ELIZA 展示对话程序的投射效应

人： 男人都一样。

程序： 哪些方面？

约瑟夫·魏岑鲍姆把这段对话印在 1966 年的 ELIZA 论文里。接下来，人继续解释自己的烦恼，程序继续索要例子、复述句子、把话题送回去。读者很容易在“哪些方面”里听见关心；ELIZA 内部却没有关于男人、关系或烦恼的模型。

它做的是一套可以逐项写在纸上的变换。程序从输入中寻找带优先级的关键词，再按“分解规则”把句子切成片段；“重组规则”交换人称、搬运原词，并把它们嵌进预先准备的回复。找不到合适关键词时，它可以说“请继续”，也可以把先前保存的一段输入稍作改写后重新拿出来。所谓记忆并不是对往事的理解，更像桌面上暂存的一张纸条。

ELIZA 与扮演的角色并非一回事。程序可以加载不同脚本，最著名的 DOCTOR 只是其中之一。它模仿罗杰斯式访谈中倾听、反问和鼓励展开的表面形式，这个角色尤其适合规则系统：谈话者负责提供人物、经历和因果，程序只需把材料以开放问题送回。每一次“能具体说说吗”都让对话继续，也把尚未完成的语义工作交还给用户。

魏岑鲍姆并没有把机制藏在舞台后。他在论文中列出规则与流程，还用一个英语有限、靠套话表示兴趣的外国人作类比。他观察到，人会先为谈话对象假定知识和推理能力；只要下一句尚能被解释成合乎情理，这个想象中的对象就不会崩塌。ELIZA 最有效的一部分“计算”，其实发生在终端另一端：人主动寻找上下文、补足动机，并原谅含糊的回答。

这也是程序最有力量、也最危险的地方。MIT 分时系统上的文字往返证明，普通语言可以成为计算机界面；用户不必学习一套命令，便能自然地继续输入。但界面的亲近感会诱使人把持续对话误认为理解，把合适的句式误认为判断。魏岑鲍姆后来在《计算机的力量与人的理性》中进一步追问的，正是人是否会把本应由自己承担的道德责任交给机器。

今天的语言模型与 ELIZA 的关键词表在技术上已经相隔数代，ELIZA 却仍提供一种有效的阅读动作：当一句机器回复显得体贴或聪明时，问一问其中有多少来自系统，又有多少是自己刚刚补进去的。1966 年那句“哪些方面？”没有理解上一句话；它只是足够准确地留下一个空位，让人把理解放了进去。

MEN ARE ALL ALIKE.

IN WHAT WAY?

Joseph Weizenbaum printed this exchange in his 1966 paper on ELIZA. The person continues with a complaint; the program requests examples, reformulates phrases, and returns the subject to the speaker. It is easy to hear concern in “In what way?” Inside ELIZA, however, there was no model of men, relationships, annoyance, or the person at the terminal.

Its procedure could be laid out line by line. ELIZA searched the input for prioritized keywords. A keyword selected a decomposition rule that divided the sentence into pieces; a reassembly rule changed pronouns, moved the user’s words, and fitted them into a prepared reply. If no useful keyword appeared, the program could ask the user to continue or retrieve a transformed fragment saved earlier. Its “memory” was not an understanding of the past so much as a note left on a desk for later reuse.

ELIZA was also distinct from the role it played. Different scripts could supply different keyword lists and transformations. The famous DOCTOR script imitated the surface form of Rogerian, person-centered interviewing: reflection, requests to elaborate, and questions that returned attention to the speaker. That role was exceptionally economical for a rule system. The user supplied the people, history, and causal links; the program kept the exchange moving while contributing almost no external facts. “Can you give me an example?” sounded responsive and quietly handed the unfinished semantic work back to the human.

Weizenbaum did not conceal this arrangement. His paper exposed the procedures and rule tables, and compared the mechanism to a foreign speaker with limited English who recognizes a phrase and uses a formula to display interest. People, he observed, approach a conversational partner with assumptions about what it knows and how it reasons. As long as the next reply can be reconciled with those assumptions, the imagined partner survives. Some of ELIZA’s most effective computation therefore took place outside the computer: the user supplied context, inferred motives, and repaired ambiguity.

That was both the demonstration’s power and its warning. On MIT’s time-sharing system, ordinary text became an approachable computer interface; a person could continue without learning a command language. But the ease of the interface encouraged a category mistake—conversation mistaken for comprehension, and an apt phrase mistaken for judgment. In Computer Power and Human Reason, Weizenbaum later pushed the moral issue further: a machine’s ability to participate in an exchange did not justify transferring human responsibility to it.

Modern language models are technically remote from ELIZA’s keyword tables, but the old program still teaches a useful reading habit. When a machine’s sentence feels caring or intelligent, ask how much is represented by the system and how much has just been supplied by the reader. The 1966 reply “In what way?” did not understand the preceding complaint. It placed an empty space with such precision that the person on the other side filled it with understanding.