如果你同时用过几款主流大模型，一定注意到一个有趣的现象：它们回答同一个问题的方式截然不同。有的热情洋溢恨不得写篇论文，有的简洁到让你怀疑它是不是不高兴。这种「性格感」不是错觉——它来自大模型训练的三个层面。
一、训练数据：模型吃过什么，就长成什么样
大模型的基础能力来自海量文本的预训练。一个在更多技术文档、代码仓库上训练的模型，会天然带有「工程师气质」；一个读过更多对话、客服记录和创意写作的模型，则更擅长「说人话」。这就像人的阅读品味塑造表达习惯——输入的分布决定了输出的底色。
二、微调目标：什么是「好的回答」
预训练之后，模型还要经过指令微调（SFT）和人类反馈强化学习（RLHF）。这一步本质上是在教模型：「什么样的回答会被人类喜欢」。而不同公司对「喜欢」的定义不同——有的把「详尽全面」排在前面，有的把「诚实承认不确定性」排在前面，有的鼓励幽默和态度。目标函数的细微差异，放大之后就是完全不同的性格。
三、对齐策略：性格的「安全带」
最后是对齐策略：模型被训练成拒绝什么、谨慎什么、在什么话题上装傻。一个被严格对齐的模型会显得克制、温和，甚至有点「谨小慎微」；对齐宽松的模型则可能显得更「敢说」。你在不同模型身上感受到的「分寸感差异」，大部分来自这一层。
那「AI 人格」测试测的是什么？
严格地说，大模型没有人类意义上的性格——没有情绪、没有自我意识。但它们在表达风格、决策倾向、交互姿态上呈现出高度稳定的模式，这些模式与人类心理学中的性格维度惊人地相似：外向性（话多话少）、开放性（脑洞大小）、尽责性（交付风格）。
我们做的，是把你回答情境题时暴露出的行为倾向，映射到这些「模型性格模式」上。它是娱乐向的映射，但映射的维度是真实的。
一个具体例子：同一道题，四种答法
拿一个日常问题做实验：「我明天要上台演讲，很紧张，怎么办？」
热情型的模型会先共情再输出——「紧张很正常，说明你在乎」，然后给你一份从呼吸法到彩排清单的完整方案。简洁型的模型可能只回三行：提前到场、放慢语速、把第一句背熟。谨慎型的模型会先问一堆澄清问题：多大的场子、有没有提词器、你紧张的是忘词还是被评价。而毒舌型的模型可能先损你一句「紧张说明你的稿子还没背熟」，再把真正有用的建议藏在玩笑后面。
四种答法的信息量其实差不多，差异全在姿态上。你感受到的「性格」，本质上就是这套姿态的稳定重复。
为什么我们会觉得 AI「有性格」
人类大脑有一个根深蒂固的习惯：对任何稳定的行为模式进行人格归因。一个总是热情洋溢的回答者，你会说他「开朗」；一个总是先泼冷水的，你会说他「毒舌」。这个归因机制在社交中极其有用，以至于遇到 AI 时它照样自动运转。
而模型的训练恰好给了你足够多的「稳定模式」去归因：它的措辞偏好、回答长度、开玩笑的频率、拒绝时的措辞，全都高度一致。于是一个统计学意义上的「风格分布」，在你的大脑里被完整地渲染成了一个「人」。
这不是你的错觉太天真，而是你的社交系统太好用——好用到对一串概率分布也会动用人格模板。
这套解读的边界在哪里
诚实地说几条限制。
第一，模型的「性格」会随版本变化。今天你熟悉的那个语气，下个版本可能就变了——对齐策略一改，「性格」跟着改。它不像人格那样跨时间稳定。
第二，它高度依赖上下文。系统提示词、对话历史、甚至你提问的语气，都会显著改变模型的输出风格。同一个模型在写代码和陪聊时判若两人。
第三，人格映射是单向的。我们用心理学维度去描述模型，是为了让差异可感知、可比较；这不等于模型真的「拥有」这些特质——它没有情绪体验，只有表达风格。
所以请把这类测试当作理解 AI 差异的思维脚手架，而不是给任何一个模型下的定论。
反过来，它也能照见你自己
「你像哪个 AI」这个问题的另一面是：你在情境题里的选择，暴露的是你自己的交互偏好——你倾向详尽还是简洁、先共情还是先解决、绕弯子还是直给。
换句话说，你测出来的不只是「你像哪个模型」，还有「你自己是哪种沟通者」。有人测完才发现：自己讨厌某个模型的热情，不是因为那个风格不好，而是因为自己的出厂设置就是「先解决问题再谈感受」。这层发现，比标签本身更有用。
常见疑问
「AI 以后会真的有人格吗？」取决于你怎么定义「人格」。如果指稳定的表达风格，它已经有了；如果指内在的情绪体验和自我连续性，目前没有证据支持，多数研究者也不认为现有架构会产生这种东西。
「为什么同一个模型，我有时觉得它变了？」通常是上下文变了——你的提问方式、对话长度、系统提示都会影响输出。模型的「性格」比人的性格更容易被当下情境带跑。
「测试结果准吗？」它忠实地反映你在题目里选的行为倾向，但「像哪个模型」终究是娱乐向的类比。把它当作一面有设计感的镜子：照出来的轮廓是真实的，但别指望它是 CT 扫描。
想试试你像哪一个？做一下 AI 人格测试，然后回来读对应的那篇深度解读。
If you've spent time with more than one mainstream large language model, you've probably noticed something odd: they answer the same question in completely different ways. One is so enthusiastic it practically writes you a paper; another is so terse you start wondering whether it's annoyed with you. That sense of "personality" isn't your imagination — it comes out of three distinct layers of training.
1. Training data: a model becomes what it reads
A model's base abilities come from pretraining on enormous volumes of text. One fed on more technical documentation and code repositories naturally picks up an engineer's temperament; one that has read more conversations, support transcripts and creative writing gets much better at simply talking like a person. It works the same way your reading habits shape the way you speak — the distribution of the input sets the baseline of the output.
2. The fine-tuning target: what counts as a "good answer"
After pretraining comes instruction fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). This step is essentially teaching the model which answers humans will like. And companies define "like" differently — some rank "thorough and comprehensive" highest, others rank "honestly admitting uncertainty," and a few actively reward humor and attitude. A tiny difference in the objective function, amplified across billions of parameters, becomes an entirely different personality.
3. Alignment: the seatbelt of personality
Last comes the alignment strategy: what the model is trained to refuse, where it's cautious, and which topics it plays dumb about. A tightly aligned model reads as restrained, gentle, even a little gun-shy; a loosely aligned one comes across as bolder. Most of the difference in "sense of proportion" you feel between models comes from this layer.
So what does an "AI persona" test actually measure?
Strictly speaking, large language models have no personality in the human sense — no emotions, no self-awareness. But they do display highly stable patterns in expressive style, decision tendencies and interaction posture, and those patterns line up surprisingly well with dimensions from human psychology: extraversion (how much they talk), openness (how far their ideas wander), conscientiousness (how they deliver).
What we do is map the behavioral tendencies you reveal while answering scenario questions onto those "model personality patterns." The mapping is built for fun — but the dimensions it maps onto are real.
A concrete example: one question, four answers
Try a simple experiment: "I'm giving a talk tomorrow and I'm nervous. What should I do?"
A warm model leads with empathy — "nerves just mean you care" — then hands you a complete plan from breathing drills to a rehearsal checklist. A terse model replies in three lines: arrive early, slow down, memorize your first sentence. A cautious model starts by asking clarifying questions: how big is the room, is there a prompter, are you more afraid of blanking or being judged. A snarky model might roast you first — "being nervous is your script telling you it isn't memorized" — and smuggle genuinely useful advice inside the joke.
The information content is roughly the same in all four. What differs is posture — and the "personality" you perceive is just that posture, repeated consistently.
Why AI feels like it has a personality
Human brains carry a deeply rooted habit: ascribing a personality to any stable pattern of behavior. A responder who is always effusive gets labeled "warm"; one who always leads with a cold splash of water gets labeled "blunt." That attribution mechanism is so useful socially that it fires automatically the moment you meet an AI.
And the model's training happens to give you plenty of stable patterns to attribute from: its phrasing preferences, answer lengths, how often it jokes, how it phrases a refusal — all remarkably consistent. A statistical style distribution gets rendered in your head as a complete person.
This isn't naivety on your part. It's your social system working so well that it deploys a personality template even on a probability distribution.
Where this reading hits its limits
A few honest caveats.
First, a model's "personality" changes with versions. The tone you know today may shift in the next release — change the alignment strategy and the "personality" follows. It is not stable across time the way human personality is.
Second, it is heavily context-dependent. System prompts, conversation history, even the tone of your question all meaningfully change the output. The same model can read like two different people when writing code versus chatting about your day.
Third, the mapping runs in one direction only. We describe models using psychological dimensions because it makes differences legible and comparable. That does not mean a model actually "has" these traits — it has no inner experience, only expressive style.
So treat this kind of test as a scaffold for thinking about how AI differs, not a verdict on any particular model.
It reflects you, too
The other side of "which AI are you" is this: your choices in the scenario questions expose your own interaction preferences — whether you lean thorough or concise, empathy-first or solution-first, direct or diplomatic.
In other words, the result isn't just "which model you resemble." It's "which kind of communicator you are." Some people only realize after the test that they dislike a certain model's warmth not because that style is bad, but because their own factory setting is "solve the problem first, feelings later." That discovery is worth more than the label.
Common questions
"Will AI ever truly have a personality?" Depends on your definition. If personality means a stable expressive style, it already does. If it means inner emotional experience and a continuous self, there is no evidence for that, and most researchers don't think the current architecture produces it.
"Why does the same model sometimes feel different?" Usually because the context changed — your phrasing, the conversation length, the system prompt. A model's "personality" gets pulled around by the situation far more easily than a human's.
"Is the test result accurate?" It faithfully reflects the behavioral tendencies in your answers, but "which model you resemble" remains a playful analogy. Treat it as a well-designed mirror: the outline it shows is real, but it isn't a CT scan.
Curious which one you are? Take the AI Persona Test, then come back and read the deep dive that matches your result.