，其源头可追溯到20世纪中期一批工程师的主动设计。他们相信，要让机器像人一样思考，一定程度的非理性不可或缺。为模仿创造性思维，这些先驱在原本用于确定性求解的算法中刻意加入随机性。

随机性为何被当作创造力

林奎斯特在访谈中举出贝尔实验室工程师约翰·皮尔斯的观点：数学的价值不在于让艺术更有序，而恰恰在于能“去除秩序”，以计算的方式加入意外，从而偏离陈词滥调、可能带来洞见。早期神经网络多由心理学家构建，他们希望这类模型能打开大脑的黑箱，让人观察“有产出的错误”如何发生。另一派则认为创造性思维只是“做得更好的正常思维”，两派之争实质是在追问人的创造力究竟是什么、又如何以机器形式重建。

从实验室到产品：怪异被“打磨”

林奎斯特指出，早期系统局限于实验室，输出怪异或错误没有现实后果；如今OpenAI或Anthropic等公司面对法律与金钱上的理由，不希望聊天机器人给出意外或创造性的答案，系统中的“怪异”已被大幅压制，但可能仍潜藏于核心。他还提到，符号主义AI曾长期占据主流，神经网络在1950年代获得关注后一度失势，心理学家弗兰克·罗森布拉特1971年意外去世时，转向符号主义的趋势已在发生；神经网络后来转入认知科学，直到互联网带来数据与算力，约2012年起超越一种统计形式的机器学习。

Generative artificial intelligence models are rooted in ideas of mid-20th century engineers who thought some amount of irrationality was essential for a machine that could think like a human, contends Benjamin Lindquist, an assistant professor in the MIT History Section, who holds an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science (EECS) and is a principal investigator in the Laboratory for Information and Decision Systems. 

To mimic creative thinking, these AI pioneers intentionally injected randomness into algorithms designed to deterministically solve problems, Lindquist writes in a new paper that appears in the journal Past and Present. In the paper, he traces the origins of AI hallucinations back to this early research and suggests there will be inevitable tradeoffs between creativity and predictability when trying to build machines that mimic human behavior. MIT News spoke with Lindquist about these ideas.

Q: Why did early computer engineers believe they needed to inject randomness and irrationality into AI systems designed for logic and calculation? 

A: There is a long history of many figures in the world of science and art theorizing that chance plays an important part in discovery and creativity — in the ability to originate ideas. They equated creativity with unexpected leaps of logic, or even a dream-like state that allows you to think about things from a new perspective. 

In the early 1950s, one popular idea from John Pierce, who was a prominent electrical engineer at Bell Labs, was about math being valuable in the world of art. The idea was shared by a lot of people working in early computing. When we think of artwork that includes math, we may think of something very orderly and precise, but Pierce came to the opposite conclusion. He concluded that the utility of math is that it could remove order from art. Then in a very calculated way, math could add a burst of the unexpected as a means to deviate from clichés and add an element of irregularity which just might lead to insight. 

One popular way to add irregularity into otherwise predictable computers was through artificial (or random) neural networks. Many early neural networks were built by psychologists, who thought that by enabling scientists to build a model of the brain, artificial neural networks allowed scientists to observe what is happening inside, including how brains make productive mistakes. Rather than being black boxes, researchers believed these systems could open this black box of the mind and let scientists peer inside and see all the messy stuff that was really happening.

But there was another camp in the world of AI who thought that creative thinking was just “normal” thinking done better. It is interesting to look at these two very simplistic heuristics for understanding intelligence and creativity, and how they do and don’t mesh. The debate, then, was about trying to determine what human inventiveness boils down to, and then how to reconstitute human inventiveness in machine form. 

Q: How did these early computer engineers view AI, and how is it different from the way we think about AI today?

A: When we think of AI systems, we think of ChatGPT or Claude. We value and use these systems because they give answers that, most of the time, are correct. One of the things that was most interesting to me in this research was finding out that a lot of the early advocates of these systems valued them for exactly the opposite reason. They wanted a system that, on occasion, gave answers that deviated from their expectations. They already had systems that could deliver reasonable and expected answers. They saw neural networks as valuable because they could give unexpected, perhaps even strange and bizarre answers.

One reason we might view these systems differently today could be because of how they operate. Early on, these systems were confined to laboratories, so there weren’t really any real-world consequences of a system that gives a strange, or wrong, output. But now there are legal and monetary reasons why OpenAI or Anthropic doesn’t want a chatbot that gives unexpected or creative answers. There has been a “sanding down” of strangeness in these systems. They still potentially have that at their core, but it is something that is now largely suppressed.

Q: How did early neural networks that incorporated randomness to simulate human creativity evolve to become the dominant paradigm in AI?

A: For a long time, a specific kind of AI called “symbolic AI” was the dominant approach. Essentially, symbolic AI says the world is based on rules. The role of the AI researcher is to try and reconstruct the knowledge of the world in facts and propositions and feed it into a computer.

Neural networks, with their use of randomness to solve problems, started to gain traction in the 1950s but fell out favor early on. Frank Rosenblatt, a psychologist who was a very important early figure in neural networks, died prematurely; he died accidentally on his 43rd birthday in 1971. He was quite a charismatic and forceful personality, who made a lot of grand promises about what systems based on neural networks could do. By the time he died, a shift toward symbolic AI was already underway. It is possible if, had he lived, things might have turned out differently.

But neural networks never went away. By the 1980s they had shifted more into the world of cognitive science. To become more powerful, and more useful, neural networks needed a lot of data and computational power. That didn’t really arrive until the advent of the internet, when a lot of data became available and computing started to power up. Around 2012, neural networks started to outperform and outpace a statistical form of machine learning, and they kept growing from there.

Q: What lessons can we learn from studying AI’s history that could inform its development? 

A: I think there is a cautionary lesson here. The ideas and experiences of these early engineers tell us that we can work toward quashing hallucinations — we can work toward making AI perform a certain way so it doesn’t produce unexpected or unwanted answers — but there are going to be other consequences of that. These historical figures were onto something. There is a connection between the surprising and creative conclusions AI can offer and errors or hallucinations. There are tradeoffs when we build our systems to behave one way or the other. Maybe it is impossible to be completely at either end of the spectrum, but when we work toward one thing robustly, we often sacrifice something else unintentionally.