为什么 Anthropic 的公开写作风格与 Claude 如此不同？
英文标题：🔶 Why is Anthropic's public writing style so unlike Claude's?
作者认为，Claude 具有极易辨认的写作风格：短促有力的句子、频繁的强调性表达，以及已经能被自动化工具捕捉的固定措辞。尽管 Claude 及受其蒸馏影响的模型常呈现这种“Claude 味”，Anthropic 的研究文章、政策立场和公开文案却采用更自然、易读且近似资深博客作者的语调。文章提出多种猜测，包括创始人偏好亲自写作、维护“人类掌控”的品牌形象，以及刻意让 Claude 输出可被识别等。作者进一步质疑，这种风格是否会降低 Claude 面向高辨识度读者的写作实用性，并好奇 Claude 在非英语语言中是否同样具有鲜明辨识度。
中文
为什么 Anthropic 的公开写作风格与 Claude 如此不同？
__ 2026 年 8 月 3 日
Claude 有一种非常鲜明的写作风格：短促、有冲击力的句子，以及不断刺入眼帘的强调表达。我们甚至有自动高亮工具，它们完全不尝试以花哨的统计方式预测是否使用了 LLM，而只是匹配一些短语模式——这些短语在 Claude 出现之前几乎无人使用，但如今因 Claude 而随处可见。
大多数时候，我甚至不需要高亮工具。读到这样的文章，或是这样的文章时，我会立刻觉得：这显然是 Claude 的声音。
“菜单布局？几乎一样。你不是在重新学习一款浏览器——你只是在使用一个更干净的版本，而它的使用方式你本来就已经熟悉。”
“K3 没有搭载任何 draft tensor——没有 MTP，也没有 EAGLE——因此唯一的推测路径是外部 block-diffusion draft：RadixArk 的 Kimi-K3-DSpark。”
不仅 Claude 本身说话是这种风格，哪怕只是轻度从 Claude 蒸馏而来的 LLM（例如 Kimi K3）也是如此。也许还有一种可能：_人类_在接触了足够多 Claude 输出后，也会在自己的写作中无意间沾染这种风格。😱
LLM 的输出并不_必须_是这种风格！GPT-5 和 Gemma 4 的写法更“正常”。因此，Claude 的声音至少是 Anthropic 半有意为之的选择。这似乎不是 Amanda Askell 的错。
也许这是几百万次 RLHF 微调逐渐涌现的结果：有人一次次选择听起来更好的 A 句，而不是 B 句？Claude 是否通过梯度下降，最终收敛到了人们整体上真正偏好的风格？我不确定自己是否相信这一点。长期以来，思想市场一直奖励那些采用读者认为有吸引力写作风格的作者；然而，数个世纪以来这些激励机制的共同作用，并没有让整体书面记录向 Claude 风格倾斜。
无论如何，下面是我觉得最有意思的一点。Anthropic 的公开写作完全不使用 Claude 的声音，连一点都没有！ 它们措辞清晰，但读起来轻松，几乎是对话式的散文风格，就像一位经验丰富的博主。去读读关于 J-Space 的文章，或者浏览一下 Claude 的宪法。即便 Claude 参与撰写过这些内容，他们也非常成功地掩盖了这一点。
我们知道 Anthropic 广泛地内部使用 Claude。Claude 现在编写 Anthropic 80% 的代码，那为什么不让它写散文？为什么 Anthropic 对研究成果的介绍和政策立场的表达方式，与其产品的表达方式_如此不同_？
以下是一些猜测，按从平淡无奇到天马行空的顺序排列。
- 我们知道 Dario 是一位热衷写文章的人。也许无论 LLM 听起来怎样，他都更愿意以自己的声音面向公众继续写作。我能理解！
- Anthropic 的品牌信任（就如今的情况而言，姑且这么说）建立在“人类仍然掌控一切”之上。让人觉得 LLM 对公司最重要的公开表述哪怕产生了一点影响，都会造成不佳观感。因此，研究和公关文章不应听起来像产品本身，可能永远都不该如此。
- Anthropic 内部存在一个群体，他们对 Claude 声音拥有足够的影响力，并且出于多种与安全、透明度以及识别“对抗性”蒸馏相关的理由，相信 Claude 的写作应该对细心的人类而言独特且可识别。因此，他们在输出中散布一些个人语体标记。这些标记不会出现在_每一句_中，但会出现在足够多的句子里，以至于只需几段 Claude 文本就能令人一眼认出。与此同时，Anthropic 内部负责润色研究和政策文章的人，发现这些相同的个人语体标记既令人疲惫又陈词滥调，因此会主动删去它们。（又或者，他们能够使用一款仅限内部使用、拥有不同声音的模型。）
如果你同意 Claude 的声音会让它在面向足够有辨识力的受众写作时变得不那么有用，那么市场似乎会随着时间推移推动 Anthropic 将这种风格“调掉”。如果 Fable 听起来比此前模型稍微更不像 Claude，也许这个过程已经开始了；但 Opus 5 肯定仍然带着这种风格。
到底是怎么回事？
我的多语言能力不足以回答这个问题，但现在我开始好奇：Claude 在非英语语言中听起来有多像 Claude？
English / 原文
Why is Anthropic's public writing style so unlike Claude's?
__ Aug 3, 2026
Claude has a very distinctive writing style: short punchy sentences and eyeball kicks. We have automatic highlighting tools that don’t even try to predict LLM use in a fancy statistical way, only match on phrase patterns that hardly anyone used prior to Claude, but which now appear everywhere thanks to Claude.
Most of the time, I don’t even need the highlighting tool. I’ll start reading something like this, or this, and think this is obviously Claude’s voice.
“Menu layout? Nearly the same. You’re not relearning a browser — you’re just using a cleaner version of what you already know.”
“K3 ships zero draft tensors — no MTP, no EAGLE — so the only speculative path is an external block-diffusion draft: RadixArk’s Kimi-K3-DSpark.”
In addition to Claude itself sounding like this, so do LLM that are even mildly distilled from Claude (such as Kimi K3). Maybe it’s also people who are exposed to enough Claude output to unintentionally pick up the style in their own writing. 😱
LLM output doesn’t have to sound like this! GPT-5 and Gemma 4 write more ’normally’. So, Claude’s voice is at least a semi-deliberate choice by Anthropic. It doesn’t appear to be Amanda Askell’s fault.
Maybe it has emerged from a few million RLHF nudges where someone picked sentence A as sounding better than B? Has Claude gradient-descended onto the style that people actually prefer in aggregate? I’m not sure I would buy that. The marketplace of ideas has long rewarded authors for using whatever style readers find compelling, and these incentives playing out over centuries did not tilt the total written record in a Claudey direction.
Whatever the case, here is what I find most interesting. Anthropic’s public writing does not use the Claude voice, not even a little bit! It’s articulate but easy-reading, almost conversational prose, like a seasoned blogger. Go read about the J-Space or skim Claude’s Constitution. If Claude helped write these at all, they did a great job of hiding it.
We know that Anthropic dogfoods Claude extensively. Claude now writes 80% of Anthropic’s code, so why not for prose? Why does Anthropic voice their research writeups and policy positions so differently than they voice their product?
Some conjectures, ordered from mundane to wildly speculative.
- We know Dario is an avid essayist. Maybe he prefers to continue writing in his own voice for a public audience, no matter what the LLMs sound like. I can relate!
- Anthropic’s brand trust (such as it is these days) is built on ’the humans are in control’. It’s bad optics to create the appearance of the LLM having even a little influence over the most important things that your company says. Thus, the research and PR pieces should not sound like the product, probably not ever.
- There is a constituency within Anthropic, having enough influence over Claude’s voice, who believe that, for a variety of reasons rhyming with safety and transparency and noticing ‘adversarial’ distillation, Claude’s writing should be distinct and detectable to observant humans. So, they pepper the output with idiolectic markers that don’t show up in every sentence, but in enough sentences that a few paragraphs of Claude text are unmistakable. Separately within Anthropic, the people in charge of copy editing the research and policy essays find these same idiolectic markers exhausting and cliché, so they actively remove them. (Or, maybe they have access to an internal-only model with a different voice.)
If you agree that Claude’s voice makes it less useful for writing to sufficiently-discerning audiences, it seems that that the market will push Anthropic to ’tune it out’ over time. Maybe this is already happening if Fable sounds incrementally less Claude-y than prior models, but Opus 5 sure has it.
What gives?
I’m not multi-lingual enough to answer this, but now I’m wondering: how Claude-y does Claude sound in non-English languages?