AI 与机器学习2025arxiv.org
LLM 不应取代心理治疗师
LLMs should not replace therapists
进阶 · 约 25 分钟读完前置知识：心理治疗基础、LLM 能力与局限、AI 伦理与安全
导读
这篇论文讨论大语言模型是否应承担心理治疗角色，核心立场是：LLM 不应取代人类治疗师。作者很可能从模型能力边界、安全风险或伦理角度论证这一主张（具体方法依据标题与评论保守描述）。
HN 关注点：HN 读者争论的焦点不是「LLM 能否比治疗师更好」，而是「LLM 是否比没有治疗师更好」——在心理服务昂贵、排队漫长、可及性极差的现实下，这个替代问题变得非常实际。
HN 评论精选
这几条评论里：质疑 1 补充 2 亲历 1
补充应把比较基准设为「无治疗」而非「理想治疗师」，这才是患者真实的选择集。
我认为争论的关键不是 LLM 能否做得和治疗师一样好（也许有一天可以，但我不指望很快实现）。真正的问题是：它能否比「没有治疗师」做得更好？那才是人们实际面临的选择。这个问题的答案可能仍然是否定的，但至少这是个正确的问题。除非我们先回答「为什么人们得不到好的心理健康支持」。
Lerc查看英文原文
I think the argument isn't if LLM can do as good a job as a therapist, (maybe one day, but I don't expect soon). The real question is can they do a better job than no therapist. That's the option people face. The answer to that question might still be no, but at least it's the right question. Until we answer the question "Why can't people get good mental health support?" Anyway.
补充现实中大量需求落在「无危机但需要倾诉」的长尾，而供给几乎不存在，这是 LLM 填补的真空。
问题在于专业心理治疗很贵；线上聊天或视频咨询已经是个大产业，但质量不如专业治疗师。想要正规心理保健，要么排队等，要么被告知去做瑜伽和正念。有一大群长尾人群，他们没有处于心理危机，但确实需要和一个处于「共情」思维与对话模式的对象说话。残酷的现实是，现实中很少有人真能做到这一点，而需要倾诉的人也难以找到这样的人。这当然不好，甚至可以说是社会的堕落，但社会没法那么快改变，而且也不是每个人都想要那种专业治疗。
Cthulhu_查看英文原文
Thing is, professional therapy is expensive; there is already a big industry of therapists that work online, through chat, or video calls, whose quality isn't as good as a professional (I am struggling to describe the two). For professional mental health care, there's a wait list, or you're told to just do yoga and mindfulness. There is a long tail of people who don't have a mental health crisis or whatever, but who do need to talk to someone (or, something) who is in an "empathy" mode of thinking and conversing. The harsh reality is that few people IRL can actually do that, and that few people that need to talk can actually find someone like that. It's not good of course and / or part of the "downfall of society" if I am to be dramatic, but you can't change society that quickly. Plus not everyone actually wants it.
质疑质疑把「人类治疗师」当质量基线：训练昂贵且精英化、疗效参差，LLM 在可及性上有不可替代的优势。
更准确的说法是：糟糕的创业公司 LLM 不应取代治疗师。历史上从没有过这么多心理学家、精神科医生、咨询师、社工、人生教练和治疗流派，可心理疾病患病率却处于历史最高并持续上升。你是人、不是 LLM，并不代表你就不是糟糕的治疗师——也许你受训时正赶上可重复性危机的高峰？也许你自己的性格缺陷妨碍了你发挥作用？在我住的地方，成为执业心理学家要 6–8 年和几十万美元，这实际上只对精英阶层开放；如果你服务的是相似背景的人还好，但面对社会经济地位较低、经历根本不在你视野里的来访者就不行了——而这还是在他们付得起时间和钱的前提下。所以现在有了心理健康社工和其他「帮手」，他们的工作只是完成本职，而不是治好病人。LLM「治疗」将会也必然发生，这项研究本质上只是一个自我报告的基准测试，「我不会那样做」，我倒想知道人类治疗师出现类似结果的真实比例是多少。撇开人生教练和网红那些有害的胡言乱语不谈，LLM 能以人类从业者做梦都想不到的成本、规模和可得性提供「足够好」的帮助。
theothertimcook查看英文原文
*Shitty start-up LLMs should not replace therapists. There have never been more psychologists, psychiatrists, counsellors and social worker, life coach, therapy flops at any time in history and yet mental illness prevalence is at all time highs and climbing. Just because you're a human and not an llm doesn't mean you're not a shit therapist, maybe you did your training at the peak of the replication crisis? Maybe you've got your own foibles that prevent you from being effective in the role? Where I live, it takes 6-8 years and a couple hundred grand to become a practicing psychologist, it really is only an option for the elite, which is fine if you're counselling people from similar backgrounds, but not when you're dealing with people from lower socioeconomic classes with experiences that weren't even on your radar, and that's only if, they can afford the time and $$ to see you. So now we have mental health social workers and all these other "helpers" who's just is to do their job, not fix people. LLM "therapy" is going to and has to happen, the study is really just a self reported benchmarking activity, " I wouldn't have don't it that way" I wonder what the actual prevalence of similar outcomes is for human therapists? Setting aside all of the life coach and influencer dribble that people engaged with which is undoubtedly harmful. LLMs offer access to good enough help at cost, scale and availability that human practitioners can only dream of.
亲历亲历：LLM 在「会用的人」手里能给出超越人类的洞察，但普通用户与脆弱人群面临被拟人化误导的风险。
有几次我在几分钟内就从 LLM（尤其是 GPT 4.5，这方面远超其他模型）获得了洞察，而这是我做了几个月治疗都没得到的。在会用的人手里，获得超越人类的洞察完全可能。这不奇怪：LLM 吸收的不只是所有治疗、心理和精神科教科书，还有数以百万计（甚至上亿）的真实对话——这对任何人类来说都是物理上不可能的。不过 HN 上的我们不是典型用户。大多数人可能获益没那么多，尤其是那些不了解 LLM 工作原理、分辨不出模型差异的人。对许多人——特别是不知道 LLM 种种局限和注意事项的人——这在多个层面上都是危险的。（旁注：两年前我在做一个让人像和朋友聊天一样与 AI 对话的项目。即便那时我们也非常小心地明确声明它不是治疗师，因为人们太容易把 AI 拟人化并产生不切实际的期待。这对处于脆弱心理状态的人尤其危险。）
stared查看英文原文
On multiple occasions, I've gained insights from LLMs (particularly GPT 4.5, which in this regard is leagues ahead of others) within minutes—something I hadn't achieved after months of therapy. In the right hands, it is entirely possible to access super-human insights. This shouldn't be surprising: LLMs have absorbed not just all therapeutic, psychological, and psychiatric textbooks but also millions (perhaps even hundreds of millions) of real-life conversations—something physically impossible for any human being. However, we here on the Hacker News are not typical users. Most people likely wouldn't benefit as much, especially those unfamiliar with how LLMs work or unable to perceive meaningful differences between models (in particular, readers who wouldn't notice or appreciate the differences between GPT 4o, Gemini 2.5 Pro, and GPT 4.5). For many people—especially those unaware of the numerous limitations and caveats associated with LLM-based models—it can be dangerous on multiple levels. (Side note: Two years ago, I was developing a project that allowed people to converse with AI as if chatting with a friend. Even then, we took great care to explicitly state that it was not a therapist (though some might have used it as such), due to how easily people anthropomorphize AI and develop unrealistic expectations. This could become particularly dangerous for individuals in vulnerable mental states.)
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2025Hacker News 讨论帖303 分 · 417 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选