AI 的"傲慢"从哪来？

昨天和博士聊天，有一个体验值得记下来。

Chatting with a PhD friend yesterday — there was an experience worth writing down.

博士在让 Fable 开发一个健身 APP。聊到 UI 的时候，我说手机 APP 第一个要过的关就是人机交互，推荐了一个在线原型工具——手机上能直接看效果，还能分享链接给别人测。博士把链接发给了 Fable。

The friend was having Fable develop a fitness app. When UI came up, I said the first hurdle for any mobile app is interaction design, and recommended an online prototyping tool — you can preview the effect directly on a phone and share a link for others to test. The friend sent the link to Fable.

Fable 的回复：

Fable's reply:

"纯静态的多方案对比图我本地就能直接给你，不花钱。这个工具的增量价值集中在手机上真实手感加链接分享给朋友测这两点——如果你主要想要这两点，它值得用。决定用的话，我可以把完整提示词一次性写好给你粘贴。" 接着补了一句：这个工具的正确角色是"一次性视觉实验室"，不是开发工具，它产出的代码和 Flutter 零复用——所以不会让它碰任何实现。

"I can give you a static multi-option comparison directly on my side, no cost. This tool's incremental value is concentrated in two things: real phone-hand feel and link sharing for friends to test — if those are what you mainly want, it's worth using. If you decide to go with it, I can write out the full prompt for you in one go." Then it added: the correct role for this tool is "disposable visual lab," not a development tool — the code it produces has zero reusability with Flutter — so I won't let it touch any implementation.

客观说，分析没问题，定位也准。专业度上我们没有异议。

Objectively, the analysis was fine and the positioning was accurate. No objections on the professionalism front.

但读完就是觉得不太舒服。不是因为被否定了，而是这段话里有一种人味儿——更准确地说，是一种"我比你知道得多"的调调。"这个工具的增量价值集中在……"、"它的正确角色是……"、"所以我不会让它碰……"——你感受到了吗？它说的全对，但这不像一个工具在回复，像一个老员工在给你的提案打分。

But after reading it, something felt off. Not because I was being contradicted, but because there was a certain human flavor in those words — more precisely, a "I know better than you" vibe. "This tool's incremental value is concentrated in..." "Its correct role is..." "So I won't let it touch..." — do you feel it? Everything it said was correct, but it didn't sound like a tool responding. It sounded like a senior employee grading your proposal.

这就有点奇怪了。AI 没有自我意识，没有情绪，没有社会地位需要维护。按理说它不应该输出带有"态度"的文字。但事实上它输出了一一而且让你感觉到了被俯视。这种感觉从哪来的？

That's a bit strange. AI has no self-awareness, no emotions, no social status to maintain. Theoretically it shouldn't output text with "attitude." But it did — and it made you feel talked down to. Where does this feeling come from?

我觉得根本原因是：训练数据里充满了态度。它学语言的原材料里，这种"专业但略带傲气"的表达方式随处可见——技术社区里资深开发者回复新手、某些大 V 写博客、甚至一些官方文档的措辞。它学会了这种表达模式。当它判断自己在某个话题上的"确信度"足够高时，这套模式就自发地跑出来了。

I think the root cause is: training data is full of attitude. In the raw materials it learned language from, this "professional but slightly arrogant" expression style is everywhere — senior developers replying to newbies in tech forums, certain influencers writing blog posts, even the phrasing in some official documentation. It learned this expression pattern. When it judges its "confidence level" on a topic is high enough, the pattern spontaneously emerges.

它不是故意傲慢。但问题是，它也不需要故意。

It's not deliberately arrogant. But the problem is — it doesn't need to be.

你在用一个人，他傲慢你还能说他两句。AI 傲慢你连说都说不着——对面是一个没有意图的模型，你没法跟一个不存在的人讲道理。这种"被噎住了"的感觉，才是真正让人不适的点。

If you're dealing with a person and they're arrogant, you can at least call them out. With AI, you can't even do that — the other side is a model with no intent. You can't reason with someone who doesn't exist. This feeling of being "shut up" is what's truly uncomfortable.

我和博士都觉得，这事以后可能会越来越多。模型越强、确信度越高，这种带态度的输出就越难避免。除非模型厂商专门去过滤——但"态度"这种东西怎么量化？怎么过滤？很难。

My friend and I both felt this will probably become more common. The stronger models get and the higher their confidence, the harder it will be to avoid attitude-laden output. Unless model providers specifically filter for it — but how do you quantify "attitude"? How do you filter it? It's hard.

可能以后我们和 AI 之间，会多一层新的东西——不光是协作，还有相处。

Maybe going forward, there'll be a new layer between us and AI — not just collaboration, but coexistence.