图灵提出“模仿游戏”

1950 年秋天，英国哲学季刊《心智》（Mind）第 LIX 卷第 236 期刊出一篇硬邦邦的标题：《计算机器与智能》。作者署名 A. M. Turing——战时破译密码的那位数学家，此时在曼彻斯特与电子计算机打交道。

他没有先给读者一份定义表，也没有急着宣布机器已经会思考。文章几乎是从一句拒绝开始的。

外面的人爱问：机器能不能思考？图灵说，先别这么问。这句话听起来像个完整问题，其实里头塞了两个还没说清的词。你若先去定义“机器”和“思考”，争论很快会滑进日常语言、人的特权，以及什么才配称为心智；你若拿街上的民意调查来定这两个词的意思，结果只会更荒唐，也帮不上研究。图灵不要一场关于名词的永久内战。他要的是：换一个还能动手讨论的问法。

他提议的是一场游戏。

游戏起初更奇怪，也比后来教科书上的简图更完整。一名询问者只靠打字问答，去辨认一男一女——其中一方还故意误导。随后，机器接替其中一人的位置。声音、外貌、触摸全部拿掉，只留下文字。机器不必长得像人，也不因没有皮肤、腿脚或嗓音而失分。要比的不是材料，而是在同一条通信线上给出的回答：回避、错误、风格、知识，以及含糊其辞，都算在内。

这一刀切得很狠。它不证明机器拥有意识，不检查回答背后有没有和人一样的经验，也没有给“智能”颁发本体论证书。它只把一个看不见的名词，改写成一组可以组织、重复和争论的行为。后来人们常说的“图灵测试”，力量和局限都来自这次删减。公众有时把它记成“聊天及格线”；论文本身从未把它写成衡量意识的金标准。

那一年的文章也远不止一场问答游戏。图灵一项一项地处理神学、数学、意识、“无头脑行为”、创造性等反对意见；他把事实和猜测分开写，还留下一个后来常被截取的估计：大约五十年后，经过约五分钟询问，普通询问者正确辨认的机会不会超过七成。这个数字未必是全文最重要的预见，也从未在后来的历史里精确兑现为同一协议下的官方考试。更耐读的部分在后头：与其逐条写完一个成人心智的程序，不如先造一台“儿童机器”，再通过教育、奖励、惩罚、语言和试验让它形成行为。教师甚至可能并不完全知道学生内部发生了什么。训练，而不是手写全部规则——这条路，被当作可以认真走的工程方向提出来。

面对“机器只能做我们命令它做的事”这种说法，图灵没有否认程序受规则约束；他追问的是：人类是否总能预见复杂程序跑起来之后的每一个结果？面对数学上的限制（可计算性与不可判定问题），他承认离散状态机有答不出的题，却指出人同样会犯错。这里没有一张通往胜利的证明表，只有对称地检查：同一把尺子，有没有只拿来量机器、不拿来量人。

于是论文完成了两次替换。先把“思考是什么”换成“在交互中能否分辨”，再把“如何写出智能”换成“如何让机器学习”。前一个替换给了公众一场著名的比赛，也影响此后聊天机器人、人机评测和媒体报道的语汇；后一个替换则悄悄改变了工程问题本身。ELIZA 一类程序后来证明：流畅的文字往返可以触发投射，而投射不等于通过了图灵的游戏，更不等于拥有心智。

模仿游戏没有替人类决定机器是否有心。它把一项更朴素、也更难摆脱的义务留给后来者：每当你确信自己认出了智能，先说清楚——你究竟观察到了什么，在何种信道、多长询问、何种评判协议之下。缺少协议的“通过图灵测试”只是修辞；有了协议，它才重新变成可以反对、可以重做的实验安排。

In the autumn of 1950, the British philosophy journal Mind published a paper with a hard title: Computing Machinery and Intelligence. The author was A. M. Turing—the mathematician who had broken wartime codes and was then working with electronic computers in Manchester.

He did not open with a table of definitions, and he did not hurry to announce that machines could think. The essay almost begins with a refusal.

People liked to ask whether machines can think. Turing’s reply was: not so fast. The question sounds complete, but it smuggles two words that have not been made clear. Define machine and thinking first, and the dispute slides into ordinary language, human privilege, and what deserves to be called a mind. Take a street poll on the meaning of the words, and the result is nearly absurd—and no help to research. Turing did not want a permanent civil war over nouns. He wanted a question that could still be discussed by doing something.

He proposed a game.

The original game is stranger, and more complete, than the classroom diagram it later became. An interrogator, using only typed questions and answers, tries to tell a man from a woman while one of them deliberately misleads. Then a machine takes one participant’s place. Voice, appearance, and touch are removed; only text remains. The machine need not look human, and it loses no points for lacking skin, legs, or a voice. What is compared is not material, but conduct on a shared channel: answers, evasions, mistakes, style, knowledge, and vagueness all count.

That cut is severe. The game does not prove that a machine has consciousness. It does not inspect whether answers rest on experiences like ours. It issues no ontological certificate for “intelligence.” It takes an inaccessible noun and replaces it with an organized encounter that can be observed, repeated, disputed, and redesigned. The later fame of the “Turing test” draws both its force and its limit from that subtraction. Public memory sometimes reduces it to a chatbot pass mark; the paper never installed it as a gold standard of mind.

The 1950 essay also ranges far beyond the game. Turing takes up objections from theology, mathematics, consciousness, “heads in the sand,” and originality. He keeps fact and conjecture apart, and he leaves an estimate often clipped out of context: that in about fifty years, after roughly five minutes of questioning, an average interrogator would have no better than a seventy percent chance of correct identification. That number is not necessarily the paper’s most durable foresight, and history never redeemed it as an official examination under a fixed protocol. The more lasting proposal comes later: instead of writing out an adult mind rule by rule, build a “child machine” and educate it—through teaching, reward, punishment, language, and experiment. The teacher might not fully understand what changes inside the pupil. Training, rather than hand-writing every rule, is offered as a route engineers could seriously take.

To the claim that a machine can do only what it is ordered to do, Turing does not deny that programs follow rules. He asks whether designers can always foresee every consequence of a sufficiently elaborate program. Mathematical limits—computability, undecidable questions—are conceded, then set beside human fallibility. There is no victory table of proofs, only a symmetrical check: whether the same yardstick is applied to machines alone and left off humans.

Turing thus made two substitutions. He exchanged What is thought? for What difference can an interrogator detect? Then he exchanged How do we write intelligence? for How might a machine learn its behavior? The first produced a public ritual that has never stopped attracting contestants, and it shaped the vocabulary of chatbots, human evaluation, and the press. The second quietly altered the engineering problem. Programs such as ELIZA later showed that fluent text exchange can invite projection; projection is not the same as winning Turing’s game, still less the same as possessing a mind.

The imitation game did not decide whether a machine could possess a mind. Its more lasting demand is plainer and harder to escape: whenever someone claims to have recognized intelligence, they must say what, exactly, they were able to observe—under which channel, after how long an interrogation, and according to what judging protocol. Without a protocol, “passing the Turing test” is rhetoric. With one, it becomes an experimental arrangement that can be opposed, redesigned, and run again.