Turing’s Wager

1. The imitation game↗

In 1950, Alan Turing opened Computing Machinery and Intelligence with a question: can machines think?

The question is short; its definitions are not. What counts as thinking, or as a machine? Rather than settle those boundaries first, Turing changed the question: consider a game.

The imitation game.

The paper begins with an interrogator identifying the sex of human participants, then asks what happens when a machine replaces one of them. The familiar human–machine version retains a crucial arrangement: participants are separated, and the interrogator uses text, without hearing voices or seeing appearances.

This directs debate towards observable performance. It does not establish a theorem that indistinguishability entails understanding. Turing also considers objections to replacing the original question this way.

More than seventy years later, chat models have brought the game into everyday life. A 2025 preprint by Jones and Bergen tests five-minute text conversations with a human and a system simultaneously. Given a humanlike persona prompt, GPT-4.5 was selected as the human 73% of the time.

That result deserves attention. Model, prompt, duration and interrogators are also part of it. It concerns performance on a defined discrimination task, while the question of understanding remains on the table.

2. The parrot’s curse↗

In their March 2023 New York Times commentary, Noam Chomsky, Ian Roberts and Jeffrey Watumull argue that human minds do more than discover statistical associations: they offer explanations and consider what is possible or impossible.

This is a position on what understanding requires, rather than a proof supplied by a training objective alone. The frequently invoked “stochastic parrot” critique has another source, in Bender and colleagues’ 2021 work. The author’s contemporary presentation distinguishes linguistic form from meaning; it should not be folded into Chomsky’s argument as though the two were identical.

The parrot image compresses the doubt into a question: when the sounds and sentences are right, must meaning be present too?

Fluency does not establish understanding.

Yet next-token prediction describes a training objective. What a model learns internally to perform it is another research question. Turning that objective directly into “mere repetition” skips a step, just as turning fluency directly into understanding does.

3. But wait↗

3.1 Is it only repeating what it has heard?↗

The 2023 Skill-Mix study turns towards composition: choose a topic and a random set of skills, then ask the model to satisfy them within a short passage. Metaphor, a form of inference and particular expressive constraints, for example, must work together.

If answers come only from recalling whole passages, growing combinations should make the task harder. A simple formalisation helps: uniformly choose an unordered subset of $k$ skills from $N$. If the seen collection $\mathcal C_{\mathrm{seen}}$ covers $m$ distinct subsets and $S$ is the sampled subset, then:

Here $1\leq k\leq N$ and $0\leq m\leq\binom{N}{k}$. This is a coverage calculation under uniform sampling, rather than Skill-Mix’s full estimate of training-corpus exposure. Topics, occurrence frequencies and dependencies between skills have not been included.

Using frequency estimates and assumptions, the paper argues that some GPT-4 performance supports generalisation beyond seen combinations. It does not directly inspect the entirety of GPT-4’s training set. Evidence of novel combinations is different from proof of understanding or consciousness. Successful composition can weaken a verbatim-recall account without settling how meaning arises.

3.2 The question turns back towards people↗

Context, memory and habit constrain the next word when people speak too. What role, then, does prediction play in human language?

There is no need to turn that question immediately into “humans are parrots too”. Prediction may be a useful description; a shared word does not establish shared mechanisms, objectives or experience between brains and language models.

The similarities deserve testing. So do the differences.

4. A question at the cliff’s edge↗

The apparent alternatives are unsettling: the machine is an elaborate symbolic deception, or human understanding is itself an illusion.

There is still ground between those judgments. A system can have real, limited capacities for inference and composition while lacking other kinds of understanding. Whether it has subjective experience is another question. Acknowledging one capacity need not settle the rest.

Treat an untested capacity as reliable knowledge, and fluent answers can become mistaken grounds for judgment. Dismiss every success as repetition, and there may be computations worth explaining that go unnoticed. In advice, teaching or companionship, the immediate questions remain: when is the system reliable, how does it fail, and who bears the consequences?

Beliefs about human uniqueness should not be defended or overturned by a single model score. AI brings several entangled questions back into view: what is thinking? How should understanding be tested? What evidence do meaning and experience each require?

5. Why language?↗

One further question remains. Progress in AI did not begin with language. Why does a speaking system nevertheless touch questions about people so directly?

Language records facts and organises explanations. It describes the world and helps people negotiate intentions. A system that enters this medium fluently acquires more than another output format: it enters a space through which people judge who is on the other side.

Information theory offers a thread to follow. Shannon’s 1948 A Mathematical Theory of Communication connects statistical structure with coding and transmission. It also explicitly sets semantics apart from the communication-engineering problem it addresses.

Shannon’s result therefore cannot be written as “prediction = compression = understanding”, nor does it establish language as the most compressible form of human knowledge. Regularities in text can indeed be learned and used for prediction and coding. Which reflect the world, and which reflect habits of expression, remain questions.

Prediction and compression admit precise connections. Moving from compression to understanding still requires an explanation.

The next essay begins with that unestablished equality.

1. 模仿游戏↗

1950年，阿兰·图灵在 Computing Machinery and Intelligence 中提出了一个问题：机器能思考吗？

问题很短，定义却很难。什么叫“思考”，什么叫“机器”？图灵没有先替这些词划定边界，而是换了一个问题：来玩一个游戏。

模仿游戏。

原文先从裁判辨别人类参与者的性别谈起，再问机器替代其中一方会怎样。后来熟悉的人机辨别版本，保留了其中关键的安排：参与者隔开，裁判通过文字提问，不能借助声音或外表判断。

这一步把争论引向可观察的表现，却没有给出“无法分辨，所以必有理解”的定理。图灵也讨论了为什么有人会反对这种替换。

七十多年后，聊天模型让这个游戏走进日常。2025年 Jones 与 Bergen 的预印本实验让裁判同时与真人和系统进行五分钟文字交谈；在人类化角色提示下，GPT-4.5 有73%的次数被选为真人。

这个结果值得认真对待。模型、提示、时长和裁判，也都是结果的一部分。它回答的是一个明确辨别任务中的表现，关于理解的问题仍然留在桌上。

2. 鹦鹉的诅咒↗

2023年3月，Noam Chomsky、Ian Roberts 和 Jeffrey Watumull 在《纽约时报》评论中提出尖锐批评：人类心智不只发现统计关联，还会提出解释，考虑什么可能、什么不可能。

这是对“什么才算理解”的一种立场，不是一个仅凭训练目标就已完成的证明。另一路常被援引的“随机鹦鹉”批评，来自 Bender 等人的2021年工作；作者当年的讲解特别区分了语言形式与意义，不能把它直接当成乔姆斯基提出的同一个论证。

鹦鹉这个比喻之所以有力，是因为它把疑问压成一句话：声音与句式都对了，意义就一定在场吗？

流利，不等于懂。

但“预测下一个 token”描述的是一种训练目标。为了把这个目标做好，模型内部学到了什么，还需要另行研究。直接把目标改写成“只会复述”，与直接把流利改写成“已经理解”，都跳过了一步。

3. 但是，等等↗

3.1 它只是重复听过的话吗？↗

2023年的 Skill-Mix尝试把问题推向组合：给定一个主题，再随机挑出几种技能，要求模型在同一段短文中满足它们。例如，比喻、推理形式与特定表达约束，需要在一起成立。

如果答案只来自整段背诵，组合一多，任务就应当变难。这个直觉可以作一个简单的形式化：从 $N$ 种技能中均匀抽取 $k$ 种，不考虑顺序。若已见集合 $\mathcal C_{\mathrm{seen}}$ 覆盖了其中 $m$ 个不同子集，以 $S$ 表示本次抽取，则：

这里假定 $1\leq k\leq N$，且 $0\leq m\leq\binom{N}{k}$。它只是说明均匀抽样下覆盖率的计算，不是 Skill-Mix完整的训练语料估计：主题、技能出现频率和技能之间的依赖还没有放进来。

论文据其频率估计与假设，认为 GPT-4 的部分表现支持超出见过组合的泛化；它没有直接查遍 GPT-4 的训练集。新组合的证据，与理解或意识的证明，不是同一件事。 组合做得好，可以削弱“只会原样背诵”的解释，却还没有回答意义怎样产生。

3.2 问题也会回到人类身上↗

当人说话时，语境、记忆和习惯当然会约束下一个词。那么，预测在人的语言使用中扮演什么角色？

这个问题不必急着变成“人类也是鹦鹉”。预测可以是一种有用的描述，却不能仅凭这个共同词语，就认定大脑与语言模型具有相同机制、目标或体验。

相似之处值得检验。差异也一样。

4. 悬崖边的问题↗

原先看似只有两个方向：机器只是符号的骗局，或者人类的理解也不过是幻觉。两边都令人不安。

但这两个判断之间，仍有可以立足的地方。一个系统可以具有某些真实而有限的推理与组合能力，同时仍不具备另一些意义上的理解。它是否有主观体验，又是另一个问题。承认前者，不必提前宣布后者。

若把尚未检验的能力当作可靠知识，就可能把流畅的回答误作判断依据。若把所有成功都叫作复述，又可能错过需要解释的计算。无论在顾问、教学还是陪伴场景，首先要问的仍是：它在什么情况下可靠，怎样失败，谁来承担失败的后果？

关于人类独特性的信念，也不该只靠机器的一次得分来保卫或推翻。AI 把几个原本纠缠的问题重新摆到面前：什么是思考？怎样检验理解？意义与体验，又各自需要什么证据？

5. 为什么是语言？↗

这一篇还留下另一个悬念。AI 的进展并不始于语言，但一个会说话的系统，为什么能如此直接地碰到关于人的问题？

语言既记录事实，也组织解释；既描述世界，也让人协商彼此的意图。一个能熟练进入这种媒介的系统，获得的不只是又一种输出格式。它也进入了人用来判断“对面是谁”的空间。

信息论提供了一条值得追的线索。1948年，香农的 A Mathematical Theory of Communication把消息的统计结构与编码、传输联系起来。但他也明确把语义问题与所处理的通信工程问题区分开来。

因此，不能把香农的结论写成“预测＝压缩＝理解”，更不能未经比较，就称语言为人类知识最可压缩的形式。文本的规律确实能被学习、用于预测与编码；其中哪些规律反映了世界，哪些只是表达的习惯，仍需要继续追问。

预测与压缩之间可以建立精确联系；从压缩走到理解，还欠一个解释。

下一篇，就从这个尚未成立的等号开始。