AI 与机器学习2021nature.com
人脸识别可 72% 准确率预测政治倾向
Facial recognition can predict person’s political orientation with 72% accuracy
入门 · 约 45 分钟读完前置知识：机器学习基础、分类模型评估
导读
这篇发表于 Nature Scientific Reports 的研究利用面部识别/机器学习模型，根据人物照片预测其政治倾向（保守派 vs 自由派），准确率达到 72%，显著高于随机猜测（50%）、人类判断（55%）和 100 题人格问卷（66%）。研究使用约会网站约 97.8 万用户的自选头像及自报政治倾向数据，并在控制种族、年龄、性别等因素后准确率依然成立。
HN 关注点：这一结果背后的机制（是面部特征还是照片选择）以及隐私与滥用风险——面部图像竟能泄露政治立场。
HN 评论精选
这几条评论里：支持 1 质疑 2 补充 1
支持为研究辩护：常见质疑其实原文都已控制，真正的局限作者自己已列出
很多人没读原文就提出方法论质疑，先澄清几点：研究确实控制了照片中可能暗示政治立场的物体（图片紧密裁剪到面部）；模型显著优于抛硬币和人类判断者；也不只是识别种族、年龄或性别——在比较同种族、同年龄、同性别人群时依然准确。作者自己指出的局限包括：需要在标准化拍摄环境下控制胡须、妆容、表情和头部朝向做更细致研究；二元的政治倾向划分只是对复杂意识形态的粗略估计，且自报标签存在参照群体效应等问题。
utborin查看英文原文
There's an avalanche of people commenting on this who didn't bother to check the article before raising their methodological objections, so let's get these out of the way here. - Yes, they controlled for objects appearing in the pictures that might indicate political affiliation. The images are tightly cropped around the face. See Methods. - Yes, this is significantly better than both a coin flip and a human classifier. They gave the same test to humans, who did much worse than the model. See Abstract, Introduction, and Results. - Yes, this is doing more than just detecting a person's race, age, and/or gender. The classifier is still accurate when they compare people with the same race, age, and gender. See Results. If you want to discuss actual limitations in the study, here are some the author points out: - "A more detailed picture could be obtained by exploring the links between political orientation and facial features extracted from images taken in a standardized setting while controlling for facial hair, grooming, facial expression, and head orientation." - "Another factor affecting classification accuracy is the quality of the political orientation estimates. While the dichotomous representation used here (i.e., conservative vs. liberal) is widely used in the literature, it offers only a crude estimate of the complex interpersonal differences in ideology. Moreover, self-reported political labels suffer from the reference group effect: respondents’ tendency to assess their traits in the context of the salient comparison group."
补充关键区分：预测的是自选头像的特征，而非面部本身
这项研究并不能证明人的长相能反映政治倾向，它证明的是：人们挑选出来公开展示自己的照片，其特征能够指示政治倾向。
karpierz查看英文原文
This isn't a study that shows that people's faces indicate their political leanings. It's a study which shows that pictures that people select to represent themselves publicly have features that indicate political leaning.
质疑系统捕捉到人类察觉不到的信号，滥用前景令人担忧
关键在于系统远超人类：72% 对比人类的 55%、随机 50%，甚至超过 100 题人格问卷的 66%。这不是'认出老年人偏保守'那种人类已知的偏见，而且在比较其他条件相似的人时准确率也没降多少。系统捕捉到了我们注意不到的东西——也许是照片本身（自选的），也许是微表情。想想如果招聘经理在角落装个摄像头，根据你对工会的态度预测来决定录用谁，这既有趣又可怕。
mabbo查看英文原文
I think the key line is how much better this system is than humans attempting the same task: > Political orientation was correctly classified in 72% of liberal–conservative face pairs, remarkably better than chance (50%), human accuracy (55%), or one afforded by a 100-item personality questionnaire (66%). This isn't a matter of "recognize that old white people are conservative", because people will do that already, and they know all those biases. And the system doesn't lose much accuracy when comparing otherwise similar people. This system is picking up on things we don't notice. Maybe it's the photos themselves (they are self-selected), maybe it's micro expressions in the face, maybe it's something else entirely. But damn it's neat and maybe frightening. Imagine if your next hiring manager had a quiet little camera in the corner and chose you based on your predicted approval of unions.
质疑质疑数据伦理：用户并未同意被用于此类研究
约会网站样本包含 2017 年 977,777 名用户的头像、位置和自报政治倾向、性别、年龄。这些人显然没有同意或意识到自己会被用于面部识别研究。不过约会网站会卖这类数据，我一点也不惊讶。
ebegnwgnen查看英文原文
"The dating website sample was provided by a popular dating website in 2017. It contains profile images uploaded by 977,777 users; their location (country); and self-reported political orientation, gender, and age." It doesn't sound like people were consenting/aware that they'll endup on a facial recognition study. But I'm not even surprised dating website would sell their such data
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。
- 视频与演讲
历次讨论
- 2021Hacker News 讨论帖341 分 · 383 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