第一次 AI 寒冬

“第一次 AI 寒冬”是后来补上的季节名称。身处其中的人更可能看到的是一份报告、一项没有续上的经费、一个被要求缩小的目标，而不是某个清楚的入冬日。

英国科学研究委员会委托应用数学家詹姆斯·莱特希尔审视人工智能研究。他的报告署于 1972 年 7 月，随后成为 1973 年英国政策争论的核心材料。报告把相关工作分为三类：面向具体用途的“高级自动化”，用计算机研究中枢神经系统，以及试图把两者连成统一人工智能事业的“桥梁活动”（Lighthill 称之为 category A、B 与中间的桥梁）。莱特希尔承认前两端各有进展，主要否定的是中间那座桥。他认为，通用机器人、场景分析和广泛问题求解没有兑现声称的目标；爱丁堡等地的机器人与视觉项目在公开争论中被点名审视。

报告把主要障碍归到规模，而非某种神秘的心智本质。小型、规则明确的演示一旦扩大，候选状态和组合会急剧增长——后来常被概括为“组合爆炸”。棋类程序依赖人工设计的评价函数与有限步搜索；自然语言系统在狭窄的积木世界中有效，扩展到更大的知识范围便迅速失据。局部成功没有自动累积成通用能力。莱特希尔也讨论了“玩具问题”与真实环境之间的落差：能在实验室桌面上完成的任务，换到开放场景后所需知识与搜索量不成比例地上升。

这份判断并非中立的自然定律。莱特希尔明确说，报告只是一个外部观察者经过约两个月阅读与访谈形成的个人看法；英国 AI 研究者也对它提出反驳，指出报告低估了部分工作的科学价值，并把过宽的期望与过窄的评估绑在一起。争论直接进入了资助决策。科学研究委员会随后大幅削减爱丁堡 AI 与机器人项目的支持，实验室人员随之流失。IEEE Spectrum 等后来的回顾文章常以爱丁堡的 Freddy 机器人作为这段英国收缩的象征性落点：一台曾经展示组装与感知能力的系统，其背后的资助环境已经改变。

美国的降温并不与英国同日发生，但有相近的问责逻辑。1966 年美国国家科学院委托的 ALPAC 报告对机器翻译的实用前景持严厉态度，联邦对自动翻译的大规模资助随之收缩。1970 年代，国防高级研究计划局（DARPA）对 AI 相关项目提出更明确的目标、里程碑与可交付物要求；未达预期的计划被收紧或取消。感知机等神经网络方向在公共叙事中也已降温，与符号 AI 的受挫叠在同一时段，却并非同一份报告造成。不同领域、机构和国家的收缩并不同步，后来才被归到同一个“寒冬”之下。

若把这段历史写成乐观者受罚，便会漏掉资金制度真正追问的内容：一个演示在哪种环境里成立？规模扩大十倍后还成立吗？任务边界由谁设定？研究者承诺的是可用系统、科学理解，还是仅仅一条值得探索的路线？早期 AI 常把这些层次叠在同一句未来宣言里——“十年内机器将……”一类表述在 1950–1960 年代并不罕见——评估到来时，它们又被一起结算。

资金下降没有证明机器智能不可能。公共预算只是不再承担承诺与证据之间不断扩大的差额。专家系统与知识工程在收缩期后仍继续积累；符号方法并未从学术地图上消失，只是换了更谨慎的验收尺度。爱丁堡之后，英国 AI 研究并未归零，而是分散、改名、缩范围。美国实验室同样在收紧的 DARPA 合同语言下继续做搜索、表示与机器人的局部工作。此后每一轮 AI 热潮都换了模型、硬件和术语，却仍要重新回答莱特希尔报告里那个不很浪漫的问题：离开精心划定的小世界，系统还能工作到什么程度？

No one could have circled the first day of the first AI winter on a calendar. The seasonal name came later. At the time, contraction arrived in administrative forms: an external review, a program whose goals were narrowed, a grant that did not continue.

The British Science Research Council asked the applied mathematician James Lighthill to survey work in artificial intelligence. His report was dated July 1972 and became central to a public policy dispute in 1973. Lighthill divided the territory into advanced automation directed at specific applications, computer-based study of the central nervous system, and a middle category that attempted to join those ends into a general AI enterprise. He acknowledged progress at the two poles. His severe judgment fell on the bridge: general robotics, scene analysis, and broad problem solving had not delivered what their advocates claimed. Laboratories such as Edinburgh’s AI and robotics work entered the public argument as concrete cases under review.

The recurring obstacle in the report was not an ineffable property of mind. It was scale. A system performed inside a small, carefully specified environment; enlarge the environment and the combinations multiplied—what later commentary often called combinatorial explosion. Chess programs depended on bounded search and hand-built evaluation. Language programs succeeded inside restricted blocks worlds and then met an explosion of possible knowledge when the walls were removed. Local demonstrations had not accumulated into general competence. Lighthill also stressed the gap between “toy” problems and open environments: success on a laboratory table did not scale proportionally when the required knowledge and search space grew without bound.

Lighthill’s assessment was not an impersonal law. He described it as one outside observer’s view formed after roughly two months of reading and interviews, and British AI researchers contested it, arguing that the report undervalued scientific contributions and tied overly broad hopes to overly narrow measures of delivery. The dispute entered a funding process. The Science Research Council then gutted support for Edinburgh’s AI and robotics program, and the laboratory lost much of its staff. Later retrospectives, including reporting in IEEE Spectrum, have often used Edinburgh’s Freddy robot as a symbolic endpoint for the British contraction: a machine that had demonstrated assembly and perception under conditions whose funding climate had already changed.

The United States did not cool on the same timetable, but similar accountability pressures were visible. The 1966 ALPAC report, prepared for the National Academy of Sciences, took a severe view of practical machine translation and contributed to a sharp reduction in large federal MT programs. In the 1970s, DARPA pressed AI-related projects toward clearer goals, milestones, and deliverables; plans that missed expectations were tightened or cancelled. Neural-network research, already damaged in public narrative after earlier overclaims, cooled in the same broad period without being the product of Lighthill’s document. Programs and countries did not freeze together. “AI winter” gathered these uneven contractions into a retrospective climate.

It is convenient to treat the episode as punishment for optimism. The harder questions were institutional. In what environment did a demonstration succeed? What happened when the state space grew by an order of magnitude? Who set the task boundary? Was a grant buying an exploratory scientific program, a deployable system, or a promise that the two would shortly become the same thing? Early AI rhetoric often carried all three in one sentence—predictions that machines would match human performance “within a decade” were not rare in the 1950s and 1960s. Reviewers eventually settled the account together.

The reduction in public spending did not prove that machine intelligence was impossible. A budget cannot establish that theorem. It established that governments no longer wished to finance the widening difference between evidence and promise. Expert systems and knowledge engineering continued after the contraction; symbolic methods did not vanish from the map, but they met stricter tests of delivery. After Edinburgh, British AI research did not fall to zero; it dispersed, renamed itself, and narrowed its claimed scope. American laboratories likewise kept doing partial work on search, representation, and robotics under tighter DARPA contract language. Later booms would arrive with different hardware, models, and vocabulary. Each would still confront the question in Lighthill’s report that has no dramatic answer: once the little world prepared for the demonstration is gone, how much of the system continues to work?