AI Model Groupthink

Ask multiple AI models for a comparable opinion, and Chinese models seem to be more likely to respond in line with the wider consensus.向多个 AI 模型询问可比的观点时，中国模型似乎更容易给出符合更广泛共识的回答。

I recently built/vibed a small web tool that asks a panel of 12 AI models any kind of “top three” question and tallies the answers like an Olympic medal table. The idea is to have a quick way to fetch hivemind recommendations on anything from a good book to read, things to do in a city, or a low-stakes purchase. 最近，我构建/尝试了一个小型网络工具，它会向 12 个 AI 模型组成的评审团提出任何类型的“前三名”问题，并像奥运会奖牌榜一样统计答案。其初衷是提供一种快速获取集体智慧建议的方法，无论是关于阅读什么好书、在某个城市做什么，还是进行小额消费。

The system prompt is simple and built around this request:系统提示词很简单，围绕以下要求构建：

You are answering a ranking question with your own independent opinion. Give your top 3 answers, ranked from 1 (best) to 3.你正在以自己独立的观点回答一个排名问题。请给出你的前 3 名答案，按从 1（最佳）到 3 的顺序排列。

The call explicitly asks each model for its own view, not to guess what the crowd thinks.调用指令明确要求每个模型表达自己的观点，而不是猜测大众的想法。

There is some normalization required: if we ask for the best castles in Ireland and Luna votes “Rock of Cashel” while Gemini votes “Cashel Rock”, we want to treat that as a consensus. The app uses Gemini Flash to adjudicate this.这需要进行一些标准化处理：如果我们询问爱尔兰最好的城堡，Luna 投了“卡舍尔岩”（Rock of Cashel），而 Gemini 投了“卡舍尔石”（Cashel Rock），我们希望将其视为共识。该应用使用 Gemini Flash 来进行裁定。

The results – from conformist to contrarian结果——从随大流到反其道而行之

A word on reading the chart. Long bars and a winner at the top make it look like a leaderboard, as if more consensus is better. It isn’t necessarily. However, having clicked around a few of the queries: outlier responses are more likely to be interesting than to be an insightfully correct response that bests the responses of all the other models. I like that Gemini recommends you paint an adult bedroom terracotta while all the other models recommend green, cream, and greige, but I understand that many people will want their bots to stick to the cultural mainstream thank you very much.关于如何解读图表的一点说明。长条图和顶部的获胜者使其看起来像排行榜，仿佛共识越多越好。事实未必如此。然而，在点击查看了几个查询后发现：离群的回答往往比那些胜过所有其他模型、被认为“深刻正确”的回答更有趣。我喜欢 Gemini 建议将卧室漆成赤陶色，而所有其他模型都推荐绿色、奶油色和灰米色，但我理解许多人希望他们的机器人坚持文化主流，谢谢。

Each model is scored against the consensus of the other eleven, getting full credit for the right item in the right podium spot and half credit for the right item in the wrong spot. A score of 1.0 would mean perfect agreement with the winning conclusions of the rest of the group; 0 would mean total contrarianism.每个模型都会根据其他十一个模型的共识进行评分，在正确的领奖台位置给出正确的项目可获得满分，而在错误的位置给出正确的项目则获得半分。得分为 1.0 意味着与小组其余成员的获胜结论完全一致；0 则意味着完全反其道而行之。

The top four conformists are all Chinese: DeepSeek, Kimi K2, Tencent’s Hy3 and GLM.排名前四的“随大流者”全部来自中国：DeepSeek、Kimi K2、腾讯的 Hy3 和 GLM。

Group the scores by country and the gap holds. The five Chinese models average 0.47, the six American models 0.38, and Europe’s lone entrant, Mistral, scores 0.36.按国家对分数进行分组，差距依然存在。五个中国模型的平均分为 0.47，六个美国模型为 0.38，而欧洲唯一的参赛者 Mistral 的得分为 0.36。

The exception is Qwen, Alibaba’s model, which scores 0.38 and sits among the American models, just behind Grok and Nemotron (both 0.40). At the other end, the two most contrarian models are Meta’s Llama 4 Scout (0.34) and Anthropic’s Claude Haiku (0.32).例外是阿里巴巴的模型 Qwen，其得分为 0.38，位于美国模型之间，仅次于 Grok 和 Nemotron（均为 0.40）。在另一端，最反其道而行之的两个模型是 Meta 的 Llama 4 Scout（0.34）和 Anthropic 的 Claude Haiku（0.32）。

Detailed table:详细表格：

Why would Chinese models agree more?为什么中国模型更容易达成一致？

I don’t know. Here are some explanations I find plausible, roughly in order.我不知道。以下是我认为合理的一些解释，大致按顺序排列。

1. Distillation1. 蒸馏

Distillation means training a model on another model’s outputs, essentially a cloning technique. US labs, including OpenAI and Anthropic, have publicly accused Chinese labs of distilling their models at scale. Earlier this year Anthropic named DeepSeek and Kimi-maker Moonshot (and also MiniMax, which is not represented in the sample here) as running large distillation campaigns against Claude. That does not establish that distillation caused the pattern here, or even that these models were trained that way. But if a model was trained heavily on outputs from other models, one expected outcome is that its preferences converge with those of source models, especially when they converge. A model built that way is partly a consensus machine. It is striking that DeepSeek and Kimi, first and second in the table, are among the labs most often named in those accusations.蒸馏是指在另一个模型的输出上训练模型，本质上是一种克隆技术。包括 OpenAI 和 Anthropic 在内的美国实验室曾公开指责中国实验室大规模蒸馏其模型。今年早些时候，Anthropic 点名 DeepSeek 和 Kimi 的开发者月之暗面（以及 MiniMax，此处样本中未包含）针对 Claude 进行了大规模的蒸馏活动。这并不能确定蒸馏导致了这里的模式，甚至不能确定这些模型就是那样训练的。但如果一个模型大量基于其他模型的输出进行训练，预期的结果之一就是其偏好与源模型趋同，尤其是在源模型本身趋同的情况下。以这种方式构建的模型在一定程度上是一台共识机器。值得注意的是，表格中第一和第二名的 DeepSeek 和 Kimi 正是那些指责中最常被点名的实验室。

2. Chinese models learning from each other2. 中国模型相互学习

The Chinese open-weight ecosystem is tightly knit. Labs release weights, publish methods and train on synthetic data produced by each other’s models. That could be giving the five Chinese models a family resemblance.中国的开源权重生态系统联系紧密。实验室发布权重、发表方法，并使用彼此模型生成的合成数据进行训练。这可能赋予了这五个中国模型某种家族相似性。

This matters because of how the score works. Five of the twelve models here are Chinese, so if they share a style, they help build the consensus they are scored against. A bloc that votes together will always look conformist. (Qwen breaking from the bloc is a point against this, or at least a sign the family isn’t that close.)这一点很重要，因为评分机制的运作方式如此。这里的十二个模型中有五个是中国模型，所以如果它们共享一种风格，它们就在帮助构建它们所依据的共识。一个步调一致的集团总是看起来很随大流。（Qwen 脱离该集团是一个反例，或者至少说明这个家族关系并没有那么紧密。）

3. Model size and tier3. 模型规模与层级

The bottom of the table is crowded with smaller, cheaper models: Llama 4 Scout, Claude Haiku and Mistral Small. Smaller models hold less of the shared cultural canon in memory, so their picks are more scattered. Gemini Flash-Lite and DeepSeek’s Flash model are counter-examples, so size can’t be the whole story.表格底部挤满了更小、更便宜的模型：Llama 4 Scout、Claude Haiku 和 Mistral Small。较小的模型在内存中持有的共享文化准则较少，因此它们的选项更加分散。Gemini Flash-Lite 和 DeepSeek 的 Flash 模型是反例，所以规模不能解释全部。

4. Knowledge cutoff4. 知识截止日期

About half of new text appearing on crawl-worthy URLs is written by humans, and half by AI. A model trained more recently has read more of the slop. Most of the Chinese models here are mid-2026 releases. Again there are counter-examples: Kimi (knowledge through 2024-12) is second, while Luna has the newest cutoff on the panel and sits ninth.出现在可抓取 URL 上的新文本中，大约一半由人类撰写，一半由 AI 撰写。最近训练的模型阅读了更多的垃圾内容。这里的大多数中国模型是 2026 年中期的版本。同样也有反例：Kimi（知识截止至 2024 年 12 月）排名第二，而 Luna 在评审团中拥有最新的截止日期，却排名第九。

5. Post-training personality5. 后训练个性

Perhaps US labs are investing in post-training in a way that gives more distinctive character, more willingness to pick the less obvious answer. If there’s a trade-off between personality and benchmark-maxxing that may also play into the flipside, where the Chinese labs are plausibly focusing more on hitting benchmarks.也许美国实验室在后训练方面的投入赋予了模型更独特的个性，更愿意选择不那么明显的答案。如果个性与基准测试最大化之间存在权衡，这也可能解释了反面情况，即中国实验室可能更专注于达到基准测试分数。

6. The English canon6. 英语准则

All prompts are in English. A model whose English training data leans on the most widely syndicated sources would tend to give the most canonical English-language answers, which is exactly what consensus rewards.所有提示词均为英语。一个英语训练数据偏向最广泛联合来源的模型，往往会给出最符合英语语言准则的答案，而这正是共识所奖励的。

Caveats注意事项

More work needed, because…需要做更多工作，因为……

• Small dataset. This is based on the first few hundred questions from my own usage and some humans in the wild who clicked from Reddit. 数据集较小。这是基于我自己的使用习惯以及一些从 Reddit 点击进来的真实用户提出的几百个问题。
• Arbitrary panel composition. The measure of which answers are normal is itself defined via the panel’s composition. 6 US, 5 Chinese, and 1 European was an arbitrary choice by me based on my sense of the industry breakdown. More practical decisions were to have max one model per lab and avoid expensive models. Plausible additions would include models from MiniMax and ByteDance. At a stretch, and in the spirit of geo-diversity, the panel could also include Aleph Alpha (Germany), Apertus (Switzerland), Cohere (Canada), Falcon (UAE), Sarvam (India) or LG’s Exaone (South Korea). Adding flagships like Claude Sonnet, alongside the cheap tiers, might also show whether price tier matters.评审团构成具有随意性。哪些答案是“正常”的衡量标准本身就是通过评审团的构成来定义的。6 个美国模型、5 个中国模型和 1 个欧洲模型是我根据对行业结构的理解做出的随意选择。更实际的决定是每个实验室最多选一个模型，并避免使用昂贵的模型。合理的补充可能包括来自 MiniMax 和字节跳动的模型。如果从地理多样性的角度考虑，评审团还可以包括 Aleph Alpha（德国）、Apertus（瑞士）、Cohere（加拿大）、Falcon（阿联酋）、Sarvam（印度）或 LG 的 Exaone（韩国）。除了廉价层级外，增加 Claude Sonnet 等旗舰模型也可能显示价格层级是否重要。
• Agreement isn’t accuracy. A high score means a model says what others say, not that it’s right. On “best” questions there is often no right answer at all.一致性不等于准确性。高分意味着模型在说别人说的话，并不意味着它是对的。在“最佳”这类问题上，通常根本没有正确答案。
• English-language questions. The questions so far lean towards an English-speaking, Western frame.英语问题。到目前为止，这些问题倾向于英语国家和西方的框架。

Conclusion结论

Back to that bedroom recommendation:回到那个卧室装修的建议：

• 11 x models recommended green, cream or greige11 个模型推荐了绿色、奶油色或灰米色
• 1 x model recommended terracotta1 个模型推荐了赤陶色

Which of these is more correct, and which is behavior we want to encourage? I suspect most of us want an AI ecosystem that isn’t a groupthink borg. We like a bit of Temperature. We want models to disagree, surprise us, occasionally push for the terracotta bedroom. I also suspect that, as users, we tend to reward models that give us answers that feel safe. These two instincts are going to be pulled apart by a lot of money over the next few years.哪一个更正确，哪一种行为是我们想要鼓励的？我怀疑我们大多数人想要的 AI 生态系统不是一个群体思维的博格集合体。我们喜欢一点点“温度”（随机性）。我们希望模型有分歧，给我们惊喜，偶尔坚持推荐赤陶色的卧室。我也怀疑，作为用户，我们倾向于奖励那些给出感觉“安全”答案的模型。在未来几年里，这两个本能将被大量的资金所拉扯。

Multiply this out to millions of users seeking suggestions and it becomes something bigger: a quiet, constant pressure towards a lab-mediated consensus middle.将这种情况放大到数百万寻求建议的用户身上，它就变成了一件更大的事情：一种向实验室中介共识中心靠拢的、安静而持续的压力。

Commercially, being included in that consensus middle is already lucrative. For two decades companies have poured huge amounts of capital into efforts to rank well in search engines (SEO), and the smart ones have already switched to competing over placement in AI responses (AIO/GEO). Self-reinforcing loops will escalate the value of this prize. If the labs are competitively exfiltrating each other’s responses, getting a product highly placed in one model’s answers could enable it to flow into the next model’s training data, and the next. Consensus that was gamed once can be copied until it looks like a fact about the world.在商业上，被纳入那个共识中心已经有利可图。二十年来，公司投入了巨额资本试图在搜索引擎中获得好的排名（SEO），聪明的公司已经转向竞争 AI 回答中的位置（AIO/GEO）。自我强化的循环将提升这一奖项的价值。如果实验室在竞争性地窃取彼此的回答，让一个产品在某个模型的回答中获得高位，就可能使其流入下一个模型的训练数据，以此类推。曾经被操纵的共识可以被复制，直到它看起来像是一个关于世界的客观事实。

Paint colours are a harmless starting point, but the same machinery will answer questions about which candidate to trust, which news source is reliable, which history is true. Decorating recommendations are obviously not equivalent to political or factual judgments, and models may behave very differently across domains, but the same basic machinery increasingly mediates all of them. When the models are copying each other’s opinions and feeding those opinions back into the real world, a consensus doesn’t need to be right. It only needs to be first.油漆颜色是一个无害的起点，但同样的机制将回答关于该信任哪位候选人、哪个新闻来源可靠、哪段历史真实的问题。装饰建议显然不等同于政治或事实判断，模型在不同领域的表现可能大相径庭，但同样的底层机制正日益中介着所有这些领域。当模型在复制彼此的观点并将这些观点反馈回现实世界时，共识不需要是正确的，它只需要是第一个出现的。

For more decorating or political advice, you can ask the robots for recommendations here.如需更多装饰或政治建议，您可以在此处向机器人寻求推荐。

---

Links链接

Run queries here: top3.best 在此处运行查询：top3.best

Stats page here: top3.best stats page统计页面在此处：top3.best 统计页面

References for model data:模型数据参考：

• Claude Haiku 4.5 overview, AnthropicClaude Haiku 4.5 概述，Anthropic
• DeepSeek V4 Flash, model trackerDeepSeek V4 Flash，模型追踪器
• GLM-5.3-Flash specifications, APIYIGLM-5.3-Flash 规格，APIYI
• Hy3 launch coverage, Unite.AIHy3 发布报道，Unite.AI
• Gemini 3.1 Flash-Lite, Google Cloud docsGemini 3.1 Flash-Lite，Google Cloud 文档
• Grok 4.3, Vercel AI GatewayGrok 4.3，Vercel AI Gateway
• Nemotron 3 Super model card, NVIDIANemotron 3 Super 模型卡，NVIDIA
• Qwen-Plus, OpenRouterQwen-Plus，OpenRouter
• Mistral Small 4, OpenRouterMistral Small 4，OpenRouter

Implementation notes: The twelve models were called through OpenRouter, mostly in their cheaper “flash” or small variants, with reasoning switched off where the provider allows it. “Default” means the model was run with whatever reasoning behaviour the provider ships. I did not touch the Temperature parameter, nor did I explore what defaults the different models use.实施说明：这十二个模型是通过 OpenRouter 调用的，大多使用其更便宜的“flash”或小型变体，并在提供商允许的情况下关闭了推理功能。“默认”意味着模型在运行时使用了提供商预设的任何推理行为。我没有触碰 Temperature 参数，也没有探索不同模型使用了哪些默认设置。