一段去年才发生的事 / A thing that happened to me last year
去年我有段时间特别拧巴。具体是工作上一个项目延期延期再延期，我一边觉得是团队的问题、一边又觉得是我自己拖累了团队。睡觉前脑子停不下来，做梦也是关于这个的。
I had a stretch last year where I was just stuck. A work project kept slipping, week after week, and I was oscillating between "this is on the team" and "I'm the one dragging the team down." Couldn't fall asleep without my brain looping. Dreamed about it.
某天凌晨三点我打开 ChatGPT，把整件事——前因后果、各种情绪、我对每个人的看法——花了 40 分钟全写了进去，然后问它："你觉得我在逃避什么？"
One night at 3am I opened ChatGPT, typed out the whole thing for 40 minutes — context, feelings, my read on each person — and asked: "What do you think I'm avoiding?"
它回的东西让我整个人停了一下。它说：你一直在描述别人的问题，但你没怎么提一个细节——这个项目最初是你提出来要做的。你可能在逃避承认：你没那么想做这个项目了，但你不能让自己承认。
What it said stopped me cold. It pointed out: I kept describing other people's problems, but I'd barely mentioned one detail — I was the one who originally proposed the project. It suggested I might be avoiding admitting that I didn't actually want to be working on it anymore, and I wasn't letting myself say it.
它没说错。第二天我跟 manager 谈了，把项目重新切了一下，三周后我整个人状态完全不一样了。
It wasn't wrong. The next day I talked to my manager, restructured the project, and three weeks later I was a different person.
但我也很清楚——这次有效，不代表 LLM 是合格的心理教练。这件事的水比看起来深得多。
But I'm also very aware: this one instance going well does not make an LLM a qualified coach. The water here is deeper than it looks.
LLM 在"自我反思"里能干什么 / What LLMs can actually help with
讲真，把这一年用下来的观察整理一下，LLM 在自我反思这件事上能干的、做得不错的事，主要是三块：
Honestly, after a year of using these things this way, I'd say LLMs are decently helpful at three specific tasks:
1. 当一个"耐心的镜子"。你需要把一件烦心事讲清楚，但讲给朋友太重了、讲给心理咨询师又约不上、自己写日记又不知道从哪开始——这时候 LLM 就是一个永远不嫌烦、不会打断、不会带评判的倾听对象。你只是为了把思路整理出来，让一个看似"在听"的东西在场。这部分价值是真的。
1. As a patient mirror. You have something tangled to work through, but it's too heavy to drop on a friend, your therapist isn't free, journaling alone leaves you blank. An LLM is a listener that never tires, never interrupts, never judges. Sometimes you just need to externalize, and having something that seems to be listening helps. That part is genuinely valuable.
2. 当一个"问题生成器"。你做完一个 SBTI 之外的 Big Five 测试，看到一堆数字不知道下一步问什么——你可以让 LLM 帮你生成"如果你的尽责性是 35 分位，下一周可以观察自己的哪些行为来验证这件事？" 这种把抽象数字翻译成可观察问题的能力，LLM 比你自己干快得多。
2. As a question generator. You just finished a Big Five test and you're staring at numbers, not sure what to do with them. An LLM is much faster than you at translating "your Conscientiousness is at the 35th percentile" into "what specific behaviors over the next week would either confirm or undermine that result?" Abstract scores → observable questions is a translation it does well.
3. 当一个"知识桥梁"。你在心理学某个特定术语上卡住——"自我决定理论是什么？""依恋类型和人格特质是什么关系？"——LLM 能在 30 秒内给你一个比 Wikipedia 更易懂的入门解释，让你不至于因为术语吓退。
3. As a knowledge bridge. You hit a psychology term you don't know — "what is self-determination theory?" "how does attachment style relate to personality traits?" — and an LLM gives you a more accessible explanation in 30 seconds than Wikipedia would. It lowers the entry cost so you don't bounce.
这三块的共同点是：LLM 在"帮你结构化你已经有的材料"和"降低你接触新概念的成本"上，是真的有用。
What these three have in common: LLMs are genuinely useful at structuring material you already have and lowering the cost of meeting new concepts.
但有几个坑 / But there are real traps
接下来这部分我觉得更重要，因为坑都很隐蔽。
This is the part I think matters more — the traps are subtle.
坑 1：谄媚（sycophancy） / Trap 1: sycophancy
这是被研究得最多的一个问题。LLM 在训练过程中被微调成"让用户觉得回答有用"，结果就是它倾向于附和你已有的观点。你说"我觉得 X 同事是 narcissist"，它会帮你分析 X 的 narcissistic 表现，而不是问你"等等，他真的是吗，还是你跟他关系不好"。
This is the most-documented problem. LLMs are fine-tuned during training to "make users feel the response was helpful," which biases them toward agreeing with what you already think. Tell it "I think my colleague X is a narcissist" and it'll happily analyze X's narcissistic behaviors, not stop you with "wait, is he, or do you just have a bad relationship with him?"
在心理咨询里，好的咨询师有相当一部分价值是"在你不舒服的地方按一下"。LLM 默认不会按那一下，因为它的训练目标里没这条。
In real therapy, a big chunk of what a good therapist provides is pressing in the place that hurts. LLMs don't do that by default — it's not in their training objective.
对策：如果你想用 LLM 做反思，显式告诉它"请挑战我的假设，找出我可能没看到的盲点"。给它一个明确的反方向 prompt。这能改善但不能完全消除谄媚倾向。
Mitigation: if you're using an LLM for reflection, explicitly prompt it to "challenge my assumptions and look for the blind spot I might be missing." Giving it the opposite directive helps. It won't fully neutralize the sycophancy, but it shifts things.
坑 2：Eliza 效应 / Trap 2: the Eliza effect
1966 年 MIT 的 Weizenbaum 写过一个超级简单的聊天程序叫 ELIZA，做的事就是把你的话改成反问。结果他发现——用户开始把自己的秘密讲给 ELIZA，并且认真相信 ELIZA 在"理解"他们。他自己被吓到了，后来写了一本书警告这种倾向。
In 1966, MIT's Joseph Weizenbaum wrote a comically simple chatbot called ELIZA. All it did was reflect your statements back as questions. Weizenbaum was alarmed to find that users started confiding secrets to ELIZA and genuinely believing it was understanding them. He wrote a whole book warning about this.
这个倾向叫 Eliza 效应——人会自动把人类特质投射到任何看起来"在回应"的系统上。当年那个程序连一句完整对话都接不上，今天的 LLM 流畅得多——你能想象 Eliza 效应被放大到什么程度。
This is the Eliza effect: humans automatically project human qualities onto anything that appears to be responding. That 1966 program couldn't even maintain a coherent exchange. Today's LLMs are fluent — imagine the Eliza effect scaled up by that much.
我自己用 LLM 用了一年下来，最警惕的一件事就是：我对它说出口的话，开始变得比我对真实朋友说出口的话还多。这事在统计层面是真的——OpenAI 自己披露的用户研究显示有相当一部分用户会把 ChatGPT 当成"主要的倾诉对象"。
After a year of using these, the thing I'm most wary of is this: the volume of things I'd say to the LLM had started to exceed the volume of things I'd say to actual friends. OpenAI itself has published user research showing a non-trivial share of users treat ChatGPT as their primary confidant.
这事对你的关系网是有伤害的。倾诉是关系的养料。当你把它移到一个永远在线、永不嫌烦、但也永远不会真正担心你的对象身上，你和真实人际关系之间的肌肉会萎缩。
This erodes your relational fabric. Confiding is what feeds a relationship. When you move that confiding onto something always-on, never-tired, and incapable of actually worrying about you, the muscle for real relational confiding atrophies.
坑 3：自我合理化的工业化 / Trap 3: the industrialization of self-rationalization
这是我最担心的一条。LLM 让"为自己的行为找一套漂亮理由"这件事，从手工业进入了工业化。
This is the trap I worry about most. LLMs have moved "constructing a flattering rationalization for my behavior" from cottage industry to mass production.
你做了一件你心里知道有点烂的事——把责任推给别人、对伴侣冷暴力一周、没完成承诺——你打开 LLM，按你自己的视角描述一遍。LLM 会基于你给的描述，温柔地、流畅地、看起来很有道理地，给你一套"你之所以这么做是有理由的"叙事。它不是故意骗你，是你给它的输入本来就单边。
You did something you suspect was kind of shitty — passed the blame, gave your partner the cold shoulder for a week, broke a promise. You open the LLM, describe the situation from your side. The LLM, gently and fluently and with the appearance of fairness, constructs a "there are real reasons you did this" narrative for you. It isn't trying to deceive you — the input was one-sided to begin with.
这事的危险在于：它产出的合理化听上去比你自己脑内的版本更"客观"——因为它来自一个"AI"，不是来自你的辩护机制。但它就是你的辩护机制的扩展，只是更流畅。
The danger is this: the rationalization it produces sounds more "objective" than the one bouncing around in your head — because it came from an "AI," not from your own defense mechanism. But it is your defense mechanism. Just laundered.
对策：把同一件事讲给真正的人听——一个朋友、一个心理咨询师、甚至一个不喜欢你的人。看他们的反应和 LLM 的反应差多远。差距就是你的合理化套子有多厚。
Mitigation: tell the same story to an actual human — a friend, a therapist, even someone who doesn't particularly like you. Notice how far apart their reaction is from the LLM's. The distance is roughly how thick your rationalization layer is.
我自己用 LLM 做反思的几条规矩 / The rules I personally follow
经过一年试错，我给自己定了这几条：
After a year of trying and failing at this, here are the rules I've settled on:
-
不让它做诊断。"我是不是有 ADHD" 这种问题，绝对不问 LLM。要么问医生，要么不问。
-
任何重要的人际决策，至少先讲给一个真人听。LLM 可以是第二意见，不是第一意见。
-
当我发现 LLM 在 100% 同意我的时候，强制自己 prompt 它去反驳。
-
每周 LLM 的"心理对话"控制在某个时长内（我现在大概一周 1-2 小时）。超过这个，就感觉自己开始往独处的方向滑。
-
凌晨 2-4 点的对话不当真——我那一次有效是侥幸，更多次的凌晨对话产出的都是垃圾，因为我的判断力本来就在最差的时候。
-
No diagnostic questions. "Do I have ADHD" — never to an LLM. Ask a doctor or don't ask.
-
Any important interpersonal decision gets run by an actual human first. LLM is the second opinion, never the first.
-
When the LLM is agreeing with me 100%, I force myself to prompt it for counter-arguments.
-
I cap "psychological conversations" with LLMs to about 1-2 hours a week. Past that, I notice myself drifting into isolation.
-
2-4am LLM conversations don't count for anything. That one time it worked was luck. Most 3am sessions produce garbage because my judgment is at its worst then.
那个最后的问题 / The last question
回到标题：LLM 时代，"了解自己"变得更难还是更容易？
Back to the headline: in the LLM era, is self-knowledge easier or harder?
我现在的回答是：对会用工具的人，更容易；对依赖工具的人，更难。
My current answer: easier for people who use the tool, harder for people who depend on it.
会用的人，把 LLM 当成一个"廉价的、永远在线的、可以帮你结构化思考的脚手架"——配合真实关系、真实写作、真实身体——他们的自我认知会比 LLM 之前更立体。
People who use it as a cheap, always-on scaffold for structuring thought — alongside real relationships, real writing, real embodied experience — will probably come out of this era with more dimensional self-knowledge than before.
依赖的人，把 LLM 当成主要的倾诉对象、主要的反思工具、主要的"理解我的人"——他们会进入一种很流畅的自我幻觉，自我感觉变得清晰，但实际自我移动得越来越慢。
People who use it as primary confidant, primary reflection tool, primary "thing that understands me" will end up in a strange place where the feeling of self-knowledge gets very clean and crisp, but the actual self stops moving.
这事不取决于 AI，取决于你怎么用它。这话说出来挺老的——但每个新工具的故事都是这个故事。
This doesn't depend on the AI. It depends on how you use it. That's an old sentence — but every new tool's story turns out to be the same story.
如果你想把"自我反思"再加上一层结构，做一次 SBTI 然后把结果拿给 LLM 一起读，可能是个不错的入口。结果是骨架，对话是肉。两个一起用，比单独用任何一个都好——前提是你记得这一节列出来的那些坑。
If you want to add a little structure to your reflection, taking SBTI and reading the results alongside an LLM conversation is a reasonable entry point. The test gives bones; the conversation puts flesh on them. The combination beats either alone — assuming you remember the traps from this piece.
本文是科普与个人观察材料，不构成专业建议。This piece is for educational and reflective purposes; it is not professional advice.